Persistent memory
State is saved at the end of each session and restored at the start. Decisions, preferences, and next steps survive window closes and tool switches.
Getting Started
Everything EGC provides out of the box.
The runtime capabilities that keep context, sessions and tool behavior aligned.
State is saved at the end of each session and restored at the start. Decisions, preferences, and next steps survive window closes and tool switches.
A user-wide scope shared across every project: save transversal preferences and lessons once with scope global, and every session in every tool starts already knowing them. Project state always wins.
Shell output is compressed up to 90% before it reaches the model: long git logs, test noise and install spam shrink while errors always survive. egc saved reports your accumulated savings at zero token cost.
Parallel sessions announce presence, split territory, take fail-fast cooperative path locks, and exchange events delivered in real time: wake-on-write push is on by default, session_wait parks idle tabs until an event arrives, and tools with native hook surfaces get a turn-boundary notice the moment the shared bus moves.
Populated memory never reaches a commit: pre-commit block, CI guard, and a git clean filter configured by egc init that stages a zeroed blob even when hooks are bypassed. Your context stays yours.
update_state propagates context to all 12 tool config files automatically. Run egc watch to also sync changes you make directly in any tool file back to state and to all other tools in real time.
Works with 28 AI coding tools. Install once and EGC sets up every tool you have installed with what it can read.
All state stays on your machine at ~/.egc/, encrypted at rest with AES-256-GCM. Nothing is sent externally.
npm install -g @egchq/egc && egc install handles everything: detects your tools, registers the servers, bootstraps the protocol. The prompt library stays opt-in.
egc-guardian enforces safety at the harness level: every shell command and file write is validated before execution. UserPromptSubmit and PreToolUse hooks block dangerous patterns without AI behavioral calls.
Content you own is cleaned of AI provenance marks automatically at write time: invisible Unicode carriers and long dashes are stripped, and AI co-authorship is removed from commit messages. An opt-in pass also strips provenance metadata from Markdown, HTML and SVG, from PNG and JPEG images, and from PDF Document Info dictionaries. Deterministic and fail-open, so a crash or binary input never blocks or corrupts a write. An opt-in statistical layer rewrites the prose to reduce token-sampling watermarks and measures how far the text moved: honestly best-effort, never claimed undetectable.
Every architectural decision is stored in SQLite with a timestamp. Query by date or tag.
Optional and off by default: the installer asks (default no), egc install --prompt-library adds it later, and the engine works exactly the same without it. Written from real experience, not generated.
Scroll the table horizontally for all columns.
| Type | Count | What it is |
|---|---|---|
| Agents | 61 | Persona and behavior definitions |
| Skills | 232 | Domain-specific workflow runbooks |
| Commands | 77 | Command definitions and lifecycle hooks |
| Rules | 109 | Constraints and governance directives |
Organized per harness: works in .cursor/, .claude/, .gemini/, and more. Switch tools and the same workflows follow you.
Scroll the table horizontally for all columns.
| Tool | MCP | Bootstrap |
|---|---|---|
| Claude Code | MCP | ~/.claude/CLAUDE.md |
| Antigravity CLI | MCP | ~/.gemini/GEMINI.md |
| Cursor | MCP | project .cursor/rules/ |
| Codex CLI | MCP | ~/.codex/config.toml |
| OpenCode | MCP | ~/.config/opencode/ (plugin, instructions, skills) |
| Devin Desktop | context | ~/.codeium/windsurf/ |
| Amp | context | ~/.amp/ |
| VS Code Copilot | context | ~/.github/ |
| Zed | MCP | ~/.config/zed/AGENTS.md |
| Kiro | MCP | ~/.kiro/hooks/ |
| Trae | context | ~/.trae/MEMORY.md |
| CodeBuddy | context | ~/.codebuddy/MEMORY.md |
| JetBrains Junie | context | .junie/guidelines.md |
| Goose | context | ~/.agents/skills/ |
| OpenHands | context | ~/.agents/skills/ |
| Aider | context | .aider/skills/ + .aider.conf.yml |
| Warp | context | .warp/skills/ + AGENTS.md index |
| Cline | context | .clinerules/ |
| Qwen Code | context | .qwen/skills/ |
EGC routes to any supported provider via ModelResolver. Fallback chains degrade automatically when a model is unavailable.
Scroll the table horizontally for all columns.
| Provider | Models |
|---|---|
| Google Gemini | Gemini 2.5 Pro / Flash / Flash-Lite, 2.0 Flash, 1.5 Pro / Flash |
| Anthropic Claude | Opus 4.5, Sonnet 4.7, Haiku 4.7 |
| OpenAI | GPT-4o, GPT-4o-mini |
| DeepSeek | deepseek-chat, deepseek-reasoner |
| Mistral AI | mistral-large-latest |
| Groq | gpt-oss-120b |
| Cohere | Command A |
| Vertex AI | Gemini through Google Cloud |
| OpenRouter | DeepSeek R1, DeepSeek Chat V3, Qwen3 235B, Qwen3 32B, Llama 4 Maverick, Llama 4 Scout, Llama 3.3 70B, Gemini 2.5 via OR, GPT-4o via OR, and more |
| Ollama | Llama 3.2, Mistral, CodeLlama (local, offline) |