Shared living project context for teams building with Claude.
Every developer using Claude starts each session from scratch. In teams this compounds — each person re-explains the same project, makes decisions the rest of the team doesn't know about, and wastes tokens on warmup every single time.
agntx fixes this. Context lives on a server, not on your machine. Your team's decisions, architecture, and current state are injected into every Claude session automatically.
Developer A pulls context → does work → saves context
Developer B pulls context → sees Developer A's decisions → continues without re-explaining anything
agntx is a Model Context Protocol server that installs in seconds and ships via git — every teammate gets it automatically when they clone the repo.
# Install in your project
npx @agntxapp/agntx initThat's it. agntx detects your stack, pre-fills your CLAUDE.md with known gotchas and rules for your technology, generates slash commands, and wires up the MCP server.
Open Claude Code and type /status to pull your team's context into the session.
| Tool | When to use |
|---|---|
/status |
Start of every session — pulls context and confirms understanding |
/save |
End of session — Claude extracts decisions and pushes to the team |
/diff |
Before starting work — shows what changed since your last pull |
/pull |
Mid-session — re-sync context with a task description for large projects |
When you run npx agntx init, agntx detects your stack from package.json and pre-fills your context with community-maintained knowledge:
$ npx @agntxapp/agntx init
✓ Detected: Expo + Supabase
✓ Generated CLAUDE.md (47 lines pre-filled)
✓ Generated .claude/commands/save.md
✓ Generated .claude/commands/status.md
✓ Generated .claude/commands/eas-build.md
✓ Generated .claude/commands/rls-check.md
✓ Generated .mcp.json
✓ Configured Stop hook
Team token: agntx_xxxx
No template for your stack? The wizard interviews you instead — same output, no community knowledge pre-filled. You can contribute your answers back as a template.
Available templates
expo+supabase- More coming — contribute yours
Every /save captures not just what changed, but why — and what was rejected.
{
"decision": "Use Supabase AI for embeddings",
"reason": "Zero external dependencies, built into Edge Functions",
"rejected_alternatives": ["OpenAI ada-002 — adds OpenAI dependency"]
}This is append-only. Nothing is ever overwritten. Six months later, when someone asks "why don't we use OpenAI here?" — the answer is already in the log.
Two developers save at the same time? agntx surfaces the conflict instead of silently overwriting:
⚠️ Conflict in current_state.next_priority:
[Yours] "Implement auth flow"
[Theirs] "Fix the onboarding bug first"
Run: /resolve A | /resolve B | /resolve both
Resolution is logged to the decisions history automatically.
.mcp.json is committed to git. Every teammate who clones the repo gets the MCP server automatically — Claude Code loads it on every session. No per-developer setup required beyond npx agntx init once.
The fastest way to get started is the hosted service at agntx.app — no infrastructure to manage.
npx @agntxapp/agntx init connects to the hosted backend by default.
Want to run agntx on your own infrastructure? See backend/README.md for the schema and reference Edge Function implementations.
The hosted service includes production features not in the reference implementation: atomic conflict detection, billing enforcement, rate limiting, and pgvector similarity search for large projects.
Templates live in templates/combos/. Each one is a JSON file with stack-specific rules, gotchas, and architecture patterns.
To contribute:
- Fork the repo
- Copy an existing template as a starting point
- Add your stack's real gotchas — the things that actually bit you
- Open a PR
Good templates come from real experience. If you've been burned by a Supabase RLS policy or a React Native Metro cache issue, that belongs in a template.
Apache 2.0 — see LICENSE
Copyright 2026 Anas Almohana