DreamGraph v7.0 — El Alarife

The autonomous cognitive layer for sovereign development.

DreamGraph is a model-agnostic, graph-grounded development agent. It replaces brute-force context dumping with architectural discipline — shipping the 2.4 kB of entity-specific knowledge that matters instead of the 800 kB of raw source that costs. Your project's logic stays persistent in the Graph regardless of which LLM you plug in.

Grounded in the Graph. Built for the Master Builder.

Seed, scan, and enrich a knowledge graph with features, workflows, data models, validated edges, and source anchors. The graph is authoritative; everything else orbits it.

68 tools and 26 resources via the Model Context Protocol. The daemon delegates to scanners, extractors, and LLMs while keeping the graph consistent and traceable.

Dream cycles generate hypotheses, validate/promote/decay edges, and converge on truth. 10 dream strategies auto-adapt. ADR recording and guard rails govern change. The 5-state cognitive machine (AWAKE → REM → NORMALIZING → NIGHTMARE → LUCID) enforces isolation between speculation and fact.

The dg command manages the full lifecycle: create, attach, start, scan, schedule, export, fork, archive, destroy. See CLI Reference below.

Zero-dependency browser UI at http://localhost:<port>/ — cognitive status, schedule management, runtime LLM config, and auto-generated knowledge docs with Mermaid diagrams.

Native VS Code sidebar with Chat, Dashboard, and Files Changed panels. The Architect LLM reasons over the graph first, makes targeted edits with verification, and syncs the graph so knowledge and code never drift. Change model and provider mid-flight without losing context — the graph persists independently of any LLM session.

• Node.js ≥ 18 · Git · An LLM API key (OpenAI, Anthropic, or Ollama for local)

# Windows
git clone https://github.com/mmethodz/dreamgraph.git
cd dreamgraph
.\scripts\install.ps1 -Force

# Linux / macOS
git clone https://github.com/mmethodz/dreamgraph.git
cd dreamgraph
bash scripts/install.sh

This builds the server, installs the dg CLI globally, and automatically installs the VS Code extension if VS Code is detected.

dg init my-project                        # Create a DreamGraph instance
dg attach my-project /path/to/your/repo   # Bind to your project
dg start my-project --http                # Start the daemon
code /path/to/your/repo                   # Extension auto-connects

The daemon auto-bootstraps: scans your project, runs an initial dream cycle, discovers implicit ADRs, and schedules follow-up dreams. Open the Chat panel and start asking immediately.

1. The extension installs automatically via install.ps1 / install.sh.
2. Open your project in VS Code — the extension discovers the instance and connects.
3. The DreamGraph icon appears in the activity bar (left sidebar). Click it to open Chat, Dashboard, and Files Changed.
4. Set your API key: Ctrl+Shift+P → DreamGraph: Set Architect API Key.
5. Configure model: VS Code Settings → dreamgraph.architect.provider / dreamgraph.architect.model.

You can change provider and model at any time — the graph-grounded context is independent of the LLM session, so switching from Sonnet to Opus (or to a local Ollama model) loses nothing.

# Engine dreamer (daemon-side) — set in instance config
# ~/.dreamgraph/instances/<uuid>/config/engine.env
DG_LLM_PROVIDER=anthropic
DG_LLM_API_KEY=sk-ant-...
DG_LLM_MODEL=claude-sonnet-4-20250514

# Extension Architect (VS Code-side)
# Settings → dreamgraph.architect.provider / model
# API key → Ctrl+Shift+P → DreamGraph: Set Architect API Key

dg schedule my-project --add --name "nightly" --action dream_cycle --type interval --interval 300

The DreamGraph sidebar gives you one-click access to the full cognitive engine.

Traditional AI coding assistants dump entire source files into the LLM context — 50–800 kB per request. DreamGraph's Graph RAG assembles entity-specific context (features, edges, data models, tensions) that typically fits in 1–5 kB. On models like Claude Opus 4.6 at $15/M input tokens, this is a 10–30× cost reduction per interaction with better results, because the model gets precisely the knowledge it needs instead of parsing irrelevant code.

68 tools · 26 resources via the Model Context Protocol.

Full parameter tables: docs/tools-reference.md

+------------------------------------------------------+
|                    VS Code Extension                  |
|  +--------+  +-----------+  +----------------------+ |
|  |  Chat  |  | Dashboard |  |  Files Changed       | |
|  +---+----+  +-----+-----+  +----------+-----------+ |
|      |              |                   |              |
|      +----------+---+-------------------+              |
|                 |  MCP over HTTP                        |
+-----------------+--------------------------------------+
                  |
+-----------------v--------------------------------------+
|              DreamGraph MCP Server (Daemon)              |
|                                                          |
|  +-------------+  +--------------+  +----------------+  |
|  |  Cognitive   |  |  68 MCP      |  |  26 MCP        |  |
|  |  Engine      |  |  Tools       |  |  Resources     |  |
|  |  (5 states)  |  |              |  |                |  |
|  +------+------+  +------+-------+  +-------+--------+  |
|         |                |                   |           |
|  +------v----------------v-------------------v--------+  |
|  |              Knowledge Graph (JSON)                 |  |
|  |        21 data stores · per-instance isolation      |  |
|  +----------------------------------------------------+  |
|                                                          |
|  +--------------+  +--------------+  +---------------+  |
|  | Web Dashboard|  | REST API     |  | Discipline    |  |
|  | (zero-dep)   |  | (HTTP)       |  | System (ADR)  |  |
|  +--------------+  +--------------+  +---------------+  |
+----------------------------------------------------------+

src/
├── cognitive/      # State machine, dreamer, normalizer, strategies, LLM, scheduler
├── tools/          # 68 MCP tools
├── resources/      # 26 MCP resources
├── instance/       # Multi-instance lifecycle, registry, policies, bootstrap
├── cli/            # dg binary — instance management
├── server/         # MCP server, HTTP daemon, web dashboard
├── config/         # Environment-driven configuration
├── api/            # REST API routes
└── utils/          # Cache, logger, mutex, metrics, paths

extensions/vscode/  # VS Code extension — Chat, Dashboard, Files Changed
scripts/            # install.ps1, install.sh
docs/               # Architecture, cognitive engine, tools, data model, workflows

• Cognitive isolation — REM and NIGHTMARE cannot write to the fact graph
• Truth Filter — multi-signal scoring with hard promotion threshold (≥ 0.62)
• Decay and TTL — stale hypotheses and tensions are automatically forgotten
• Tension cap — max 200 active tensions prevents runaway speculation
• Interrupt safety — any state returns to AWAKE with in-progress data quarantined
• Discipline system — phase-locked tool permissions with data protection tiers

Contributions welcome in: normalization strategies, graph modeling, tension heuristics, dream strategies, adversarial scan patterns, temporal analysis, federation protocols, and performance.

DreamGraph is source-available under the DreamGraph License (BSL-based) — see LICENSE.

• Free for personal, research, and internal commercial use
• Not allowed to offer as a competing service or platform

Commercial licensing: mika.jussila@siteledger.io