There are two recommended ways to use lilbee. Run lilbee for the TUI if you're the one driving; wire it into your agent over MCP if an AI agent is. Both cover the everyday workflow: pick models, index files, search, chat, manage the wiki. You should not normally need to touch a CLI flag, an environment variable, or config.toml by hand for either path.

Everything in Reference for advanced users further down (CLI commands, the HTTP server, env vars, the config file) is for users who need to drive lilbee from outside those two paths: CI, scripts, headless boxes, custom integrations. Skip past it unless that's you.

• The TUI
  • First run
  • Adding documents
  • Search vs Chat mode
  • Model bar
  • Catalog screen
  • Settings screen
  • Slash commands
  • Task Center
  • Wiki
• Agent integration (MCP)
• Per-project libraries
• Keeping files out of the index
• Cloud models
• Reference for advanced users
  • CLI commands
  • HTTP server
  • Data locations
  • Config file (config.toml)
  • Environment variables
  • Settings reference
  • Optional extras
  • Cross-encoder reranking
  • Semantic chunking
  • OCR

lilbee (no args) launches the full Textual app: streaming chat with clickable citations, a model bar, a Task Center for background jobs, and screens for the catalog, settings, and the generated wiki.

Press ? at any time for the keybinding cheat sheet, Ctrl+P for the Textual command palette, and /help for the slash-command catalog. Every action lilbee can take is reachable from one of those three.

The TUI is a full-screen Textual app. It runs everywhere, including macOS Terminal.app, plain xterm and over SSH: every screen, panel and key works the same, and nothing is hidden on a lesser terminal.

Two things adapt to what the terminal can do, both detected automatically. Panel edges are drawn with partial-block glyphs, which only tile seamlessly in fonts that draw them cell-exact, so a terminal that does not gets box-drawing edges instead. And a terminal with 24-bit "true color" shows the themes as designed, while a 256-color one has no dark tinted colors to offer, so surfaces come out as neutral greys and the tint shows only in the accents. Layout, contrast and layering are the same either way.

• Give it room. The model bar and layout want about 100 columns; narrower and panels start to wrap.

• Over SSH is fine, but redraws are only as fast as the link, so a laggy connection makes scrolling and model switches feel heavier than they are.

• One multiplexer layer, not two. tmux and screen work well. Nested tmux (a local tmux, then SSH into a second tmux on the remote) is the one setup to avoid: two multiplexers fight over the same escape sequences and mangle box-drawing and the cursor. Attach to the remote session directly, or start lilbee outside the inner tmux.

• Want the full palette over SSH? SSH does not forward COLORTERM, so a remote lilbee uses 256 colors even when you are sitting in a truecolor terminal. If your local terminal really does render 24-bit color (iTerm2, kitty, WezTerm, Ghostty, Windows Terminal), forward it: SendEnv COLORTERM in ~/.ssh/config plus AcceptEnv COLORTERM in the server's sshd_config.

  Do not do this from macOS Terminal.app. It sets COLORTERM=truecolor while being unable to draw 24-bit color, and lilbee only knows to ignore that claim by reading TERM_PROGRAM, which SSH does not forward either. Forwarding COLORTERM from Terminal.app tells the remote to send color the terminal will render as garbage. Left alone, SSH from Terminal.app is already correct.

A fresh install opens on the model catalog. Each card carries a fit chip that says whether the model runs on this machine. Pick a chat model and pull it; the first model you install becomes the active one, and a toast tells you chat is ready. Models pulled earlier from the CLI are adopted automatically, and that launch lands straight in chat.

An embedding model works the same way, but nothing asks for one up front. The first feature that needs one (Search mode, /remember, indexing) sends you to the catalog's embedding tab.

After that you're in chat. /add indexes documents, /crawl indexes a website (crawler extra required), /settings opens the settings tabs.

Press /add to add files, directories, or web pages. Paths tab-complete; the job runs in the Task Center, so you can keep chatting while indexing happens.

If a file with the same name is already indexed, add skips it. To re-index in place, remove the document first with /delete name, or pass --force from the CLI.

Adding a code repository pulls in more than you want? See Keeping files out of the index.

The bar above the prompt has a two-state pill that toggles between Search and Chat. F3 flips it.

• Search (default). Every prompt goes through document retrieval first. Relevant chunks are passed to the chat model as context, and the reply ends with a Sources block of clickable citations. If retrieval finds nothing usable, lilbee says so plainly instead of answering from the model's general knowledge; switch to Chat mode when you want an off-corpus answer.
• Chat. Retrieval is skipped entirely; the model answers directly from whatever it already knows. Useful for conversational follow-ups that don't need your library, or for talking to a model when you haven't indexed anything yet.

The toggle is disabled and forced to Chat when no embedding model is configured (Search has nothing to search against).

The bar above the prompt shows what's active for chat and embedding, plus the mode toggle:

Chat [Qwen3 0.6B]   Embed [Nomic v1.5]   [Search | Chat]

The chat and embedding labels are searchable pickers. Click one (or focus it with Tab and press Enter) to open a modal with a search input above a virtualized list of every model that role can use. Type to filter; Enter picks the highlighted row; Escape cancels. The active model's display label sits on the button at all times so you can see what's loaded without opening the picker.

The pickers list everything the role can run: native GGUFs you already have installed plus, when the litellm extra is installed and an API key is set, whatever the SDK backend exposes for that provider. There is no separate "local-only" picker; routing happens automatically once the model is selected.

/models (or /m, /catalog) opens the catalog. The top of the screen has two sub-tabs:

• Local. Everything you can run on this machine. Native GGUF models from the developer's picks, the full HuggingFace catalog browsable by task and size, and any locally-running OpenAI-compatible backend that exposes models on its REST endpoint. Toggle between a card grid and a dense list view; both share the same search box. Pulling a model installs it; selecting an installed model assigns it to the matching role.
• Frontier. Cloud chat models grouped by provider (Anthropic, Gemini, OpenAI, and so on). Only appears when at least one provider API key is configured, either via the Settings screen's API-Keys tab or the provider's standard environment variable. The list shows whatever the provider's SDK exposes for that key with no curation on lilbee's side. Selecting a row makes that model the active chat model and routes subsequent prompts through the cloud provider; a persistent warning appears in the model bar so it's clear when chunks are leaving the machine.

/settings opens a tabbed settings editor. Every value persists to config.toml (see Config file); the equivalent env vars override individual values at runtime without touching the file.

Tabs for features that aren't installed are hidden, not greyed out:

• API-Keys appears only when the litellm extra is installed.
• Crawling appears only when the crawler extra is installed.
• Wiki appears only when the wiki layer is enabled (cfg.wiki = true in config.toml or LILBEE_WIKI=1).

Install the relevant extra (or flip the wiki flag) and the tab shows up on the next visit.

All slash commands available from the TUI. Slash commands and paths tab-complete; /help opens the same catalog live.

A spinner shows while waiting for the first token from the LLM.

Background jobs (sync, crawl, wiki build, model pull) appear in the Task Center with live progress. /cancel cancels the active operation. The TUI stays responsive throughout; each inference role (chat, embed, rerank, vision) runs in its own subprocess so a stuck model doesn't lock the chat.

lilbee analyzes the documents you've indexed and writes a wiki about them, inspired by Andrej Karpathy's LLM Wiki. Pages are per concept or entity rather than per document, so something that shows up repeatedly in your library gets its own page, written and cited from the source that mentions it most. Coverage across sources comes from synthesis pages (lilbee wiki synthesize) and the [[wiki link]] graph.

A build spends one LLM call per source document, so it is GPU-heavy and its cost scales with the size of your library, not with how many pages you read. A few dozen documents takes minutes; a few thousand takes hours. Nothing generates until you ask for it.

Open it with /wiki. Pages live under $LILBEE_DATA/wiki/:

Every section is citation-verified against the source chunks and scored for embedding faithfulness; low-confidence output routes to drafts/. Plain-text concept slugs inside page bodies are rewritten to Obsidian [[wiki links]] so the graph view shows how ideas connect. The directory is Obsidian-compatible out of the box.

Turning the wiki on does not generate anything by itself. You wikify explicitly: run lilbee wiki build (wiki update is the same full rebuild), press b on the wiki screen in the TUI, pick Wikify from the command palette, or call the build endpoints over HTTP or MCP. In the TUI the run happens in the background with progress in the Task Center; over HTTP the build, update, and synthesize routes stream per-phase and per-page progress as server-sent events.

If you would rather the wiki keep itself current, set wiki_auto_update (or LILBEE_WIKI_AUTO_UPDATE=1). Every sync then regenerates the pages its changed documents touched, without proposing new concepts, so day-to-day re-ingest never churns existing concept slugs. A sync that touches more than LILBEE_WIKI_INGEST_UPDATE_CAP pages (default 20) skips the regeneration and tells you to run lilbee wiki update yourself, so a bulk import cannot fire hundreds of LLM calls behind your back.

A build writes every page at once and spends one LLM call per source document, so its cost tracks how large your library is rather than how much of it you read. The browse index is the cheap half: lilbee wiki index (and every sync) extracts the names your documents mention with no LLM call at all, so the tree lists a page as soon as its document lands. lilbee wiki generate <slug> then writes one of those pages for one call, drawing on that subject's chunks across every source naming it, which is more evidence than a build gives it.

Turning the wiki back off stops new pages being written, but the pages already generated stay on disk and their rows stay in the index. Remove them with lilbee wiki wipe, the Delete wiki command in the TUI palette (or W on the wiki screen while the wiki is still on), DELETE /api/wiki, or the wiki_wipe MCP tool. The TUI offers it for you when you switch the setting off. A wipe deletes only what the wiki generated; your documents are untouched. While the wiki is off, generated pages are excluded from search results even if you never wipe them.

Lint, drafts review, and prune are available as CLI commands (see Wiki commands) and as MCP tools.

Conversations save automatically. The first message opens a session and names it after what you asked; each turn is appended as it lands. There is no save step.

• The drawer (ctrl+o or /sessions) docks the session list beside the live chat. Type to filter, enter resumes, ^n new chat, ^r rename, ^d delete.
• The Sessions tab is the same list full-screen.
• The CLI: lilbee sessions list / show / rename / delete (see Sessions commands).

Resuming restores the transcript and switches back to the model the conversation used, if it is still installed; otherwise lilbee keeps the current model and says so.

Sessions are append-only JSONL files under <data_dir>/sessions/, one per conversation. No database; back them up or sync them like any other file. The same surface exists over HTTP (list, read, create, append, rename, delete), so a script can own a conversation the way the TUI does. The TUI, HTTP server, and CLI are one conversation space: start a chat in Obsidian, continue it in the terminal.

Agents get the same tools over MCP, but they are off by default: most agent hosts already track their own history, and the tools cost context on every request. Turn on mcp_sessions_enabled if you want an agent keeping its own saved conversations. Agent sessions stay separate from yours: they never appear in your session list, and agents cannot list or read yours. The one bridge is explicit: an agent can take over a session whose id you hand it (claim=true, which moves it to the agent's space), and POST /api/sessions/{id}/claim brings one back.

To keep nothing, turn sessions_enabled off in Settings: nothing is written to disk, ctrl+o leaves the footer, and the Sessions view says sessions are off instead of showing an empty list. That covers the TUI, HTTP server, and CLI; mcp_sessions_enabled governs the agent half independently.

By default, turns that no longer fit the model's context window stop being sent. They stay on screen; a context gauge by the prompt shows the window filling, and a rule in the transcript marks where the model's view now begins.

To keep the old context instead, turn on chat_compaction in Settings. When the window fills, lilbee folds the oldest turns into short notes the model keeps reading. The fold is a model call (the gauge shows condensing… while it runs), around a second or two with a small model, even on CPU. Either way the transcript on screen and on disk is never touched; only what the model is sent changes.

lilbee loads the model lazily, the way Ollama and LM Studio do: the TUI never waits for the engine, a real llama.cpp server that starts loading in the background the moment the app opens. Ask something before it's ready and the answer bubble carries the load until your answer streams. Nothing freezes and nothing is silently queued.

One engine serves every lilbee process on the machine: open a second TUI, point a coding agent at lilbee mcp, or run lilbee serve alongside, and they all bind to the same loaded models instead of building their own. The engine stops when the last lilbee process exits, so the machine is left clean by default; no VRAM, RAM, or stray processes are held while nothing is running. Two knobs adjust that:

• Keep engine warm lets the engine outlive lilbee entirely, so the next session's first answer skips the load. Recommended if you relaunch often or use lilbee mainly through coding agents, whose sessions come and go.
• Engine idle ttl minutes bounds how long idle weights stay in memory in every mode, warm included, five minutes by default (the same idea as Ollama's keep_alive). The memory frees itself after that many idle minutes; a small proxy process stays behind, a few tens of MB and no VRAM. 0 keeps weights loaded until the engine stops.
• lilbee engine stop frees the shared engine immediately from any terminal, no TUI needed, whichever process started it. It reaches the machine slot and this project root's own overflow engine; a private overflow engine built for a different project root is stopped by running the command from that root.

Both knobs live in the TUI Settings screen, MCP lilbee_settings_set, the HTTP config API, and config.toml. Changing a model takes effect on your next prompt, from whichever surface you changed it. If no one else is using the engine it restarts on the new configuration; if other lilbee processes are serving from it, theirs keeps running and yours either rebinds, when the running engine already serves what you asked for, or starts its own alongside it. The first launch after a reboot is always a cold one.

To stop the engine without opening the TUI:

lilbee engine stop

It reports whether anything was running, frees the GPU immediately, and is safe to run at any time.

lilbee can remember durable facts about you and standing preferences for how you want answers, and recall them into context on later turns regardless of which conversation they came from. Memory is off by default, so users who don't want it pay nothing and never expose the injection surface. Turn it on once:

/set memory_enabled true

Then save things to remember. A leading pref: stores a standing preference (always recalled); anything else is stored as a fact (recalled by relevance):

/remember pref: keep answers terse, show code first
/remember the Crown Vic manual is my main brake-work reference

Next time you ask a question, the relevant preference and facts are folded into the system prompt before the model answers. Memory is never mixed into the document citations: it shapes the answer but only your actual sources show up under Sources:. Open /memories to list, delete, or share entries.

Memory lives in the active library's data directory, so it follows per-project libraries automatically. A factory /reset clears it; a /rebuild (which only re-indexes documents) leaves it alone. If you switch embedding models, your memories are re-embedded from their stored text during the rebuild, so recall keeps working.

Sharing model. Your own memories are private to you unless you mark them shared. Agent memories (see below) are private to that agent by default and are never folded into your prompt. This asymmetry keeps an agent's notes out of your chat unless you opt in.

Auto-extraction (optional, off by default). With memory_auto_extract on, the TUI runs a small background pass after each answer that saves durable facts and preferences from the exchange, so memory builds up as you chat. The saved memories are recalled like any others; review and prune them anytime in /memories.

/set memory_auto_extract true

The lilbee memory CLI group (add / list / recall / remove) and the REST /api/memories routes operate on the same memories from outside the TUI.

lilbee is also the retrieval backend for AI coding agents. Wire it into any agent that speaks MCP (Claude Code, opencode, Cursor, anything else) and the agent calls lilbee_search / lilbee_add and gets back cited snippets it can quote back. Your files, the embeddings, and the index stay on your computer. See the lilbee-mcp skill for the full MCP tool list and workflows. Non-MCP agents can use the JSON CLI fallback below.

lilbee agent-config <client> prints a paste-ready config for one client, filled in with the running server's address and session token:

lilbee agent-config claude     # Claude Code mcpServers block
lilbee agent-config opencode   # opencode.json provider + MCP server
lilbee agent-config hermes     # hermes config.yaml fragment
lilbee agent-config litellm    # LiteLLM proxy config.yaml snippet

opencode and hermes get lilbee as a model provider and as an MCP server. Claude Code brings its own model, so it gets the MCP tools only. The claude block registers the running server over HTTP; the alternative block printed on stderr has Claude Code start lilbee mcp itself instead.

The same blocks come off the running server over HTTP, which is what a GUI client should use. GET /api/agent-config lists the supported clients and whether each one's CLI is installed on this machine:

{
  "clients": [
    { "client": "claude", "cli_detected": true, "cli_path": "/opt/homebrew/bin/claude" },
    { "client": "hermes", "cli_detected": false, "cli_path": null },
    { "client": "opencode", "cli_detected": true, "cli_path": "/usr/local/bin/opencode" }
  ]
}

GET /api/agent-config/<client> returns that client's document. JSON clients get it in config, hermes gets its rendered YAML in content, and claude also carries the subprocess form in stdio_config. Both routes need the session token like every other route. Fetch the document when you need it rather than saving it: the token is minted fresh on every server start, so a copy written to disk stops working after a restart.

curl -H "Authorization: Bearer $(jq -r .token <data-dir>/server.json)" \
  http://127.0.0.1:8080/api/agent-config/claude

One daemon serves many agents at once, and two limits bite before the GPU does. Synchronous work is run off the event loop in a thread pool sized by mcp_tool_threads (40 by default, which is what the async runtime uses when nothing says otherwise); past that, retrieval calls queue while the disk and CPU sit idle, so raise it when a lot of agents search at the same time. Each connected agent also holds a socket, and macOS still defaults to 256 open files (most Linux distributions to 1024), which shows up as connection failures rather than slowness. Raise it in the shell you start the server from:

ulimit -n 4096
lilbee serve

The server logs the current limit at startup when it is low enough to matter.

Agents get their own memory through the lilbee_memory_remember and lilbee_memory_recall MCP tools (memory must be enabled first, e.g. the agent calls lilbee_settings_set({"memory_enabled": true})). An agent's memories are scoped to that agent and private by default: another agent won't see them, and they're never folded into the human TUI's prompt. An agent can recall its own memories plus any of yours you've marked shared.

Each agent's memories are namespaced by an owner like agent:opencode. lilbee derives the agent id from the MCP client name when it can, but pin it explicitly so it stays stable across sessions and clients. Set LILBEE_AGENT_ID in the agent's MCP server config:

{
  "mcpServers": {
    "lilbee": {
      "command": "lilbee",
      "args": ["mcp"],
      "env": { "LILBEE_AGENT_ID": "opencode" }
    }
  }
}

lilbee_memory_remember({"text": "retrieval knobs live in core/config/model.py", "kind": "fact"})
# -> stored under owner agent:opencode
lilbee_memory_recall({"query": "where are retrieval settings"})
# -> returns that note next session; lilbee_search results never mix in memory

Every lilbee CLI command accepts --json (or -j) before the subcommand for structured output. Use this as the shell-out path for agents that can't speak MCP. The shape mirrors the MCP tools: one JSON object per stdout line, errors return non-zero exit with {"error": "..."}, and distance scores are lower-is-more-relevant. Vectors are stripped from output.

Read (inline, no LLM):

lilbee --json status                           # indexed sources, models, totals
lilbee --json search "query" --top-k 12        # cited chunks (no LLM at query time)
lilbee --json chunks manual.pdf                # inspect how one source was chunked
lilbee --json topics "auth"                    # concept-graph view of a query
lilbee --json model list                       # installed models
lilbee --json model show <ref>                 # catalog + installed metadata for a model
lilbee --json version
lilbee --json self-check                       # runtime + model self-check

Write (LLM calls or long ops):

lilbee --json add ~/docs ~/notes               # register files / dirs, index them in place
lilbee --json add https://example.com/page     # URL becomes a markdown source
lilbee --json sync                             # re-index after edits to the documents directory
lilbee --json rebuild                          # nuke the index and re-ingest everything
lilbee --json remove manual.pdf                # drop chunks (source files are always kept)
lilbee --json remove reports/2024              # a folder: remove everything indexed beneath it
lilbee --json remove '**/*.log'                # a glob: remove every matching source
lilbee --json ask "question"                   # full local RAG (local llama-server fleet or remote backend)
lilbee --json model pull <ref>                 # download a model, streams JSON progress events
lilbee --json model pull <ref> --allow-unsupported  # override the architecture-compat check
lilbee --json model rm <ref>                   # delete an installed model
lilbee --json reset --yes                      # factory reset (destructive, requires --yes)
lilbee --json init [path]                      # create a .lilbee/ in a directory

add is the most common entry point: files, directories, and URLs all go through it, and indexing happens in the same call. Long ops take seconds to minutes; the final JSON includes per-file outcomes and counts.

Wiki (opt-in):

lilbee --json wiki status                      # page counts + wiki_enabled flag
lilbee --json wiki build                       # generate the topic / entity wiki
lilbee --json wiki update                      # refresh after a sync (full rebuild today)
lilbee --json wiki synthesize                  # cross-source synthesis pages
lilbee --json wiki lint                        # orphans, stale citations, pending drafts
lilbee --json wiki citations <page>            # citations a page makes
lilbee --json wiki citations --source <doc>    # pages citing a source document
lilbee --json wiki drafts list                 # pending drafts with drift + faithfulness
lilbee --json wiki drafts diff <slug>          # unified diff between a draft and the live page
lilbee --json wiki drafts accept <slug>        # publish a draft (concepts/, entities/, synthesis/, or summaries/)
lilbee --json wiki drafts reject <slug>        # discard a draft
lilbee --json wiki index                       # rebuild the browse index (no LLM call)
lilbee --json wiki generate <slug>             # write one indexed page (one LLM call)
lilbee --json wiki prune                       # archive stale pages
lilbee --json wiki wipe --yes                  # delete every generated page and its rows

Two patterns worth knowing:

• search vs ask. search returns raw chunks without an LLM call. Use it when your agent has its own LLM and just needs context from your files. ask runs lilbee's local RAG end-to-end and returns an answer with sources. Most non-MCP agents want search.
• Citation rule still applies. Every fact stated from search results must trace back to a chunk's source + line range, exactly as returned. Don't invent.

Output shape:

// lilbee --json search "oil change interval" --top-k 3
{"command": "search", "query": "oil change interval", "results": [
  {"source": "manual.pdf", "chunk": "Change oil every 5,000 miles...", "distance": 0.23, "chunk_type": "raw"}
]}

// lilbee --json status
{"config": {...}, "sources": [{"filename": "manual.pdf", "chunk_count": 42}], "document_count": 1, "total_chunks": 42, "index": {"embedding_model": "nomic-ai/nomic-embed-text-v1.5-GGUF/nomic-embed-text-v1.5.Q4_K_M.gguf", "embedding_dim": 768}, "skipped": [], "skipped_total": 0}

// lilbee --json model pull <ref>  (streams events, then a final "done" line)
{"event": "progress", "model": "...", "bytes": 12345678, "total": 999999999}
{"event": "done", "model": "...", "installed": true}

Gaps vs MCP. The CLI doesn't expose crawl (non-blocking URL crawling) or per-key settings management. Use add <url> for one-shot URL ingest. For continuous crawling or programmatic settings, the HTTP server exposes both: see the REST API reference.

lilbee uses a git-like per-project model. Running lilbee init from a project directory creates a .lilbee/ folder there, just like git init creates .git/. Once initialized, every lilbee invocation from that directory (or any subdirectory) automatically uses the local database, both in the TUI and from the CLI:

cd ~/projects/my-engine
lilbee init                  # creates .lilbee/ here
lilbee                       # launches the TUI scoped to this library

If there's no .lilbee/ in the current directory, lilbee walks up the tree looking for one; if none is found, it falls back to the global database at the platform default location (see Data locations). /status in the TUI (or lilbee status from the shell) shows which database is active.

Before you write a pattern, check what sync already excludes on its own:

• Formats it cannot read. lilbee indexes the file types it can extract, plus source code, and nothing else.
• Every dot-file and dot-directory, including .git/.
• A built-in directory list: node_modules, __pycache__, venv, build, dist, target, vendor, _build, coverage, htmlcov, and any name ending in .egg-info.

LILBEE_IGNORE_DIRS adds more directory names, comma-separated. It only adds names. There is no way to take a built-in name off the list.

LILBEE_IGNORE_DIRS=generated,fixtures lilbee sync

For what that list misses, write a .lilbeeignore. The syntax is .gitignore's: * and ** globs, a trailing / to match directories only, and a leading ! to add something back.

Use it when the problem is a supported file type rather than a directory name: vendored JavaScript, test fixtures, generated Markdown, a checked-in copy of site-packages.

lilbee never reads a repo's .gitignore. Repos commit binaries and vendor trees, and a repo that excludes docs/ from git holds the content you most want indexed.

One file covers everything you index. It sits in your data directory, so in a lilbee init project it is .lilbee/.lilbeeignore, and otherwise it is the global data directory. lilbee init writes a commented copy for you.

A .lilbeeignore inside a tree covers that directory and everything below it. Put one at the top of a repo to scope patterns to that repo, or deeper to narrow them further.

The deeper file wins, so it can add back what the wider one excludes:

# .lilbee/.lilbeeignore -- everything you index
*.min.js
*.parquet

# ~/code/myrepo/.lilbeeignore -- this repo only
testdata/
!jquery.min.js

sync never indexes the ignore files themselves.

The ! above works because *.min.js matches files. An excluded directory behaves differently: sync never walks into it, so no pattern below it can add a file back. !testdata/keep.md under testdata/ does nothing. This is git's rule.

To keep one file out of a tree you otherwise index, exclude the files rather than the directory.

The patterns govern what sync takes in. A pattern you add today stops those files being indexed from now on; it does not touch what an earlier sync already indexed. This matches .gitignore, which does not untrack a file you have already committed.

To drop what a new pattern excludes, ask for it:

lilbee sync --prune-ignored

That removes the matching documents from the index and reports them as Removed. Source files are never touched. The same option is on every surface: /prune-ignored in the TUI, {"prune_ignored": true} to /api/sync, and prune_ignored=true on lilbee_sync.

Nothing is written to hold a pruned file out. Delete the pattern and the next sync indexes it again, so a prune you regret costs a re-index, not the document.

lilbee always indexes a file you name yourself. lilbee add ~/notes/report.pdf beats a pattern that would exclude it, because naming a file is the stronger instruction.

lilbee runs entirely on your machine by default. There are two ways to use cloud models when you want to:

• Bring your own key, inside lilbee. Install the [litellm] extra and add an API key in /settings → API-Keys, then pick a cloud model from the model bar picker or the Frontier tab in /catalog. The TUI shows a persistent warning whenever a cloud role is active.
• Pair lilbee with a cloud agent over MCP. lilbee stays the local part: your files, the embeddings, the search index. The agent (Claude Code, opencode, anything that speaks MCP) calls lilbee_search / lilbee_add and gets back cited snippets. See Agent integration.

Either way your files and the index never leave the machine; only the queries and the snippets the model needs to answer cross the wire when you opt in.

Everything from here on is for users who want to script lilbee, run it headless, integrate it with an agent over MCP, or override individual settings without opening the TUI. None of it is required for everyday use; the TUI exposes all of these through pickers, /settings, and slash commands.

Every TUI action is also a CLI command. Indexing and search need an embedding model installed first (the default Nomic v1.5 GGUF is fine):

lilbee model pull nomic-ai/nomic-embed-text-v1.5-GGUF   # required before `lilbee add`
lilbee model browse                                     # pick a chat model interactively

Without an embedding model installed, every lilbee add call reports the file as failed with Model '…' not found in registry.

lilbee add ~/Documents/manual.pdf       # add a file
lilbee add ~/notes/                     # add a directory
lilbee add ~/docs/*.md ~/data/report.pdf
lilbee add manual.pdf --force           # re-index in place (default is skip-if-present)

lilbee search "oil change interval"
lilbee search "oil change interval" --top-k 20

lilbee ask "What is the recommended oil change interval?"
lilbee ask "Explain this" --model qwen3

search only needs the embedding model; ask also needs a chat model.

lilbee wiki build                      # build the wiki from the current index
lilbee wiki build --dry-run            # preview the entity candidates, no LLM calls
lilbee wiki list                       # list pages with type, source count, date
lilbee wiki read <slug>                # print a page's markdown
lilbee wiki citations <page>           # citations a page makes
lilbee wiki citations --source <doc>   # pages citing a source document
lilbee wiki lint                       # find orphan pages, stale links, pending drafts
lilbee wiki synthesize                 # generate cross-source synthesis pages
lilbee wiki drafts list                # list pending drafts
lilbee wiki drafts accept <slug>       # publish a draft (concepts/, entities/, synthesis/, or summaries/)
lilbee wiki drafts reject <slug>       # discard a draft
lilbee wiki index                      # rebuild the browse index (no LLM call)
lilbee wiki generate <slug>            # write one indexed page (one LLM call)
lilbee wiki prune                      # move stale pages to archive/
lilbee wiki wipe                       # delete every generated page and its rows

MCP tools mirror the CLI: wiki_list, wiki_read, wiki_status, wiki_build, wiki_update, wiki_synthesize, wiki_lint, wiki_citations, wiki_drafts_list, wiki_drafts_diff, wiki_prune. Accepting and rejecting drafts is left off MCP on purpose: deciding whether a low-confidence page is good enough to publish is your call, so use the CLI, the TUI, or the authenticated HTTP API.

Requires memory enabled (LILBEE_MEMORY_ENABLED=1, or /set memory_enabled true in the TUI). See Memory for the full model.

lilbee memory add "the project uses rust"        # remember a fact
lilbee memory add "answer tersely" --preference  # remember a standing preference
lilbee memory add "uses rust" --shared           # also expose it to agents
lilbee memory list                               # show stored memories
lilbee memory recall "what language"             # recall facts by relevance
lilbee memory remove <id>                        # delete a memory by id

See Sessions. Ids accept any unique prefix.

lilbee sessions list                   # saved conversations, newest first
lilbee sessions show 3f2a              # print a transcript by id prefix
lilbee sessions rename 3f2a "Brake specs"
lilbee sessions delete 3f2a            # asks first; --yes skips the prompt
lilbee --json sessions show 3f2a       # transcript + compaction summary as JSON

lilbee init                            # create .lilbee/ in cwd
lilbee status                          # show active data dir, models, document count
lilbee --global status                 # skip any .lilbee/ and use the platform default
lilbee --data-dir ~/kb status          # use an explicit data dir

--model / -m, --data-dir / -d, --global / -g, --vision, --vision-timeout, --log-level, --json / -j, --version / -V.

The HTTP server exposes a REST API that any tool or GUI can hit. By default it picks a random port and writes it to <data_dir>/server.port so callers on the same machine can discover it. Start it with lilbee serve:

lilbee serve                           # random port
lilbee serve --port 8080               # fixed port
lilbee serve --host 0.0.0.0            # bind all interfaces (default: 127.0.0.1)

One server runs per data dir: a second lilbee serve against the same --data-dir waits up to fifteen seconds for the first to exit, then stops with an error (exit code 3) instead of competing for the port file and the model engine. A supervisor that manages several data dirs (the Obsidian plugin's shared root) can pass LILBEE_EXCLUSIVE_SCOPE=<dir> to extend the same guarantee across all of them; the refusal then names the data dir the running server is serving. To replace a running server, ask it to stop with POST /api/shutdown (token-authed): it shuts down exactly as an external SIGTERM would, and its locks release the moment it exits.

Every route needs the session token, reads included, GET /api/health among them; the daemon writes the token to server.json (mode 0600) next to the port file, and lilbee agent-config hands it to local clients. The surface covers search (with SSE streaming variants for ask and chat), document lifecycle, crawling, model management, memory (GET/POST/PATCH/DELETE /api/memories, when memory is enabled), saved conversations (/api/sessions: list, read, create, append, rename, delete, and the compaction summary), configuration (including a defaults endpoint that powers per-setting reset), and status/health. The Obsidian plugin uses the /api/source endpoint for vault-aware source retrieval. Interactive REST API docs live at /schema/redoc when the server is running, and the full OpenAPI schema is published at the REST API reference. (Note: this is the HTTP server reference. A Python-library API reference is still being written; for now, the source under src/lilbee/ is the canonical reference.)

Server-specific env vars live in the Server table below.

For tools that talk to lilbee's HTTP REST API all day (the Obsidian plugin, custom GUIs, anything hitting /api/*), your OS launcher can keep the HTTP server warm so requests skip the cold-start. This is the only lilbee surface designed for that pattern; the TUI, lilbee chat, the MCP server, and the rest of the CLI cold-start and exit on every invocation by design.

Pull at least one chat and embedding model first (lilbee model pull <name>). The daemon will start without one, but /api/* requests fail until a model is available. All recipes pin the server to 127.0.0.1:42697.

macOS (Homebrew):

brew services start lilbee

Linux (Arch / AUR, systemd):

systemctl --user enable --now lilbee

On a headless server (no graphical login session), also run loginctl enable-linger $USER so the service survives logout.

NixOS: add the module to your configuration.nix:

{
  imports = [ lilbee.nixosModules.lilbee ];
  services.lilbee.enable = true;
}

lilbee resolves the data directory in this order (highest priority first):

Run lilbee init to create a .lilbee/ directory in your project. It contains documents/ (your indexed files), data/ (the search index), a .gitignore that excludes derived data, and a commented .lilbeeignore to fill in.

Every persisted lilbee setting lives in a single TOML file at <data-dir>/config.toml. For the platform default that's ~/.local/share/lilbee/config.toml; for a project, it's .lilbee/config.toml next to your code. Same file, same shape, regardless of which data dir you're on; the path just follows wherever the data dir is resolved to.

Sections mirror the setting groups in the settings reference: [general], [retrieval], [chunking], [generation], [server], etc. Example:

# .lilbee/config.toml

[general]
chat_model = "Qwen/Qwen3-8B-GGUF/Qwen3-8B-Q4_K_M.gguf"
theme = "rose-pine"

[retrieval]
top_k = 12
max_distance = 0.6
reranker_model = "Qwen/Qwen3-Reranker-0.6B-GGUF/qwen3-reranker-0.6b.Q8_0.gguf"

[generation]
temperature = 0.2
num_ctx_max = 32768

You don't have to write the file by hand: the TUI's /settings screen and the /set <key> <value> slash command both persist to it, and the equivalent LILBEE_<KEY>=... env vars override individual values at runtime without touching the file.

Every setting is also an environment variable, named LILBEE_ plus the setting in upper case: top_k is LILBEE_TOP_K, wiki is LILBEE_WIKI. An environment variable applies to the process you launch and does not persist; /settings and config.toml persist.

A handful of variables are not settings and so have no config.toml entry: LILBEE_DATA (the data directory), LILBEE_LOG_LEVEL, LILBEE_ENGINE_DIR, LILBEE_TOKEN, and a few ingest and debugging knobs. The settings reference lists them under Environment-only variables.

docs/settings.md lists every setting, with its type, default, help text, and a column per surface saying whether the TUI, MCP, HTTP API, or CLI can change it. It is generated from lilbee's own configuration fields, so it never drifts from the code.

Extras exist only on a pip or uv install. If you installed a bundled build (standalone binary, Homebrew, AUR, Nix, Docker, Flatpak, Snap, Scoop) all four are already inside it and none of this applies to you.

[engine] is the llama-server that runs your models, and it is the one extra that is not really optional. [graph], [crawler] and [litellm] add capabilities you can happily live without.

pip install --pre 'lilbee[engine]' --extra-index-url https://lilbee.sh/cu125/    # NVIDIA (CUDA)
pip install --pre 'lilbee[engine]' --extra-index-url https://lilbee.sh/rocm/     # AMD (ROCm)
pip install --pre 'lilbee[engine]' --extra-index-url https://lilbee.sh/metal/    # Apple silicon
pip install --pre 'lilbee[engine]' --extra-index-url https://lilbee.sh/vulkan/   # other GPUs
pip install --pre 'lilbee[engine]' --extra-index-url https://lilbee.sh/cpu/      # no GPU

# pip
pip install --pre 'lilbee[graph]'      # concept graph: topic clustering + search boosting
pip install --pre 'lilbee[crawler]'    # web crawling: index websites alongside local docs
pip install --pre 'lilbee[litellm]'    # remote providers: connect to any SDK-backed provider

# uv tool
uv tool install --prerelease=allow 'lilbee[graph]'
uv tool install --prerelease=allow 'lilbee[crawler]'
uv tool install --prerelease=allow 'lilbee[litellm]'

pip install --pre 'lilbee[engine,graph,crawler,litellm]' --extra-index-url https://lilbee.sh/cu125/
uv tool install --prerelease=allow 'lilbee[engine,graph,crawler,litellm]' --extra-index-url https://lilbee.sh/cu125/

NVIDIA users: the default Vulkan build works, but the CUDA flavour is faster and dodges the Vulkan-loader crash that affects NVIDIA-on-Windows setups. Same lilbee command, links straight against your NVIDIA driver:

brew install tobocop2/lilbee/lilbee-cuda
paru -S lilbee-cuda
nix run github:tobocop2/lilbee#lilbee-cuda
pip install --pre 'lilbee[engine]' --extra-index-url https://lilbee.sh/cu125/   # devs

The lilbee-cuda AUR package conflicts with lilbee and provides it, so paru -S lilbee-cuda swaps automatically. On Homebrew, run brew uninstall lilbee first because the two formulas both ship a lilbee binary. For older drivers swap cu125 for cu124 or cu121 (run nvidia-smi to see which matches). cu121 and cu124 are wheel-index and direct-download only; cu125 is the variant fanned out to the package managers.

While 0.6.66 is in beta, the --pre flag (or uv's --prerelease=allow) is required on every install.

Cross-encoder reranking is built in (no extra required); see Cross-encoder reranking below.

Builds a topic map of your documents at index time. Related concepts are linked in a co-occurrence graph, which is used to boost search results and expand queries with related terms, all without extra LLM calls.

What it does: Extracts noun phrases from every chunk using spaCy, computes PMI co-occurrence weights between concepts, and clusters them with the Leiden algorithm. At search time, queries are expanded with graph neighbors and results overlapping query concepts get a relevance boost.

When to use it: Large corpora (100+ documents) where the same topics appear across multiple files. The graph helps surface connections that pure vector similarity misses. For example, finding "deployment" documents when searching for "CI/CD" because those concepts co-occur frequently.

Install: pip install --pre 'lilbee[graph]' or uv tool install --prerelease=allow 'lilbee[graph]'

Configuration:

export LILBEE_CONCEPT_GRAPH=true              # enable (default: true when deps installed)
export LILBEE_CONCEPT_BOOST_WEIGHT=0.3        # how much concept overlap matters (0.0-1.0)
export LILBEE_CONCEPT_MAX_PER_CHUNK=5         # max concepts extracted per chunk (default)

The graph is built automatically during lilbee sync. No extra commands needed; search results are boosted transparently.

Based on: Microsoft Research's LazyGraphRAG technique, Church & Hanks 1990 (PMI), Traag et al. 2019 (Leiden).

Index web pages alongside your local documents. Crawl single pages or follow links recursively.

What it does: Fetches web pages using a headless browser (Playwright), extracts markdown content, and indexes it. Supports recursive crawling with configurable depth, concurrent fetching, live progress, cancel, per-domain rate-limit + retries on HTTP 429/503, and SSRF protection against internal network access.

When to use it: When your library spans both local files and web content such as documentation sites, wikis, or internal tools. Crawled content is hash-tracked so re-crawling only re-indexes changed pages.

Install: pip install --pre 'lilbee[crawler]' or uv tool install --prerelease=allow 'lilbee[crawler]'

Usage:

# Single page (no --crawl)
lilbee add https://docs.example.com/guide

# Whole-site crawl (recursive, unbounded by default)
lilbee add https://docs.example.com --crawl

# Cap depth or page count
lilbee add https://docs.example.com --crawl --depth 2 --max-pages 200

# Multiple URLs
lilbee add https://docs.example.com https://wiki.example.com

Also available via MCP (crawl), REST API (POST /api/crawl), and TUI (/crawl). The MCP tool defaults to a single page (depth=0); agents pass a depth, or depth=null for the whole site.

Configuration (all optional):

# Global ceilings. Unset = no cap. Explicit --depth/--max-pages always win.
export LILBEE_CRAWL_MAX_DEPTH=3          # cap link-following depth
export LILBEE_CRAWL_MAX_PAGES=1000       # cap total pages

# Pacing within a single crawl.
export LILBEE_CRAWL_MEAN_DELAY=0.5       # seconds between requests
export LILBEE_CRAWL_MAX_DELAY_RANGE=0.5  # random jitter on top
export LILBEE_CRAWL_CONCURRENT_REQUESTS=3

# Per-domain rate-limit + retries on HTTP 429/503.
export LILBEE_CRAWL_RETRY_ON_RATE_LIMIT=true
export LILBEE_CRAWL_RETRY_BASE_DELAY_MIN=1.0
export LILBEE_CRAWL_RETRY_BASE_DELAY_MAX=3.0
export LILBEE_CRAWL_RETRY_MAX_BACKOFF=30.0
export LILBEE_CRAWL_RETRY_MAX_ATTEMPTS=3

# Other.
export LILBEE_CRAWL_TIMEOUT=30           # per-page timeout (seconds)
export LILBEE_CRAWL_MAX_CONCURRENT=0     # 0 = CPU count (top-level concurrency)
export LILBEE_CRAWL_SYNC_INTERVAL=30     # seconds between periodic syncs during crawl

Connect to hosted or local OpenAI-compatible LLM backends alongside lilbee's managed local llama-server engine.

What it does: Routes chat and embedding calls to any provider reachable via the SDK backend. The routing provider automatically detects which models are available locally vs. remotely and routes each call to the right backend.

When to use it: When you want a frontier model for chat while keeping embeddings local for privacy, or to surface models from a local OpenAI-compatible daemon alongside lilbee's native GGUF models.

Install: pip install --pre 'lilbee[litellm]' or uv tool install --prerelease=allow 'lilbee[litellm]'.

Configuration:

export LILBEE_LLM_PROVIDER=auto          # "auto" routes between local and remote
export LILBEE_OLLAMA_BASE_URL=http://localhost:11434  # Ollama URL (blank = default)
export LILBEE_LM_STUDIO_BASE_URL=http://localhost:1234/v1  # LM Studio URL (blank = default)
export LILBEE_LLM_API_KEY=sk-...         # API key for your provider
export LILBEE_CHAT_MODEL=your-model      # any remotely-supported model name

Provider options: auto (default; native GGUF models run on the local managed llama-server fleet, remote model names route to the SDK backend) and remote (everything goes to an external OpenAI-compatible endpoint).

By default lilbee decides which GPUs each model goes on automatically. It bin-packs all four roles (chat, embed, rerank, vision) across your available GPUs and tensor-splits anything too large for one card.

If you want to pin specific models to specific cards, you can set a placement spec. The spec is a JSON object with one entry per role you want to control:

{
  "chat":   { "devices": [0, 1], "tensor_split": [1, 1] },
  "embed":  { "devices": [2] },
  "rerank": { "devices": [2] },
  "vision": { "devices": [3] }
}

Each role takes devices (GPU indices), an optional tensor_split list for spreading a single model across cards, and an optional replicas count for the embed and vision roles. Omit a role and the auto planner handles it.

To see what the auto planner would assign (without changing anything):

lilbee placement preview

To apply a spec from a file:

lilbee placement set --spec placement.json

To see the current spec (or confirm auto is active):

lilbee placement show

To go back to automatic:

lilbee placement clear

The same operations are available in the TUI under the Placement screen and over MCP (set_placement, clear_placement). Applying or clearing placement rebuilds the shared fleet, so over HTTP PUT/DELETE /api/placement are refused by default. Set allow_http_placement (or LILBEE_ALLOW_HTTP_PLACEMENT=1) to enable them on a single-client or owned deployment, such as the Obsidian plugin's managed server or a personally-owned pod where you run lilbee serve yourself. The spec persists across restarts.

If a pinned placement no longer fits the card it names (after a hardware change, for example), lilbee surfaces an error naming the card rather than starting in a broken state. lilbee placement clear returns to automatic placement in that case.

How much of each card, and of host RAM, the planner is allowed to promise. Defaults suit most machines; these exist for the cases where they don't.

Raise usable_vram_fraction toward 1.0 on a dedicated box with nothing else on the card. Lower it if loads fail with out-of-memory despite the planner reporting the model fits.

Lower system_memory_reserve_gb on a machine with a lot of RAM and nothing else running, if a model that offloads to host memory is being refused.

Both are writable at runtime through /settings, lilbee config set, and MCP.

If a model is refused for memory it looks like it should have: run lilbee placement preview to see what the planner budgeted. The planner charges free VRAM rather than the card's total, so another process holding memory shows up here. On a partially offloaded model, check system_memory_reserve_gb too: the host side is budgeted separately from the card.

If a load dies of out-of-memory, lilbee re-plans that role smaller and retries rather than respawning the same launch. Repeated shrink-and-retry messages for one model mean the estimate is too optimistic for that card; lowering usable_vram_fraction fixes it more directly than pinning placement.

Built-in. Re-scores retrieval candidates with a cross-encoder for precision on the top results. Unlike the extras above, no extra install is required; reranking is off by default and turns on as soon as you set LILBEE_RERANKER_MODEL (or pick a reranker from /settings).

What it does: After hybrid search (BM25 and vector arms fused into one canonical relevance score) returns candidates, a GGUF cross-encoder scores each (query, chunk) pair and results are blended with position-aware weights. Top-ranked candidates keep more of the original ranking; lower-ranked candidates trust the reranker more.

When to use it: When you need high-precision answers and are willing to trade roughly 200 to 500 ms per query. Most useful with large candidate sets where top-5 ordering matters.

Configuration:

export LILBEE_RERANKER_MODEL="bge-reranker-v2-m3"   # any GGUF reranker
export LILBEE_RERANK_CANDIDATES=20                  # how many candidates to rerank

Without a reranker set, fused hybrid search already provides good results for most use cases.

Based on: Nogueira & Cho 2019 (Passage Re-ranking with BERT), Burges et al. 2005 (Learning to Rank).

Off by default. lilbee ships with two chunking strategies; which one serves you depends on what you're indexing.

Fixed-size (default). Breaks documents into roughly equal token windows with overlap. Fast, deterministic, works well on code, reference manuals, user guides, API specs, and anything with clear structural boundaries. The assumption is that each chunk only needs to be coherent enough for retrieval, and the model will handle the rest from a small window of context.

Semantic. Uses embedding similarity to detect topic boundaries and splits there instead of at fixed sizes. Each chunk tends to represent one coherent thought rather than an arbitrary slice through one. The benefit shows up on prose-heavy material: novels, essays, long-form research papers, interview transcripts, qualitative research notes, anything where an argument develops across paragraphs. When you ask a question, the retrieved chunk is more likely to contain the full section that matches rather than the first half of it plus unrelated setup.

Trade-off: Enabling semantic chunking triggers a one-time download of xberg's ONNX embedding model (separate from the chunk-to-vector embedder) and runs roughly 9x more downstream embedding calls during indexing. Indexing takes longer; retrieval latency is unchanged.

Three equivalent paths:

# Environment variable
export LILBEE_SEMANTIC_CHUNKING=true

# TUI /set command (interactive)
/set semantic_chunking true

# config.toml in your .lilbee/ vault
[general]
semantic_chunking = true

After enabling, run lilbee rebuild so existing documents are re-chunked under the new strategy. New documents added from that point use semantic chunking automatically.

export LILBEE_TOPIC_THRESHOLD=0.75   # cosine threshold for topic boundaries (0.0-1.0)

OCR is how lilbee gets text out of scanned PDFs and images. It's one step in indexing, not a substitute for it: however the text comes out (a native parser, Tesseract, or a vision model), it still gets embedded, so you still need an embedding model installed.

For PDFs without embedded text, lilbee supports two OCR backends. When a vision model is configured, it takes precedence.

Tesseract is used automatically when no vision model is configured. No flags needed.

brew install tesseract          # macOS
sudo apt install tesseract-ocr  # Ubuntu/Debian

By default Tesseract reads English (eng). To OCR other languages, set the language codes; xberg downloads the Tesseract data on first use. Join multiple languages with + (Tesseract's own convention), for example eng+deu for English and German. The setting is the same value everywhere:

export LILBEE_OCR_LANGUAGE=eng+deu   # environment, this process only

# .lilbee/config.toml -- persisted
ocr_language = "eng+deu"

You can also set it in the /settings screen in the TUI, or over the HTTP and MCP settings APIs. Commas work in place of + if you prefer (eng,deu).

lilbee runs vision OCR in one of two ways:

1. Local vision model. Point LILBEE_VISION_MODEL at a GGUF vision model (e.g. lightonocr) and lilbee serves it on llama-server with an --mmproj projector. This is the recommended path and supports an SSE heartbeat for long scans.
2. Remote vision model. With pip install --pre 'lilbee[litellm]' (or uv tool install --prerelease=allow 'lilbee[litellm]'), set the vision model to any remote name your SDK backend understands. lilbee will route vision calls accordingly.

lilbee add report.pdf --vision                # prompts for model if none set
lilbee add report.pdf --vision-timeout 30     # per-page timeout (default: 120s, 0 = no limit)
export LILBEE_VISION_MODEL=lightonocr         # persist across runs (GGUF via mtmd)

Pick or change a vision model interactively via /settings in the TUI; the selection is saved to config.toml and persists across sessions.