MCP Hub
Back to servers

antigravity-context-mcp

MCP server that gives AI agents persistent memory across sessions. Context recall, encrypted credentials, session notes, export/import — zero external requests.

Stars
1
Forks
1
Updated
Feb 19, 2026

Quick Install

npx -y antigravity-context-mcp

antigravity-context-mcp

You've been coding with your AI agent for two hours. The feature is almost done. You're in the flow.

Then the context window overflows. The session crashes.

You open a new chat. The agent looks at you like you've never met. "Hi! How can I help you today?"

Everything you built together — the decisions, the architecture, the twenty files you touched — gone. You start explaining from scratch. Again.

This MCP server gives Antigravity its memory back.

🇷🇺 Читать на русском


How it works

Session crashed. New chat. You type:

"continue"

The agent calls recall and gets the task checklist from the last session — what was done, what's in progress, where things stopped.

Need more detail?

"show me the details of that session"

The agent fetches the full context — walkthrough, plan, decisions made.

Want to look back at the week?

"show recent sessions"

You get a list with dates and titles. Pick one — the agent loads details only for that session.

Context loads progressively — the agent decides how much it needs. Not everything at once, but on demand, like a normal conversation.

Install

Paste the link to this repo into an Antigravity chat:

"install this MCP server: https://github.com/eglebov133-sudo/antigravity-context-mcp"

The agent will clone it, install dependencies, and configure everything. Restart Antigravity — done.

Manual installation
git clone https://github.com/eglebov133-sudo/antigravity-context-mcp.git
cd antigravity-context-mcp
npm install

Add to ~/.gemini/antigravity/mcp_config.json:

{
  "mcpServers": {
    "context": {
      "command": "node",
      "args": ["/path/to/antigravity-context-mcp/server.js"]
    }
  }
}

Restart Antigravity.

What it reads

The server taps into what Antigravity already stores but doesn't use between sessions:

  • Task checklists — what's done, what's in progress
  • Walkthroughs — summaries of completed work
  • Implementation plans — architectural decisions made

All local. No network. Just files on your disk that were always there.

Beyond context

Passwords between sessions. You know the drill — you give the agent your hosting password, the session dies, the next agent asks for it again. The server stores credentials encrypted with AES-256-GCM (key derived from your machine), auto-excluded from Git.

Persistent notes. Save code words, instructions, decisions with save_note — search them later with recall_notes.

Context snapshots. The agent can save an AGENT_CONTEXT.md into your project — persistent memory that survives any number of crashes.

Project list. Shows what Antigravity knows about, so you can switch projects without orientation time.

Diagnostics. context_status shows server health — sessions, notes, disk usage, encryption status.

What's new in v5.0

  • 📦 Export/import memorycontext_export dumps all notes to an encrypted file, context_import restores on another machine
  • 💾 Auto-backups — notes are automatically backed up before every modification (.backups/, 30-version retention with auto-pruning)
  • 🔐 Encrypted credentials — AES-256-GCM, machine-bound key, auto-migration from plaintext
  • 🛡️ Error handling — try/catch on every tool with clear error messages
  • 📏 Response limits — 50,000 char cap with [TRUNCATED] warning
  • 🩺 Diagnosticscontext_status for server health monitoring

Full changelog in CHANGELOG.md.

Security

  • Runs locally via stdio — no open ports, no network
  • Read-only by default — writes only when you explicitly ask
  • Credentials encrypted with AES-256-GCM, key bound to machine
  • Auto-migration from plaintext to encrypted
  • Zero telemetry, zero external requests

Requirements

  • Node.js 18+
  • Antigravity

License

MIT


Built because we got tired of introducing ourselves to our own AI agent.

Reviews

No reviews yet

Sign in to write a review