EGC

New session? > egc restore ✓

Context that never resets.

EGC gives the AI tools on your machine the same project memory. Decisions, preferences and next steps come back in every session, in any tool.

npm install -g @egchq/egc
  • v1.1.22
  • Apache-2.0
  • 28 tools
  • 100% local

01 The problem

Why start over every time?

Close the window and the context is gone: your stack preferences, last week’s architecture decisions, the approach that failed after three attempts. Switch from Cursor to Claude Code and you start over again.

  • you ›We use PostgreSQL with row-level security…
  • you ›I prefer functional components, no classes…
  • you ›We moved from Prisma to Drizzle because…
Illustrative comparison: without EGC, the context available to the AI drops to zero at every new session; with EGC it is restored and builds up across sessions.

Without EGC

  • Repeating context in every session
  • Losing important decisions
  • Re-explaining project structure

With EGC

  • Context restored in every new session
  • Decisions kept with the project
  • Switch tools, keep going

02 Demo

One decision, two tools.

Follow one decision: recorded in Cursor, kept by EGC and restored in Claude Code, without you repeating anything.

Illustrative demoA simulated interface that explains the flow. Not a real recording.

Step 1 of 5

The session records a decision

In Cursor you decide to switch ORMs. The AI saves the decision with store_decision.

Step 2 of 5

The session is closed

You close the window. The decision stays encrypted in ~/.egc, on your machine.

Step 3 of 5

Another tool is opened

Days later you open Claude Code on the same project and just say “let’s continue”.

Step 4 of 5

EGC restores the decision

EGC loads the project state and hands the decision to the new session.

Step 5 of 5

Work continues

The AI picks up where you left off, already knowing what was decided.

Watch the real recording

The simulation above explains the flow. To see EGC for real, watch the recorded demo.

When you press play, the video loads from YouTube.

03 How it works

One brain, four faculties.

One local engine, on from the first install in every supported tool. Each circle in the mark is one of the parts working together for you.

#01 Memory

What one agent learns, every agent knows.

Decisions and lessons are kept and come back in any tool.

Technical details

Decisions, session context, working memory and learned lessons are captured as you work and are available in any terminal, IDE or agent. You speak naturally, in any language: “save this session”, “what did we decide about auth?”. There is no command to memorize.

get_state · update_state · store_decision · search_history · lesson_save — per project and per branch

How memory works

#02 Session Mesh

Your open sessions see each other.

Parallel sessions announce what they’re on and cooperate instead of colliding.

Technical details

Two Cursor tabs, a Claude Code terminal and an Antigravity session share one live bus. They announce what they are working on, claim the files they edit and hand work to each other.

session_announce · claim_path · session_send · session_wait — events delivered exactly once

See the MCP servers

#03 Guardian

Safety before every command.

Risky commands and writes are checked before they run, in the background.

Technical details

Guardian validates commands before they run, gates risky writes and keeps context from overflowing, without you invoking anything.

Caveat: coverage depends on each tool’s own hook support. The exceptions are documented in the project’s security assessment.

validate_command blocks shell injection · validate_write protects paths like ~/.ssh and /etc · validate_content flags prompt injection (advisory only)

Meet Guardian

#04 Token Crusher

Noise never reaches the model.

Long shell output is compressed by up to 90%, keeping errors and warnings.

Technical details

Before shell output reaches the model, Token Crusher compresses git logs, test spam, install noise and giant JSON while keeping every error and warning. Ask “how much did I save?” and the answer comes from your local ledger.

One 637 KB git log became 31 KB in a single crush · egc gain shows your own savings

Read about token optimization

04 Integrations

Works with the tools you already use.

egc install detects what is installed and sets up each tool with what it can read.

Model providers10

Bring your own model. Works natively with these providers, plus OpenRouter as a broker.

  • Claudenative
  • ChatGPTnative
  • Gemininative
  • DeepSeeknative
  • Mistral AInative
  • Ollamanative
  • Groqnative
  • Vertex AInative
  • Coherenative
  • OpenRouterbroker

05 Efficiency

EGC project estimates

87%fewer tokens to rebuild context every session.

  • Context overhead per session

    without EGC~1,500 tokens
    with EGC~200 tokens
  • Time re-explaining the project

    without EGC10 min/session
    with EGC0
  • Cost for 20 sessions per month

    without EGC~$0.08–0.09
    with EGC~$0.011–0.012
  • Shell output sent to the model

    without EGCfull output
    with EGCup to 90% smaller

Bars are proportional to the figures published by the project (ranges at their midpoint). Run egc gain to see the savings recorded on your machine.

06 Privacy & security

Your memory lives on your machine.

  • Local-firstNothing leaves your computer.The MCP servers talk over stdio only to your local AI tool.
  • EncryptionAES-256-GCM at rest.GCM authentication detects corruption on read.
  • ScopePer project and per branch.Switching branches restores that branch’s context.
  • RepositoryNever committed to git.Memory stays out of your source code.
  • LicenseOpen source, Apache-2.0.The engine stays free.

07 Install

Two commands. No step three.

  1. Install the package

    Requires Node.js 20 or later. Works on Windows, macOS and Linux.

  2. Run egc install

    Detects your tools, registers the local MCP servers and writes the memory protocol. It only asks whether you want the optional prompt library.

  3. Just talk

    Open your tool and say “let’s continue”. Memory loads on its own.

Linux, macOS and Windows. No git clone needed.

$ npm install -g @egchq/egc
$ egc install
# verify
$ egc doctor
egc-guardian running
egc-memory   running

08 FAQ

Frequently asked questions.

Didn’t find what you were looking for?
See the full FAQ

Does EGC send my project data anywhere?

No. All state is stored locally in ~/.egc. Nothing is sent to external servers: the MCP servers talk over stdio only to the AI tool on your machine.

Which tools are supported?

28 tools: Claude Code, Cursor, Codex CLI, Antigravity, OpenCode, Devin Desktop, Amp, VS Code Copilot, Zed, Kiro, Trae, CodeBuddy, JetBrains Junie, Goose, OpenHands, Aider, Warp, Cline, Qwen Code, Kimi Code CLI, Crush, Auggie, Mistral Vibe, Grok Build, Kilo Code CLI, MiMo Code, Pi and Droid. The installer auto-detects which ones you have.

Do I need to ask the AI to load context?

No. egc install writes the memory protocol into each tool’s global instruction file, and the AI is told to load state at the start of every session. For tool versions that don’t read that file automatically, you may need to add it to the session context manually.

How is this different from a CLAUDE.md or .cursorrules file?

Those files are static and committed to the repo. EGC state is dynamic and project-local: the AI writes to it every session, building up decisions and preferences. They complement each other.

Does it work with multiple projects?

Yes. State is stored per project and per git branch. Each project gets its own memory.

Is EGC free?

Yes. The engine is Apache-2.0 and stays free. If EGC ever offers something paid, it will be a team layer on top of the engine, never the memory on your machine.

Is the prompt library required?

No. It is optional and off by default: the installer asks, and the engine (memory and Guardian) works the same without it.

Same project. New session.

Install EGC
npm install -g @egchq/egc && egc install

EGC · Context That Never Resets