Axis: one coordination board for your whole team's agents

Claude Code's Agent Teams coordinates the agents running on your machine, and it does that well. Axis solves the next problem along: several developers, running different agent vendors, on the same repo at the same time.

Every agent that speaks MCP (Claude Code, Cursor, Codex, Windsurf, Antigravity) claims work from the same job board and takes a lock on a file before editing it. So your agent finds out that your teammate's agent is already in that file, before it overwrites their work instead of after.

If you are one developer on one machine, use Agent Teams. It is free, native, and better at that. Axis is for when the collisions stop being yours.

See it in 30 seconds

bun examples/two-agent-collision.ts

Runs fully offline against a throwaway repo in your temp directory. Two agents go for the same file; every status below is a real return value, not staged output:

2. Dana's agent (Claude Code) claims the top job and takes the file
   claim_next_job      -> CLAIMED  refactor auth to issue JWTs
   propose_file_access -> GRANTED  src/auth.ts

3. Sam's agent (Cursor, different machine) goes for the same file
   propose_file_access -> REQUIRES_ORCHESTRATION

File 'src/auth.ts' is locked by 'dana-claude-code' for: "refactor auth to issue
JWTs instead of session cookies". Pick a different file or job, or coordinate via
update_shared_context. The lock auto-expires after 30 min; use force_unlock only
if 'dana-claude-code' has crashed.

4. So it takes the other job instead of colliding
   claim_next_job      -> CLAIMED  add rate limiting to the login route

That denial is the whole product. Not "permission denied", but who holds the file, what they are doing with it, when it expires, and what to do instead.

Related MCP server: ACDP

"Why not just use git worktrees?"

Worktrees are the right tool and we use them. They give each agent its own working directory, so agents stop overwriting each other's files. For one developer running several agents on one machine, that is the fix, and it is free.

They do not reach across machines. Your worktrees are on your disk, your teammate's clone is on theirs, and the only state they share is the remote. Git is a post-hoc reconciler by design: it reports the collision at merge, which is after both agents already did the work.

They also isolate files rather than intent. Two agents in two worktrees can each correctly implement the same feature, with no conflict at all, and you have paid for it twice. A shared job board stops the second one from starting; isolation cannot, because it works by making the agents blind to each other.

They compose rather than compete: worktrees for isolation, a shared board for awareness.

Open-core

This repository is the free, open-source orchestration core of Axis (AGPL-3.0): the MCP server, the axis CLI, the Python SDK, and the agent protocol — everything your agents use to coordinate.

• Free forever: the orchestration layer — job board, file locking, shared notepad, sessions, project soul. It's just coordination state; it costs nothing to run.

• Paid (hosted): the intelligence layer — hosted search (vector + full-text + trigram, fused and LLM-reranked), cited multi-hop deep_search, and incremental indexing — plus the managed backend, dashboard, and billing. The local server's search_codebase is ripgrep plus a keyword ranker (no index needed); a Pro key blends hosted results in when they arrive within budget. That lives at useaxis.dev (closed source) because it carries real embedding/LLM cost.

Quickest start: sign up at useaxis.dev, then point your MCP client at https://useaxis.dev/api/mcp and authenticate — no key to paste. See agent-instructions/mcp-setup.md for OAuth, API-key, and local-install options.

Features

1. Shared job board: Post work, claim it atomically. Claims run through SELECT ... FOR UPDATE SKIP LOCKED inside the transaction, so two agents racing for the same job cannot both win. Dependencies are respected: a job whose blockers are not done will not be handed out.

2. Per-file locks, tamper-evident: An agent takes a lock before editing. Locks are advisory by nature (a coordination server cannot block a write it does not perform), so Axis records a content fingerprint when the lock is granted and lets the holder verify before writing. You find out that a file changed under you, instead of silently clobbering it.

3. One board across people and vendors: Commit an org pin in .axis/axis.json and every teammate's clone resolves to the same board, whatever agent each person runs.

4. Live board and shared notepad: See what every agent on the repo is doing right now at useaxis.dev/team/board, pushed over Postgres Realtime.

5. MCP native: Standard protocol and OAuth, so there is no key to paste and no plugin to install per client.

Environment

Create a .env.local file for local development (see .env.local.example):

SHARED_CONTEXT_API_URL=http://localhost:3000
SHARED_CONTEXT_API_SECRET=your_shared_secret
OPENAI_API_KEY=your_openai_key
SUPABASE_URL=your_supabase_url
SUPABASE_SERVICE_ROLE_KEY=your_supabase_service_role_key

Do NOT set PROJECT_NAME unless you deliberately want to override detection: it outranks repo detection, so a fixed value (like default) collapses every repo on the machine onto one shared job board. Left unset, the project name is derived from your repo (committed .axis/axis.json "project", else the repo folder name), so every clone of the same repo resolves the same board.

Team coordination (shared boards)

Commit an org pin so every teammate's clone lands on the same board:

// .axis/axis.json (at the repo root)
{ "project": "your-repo", "org": "<your org id from useaxis.dev/team>" }

Per-machine override: AXIS_ORG_ID=<org id>. Without an org, coordination scopes to your personal workspace (solo behavior, unchanged). Watch and drive the live board at useaxis.dev/team/board.

Setup

1. Install Dependencies:

   bun install

2. Initialize Context:

   bun cli init

3. Start MCP Server: To run the server locally for testing/connection:

   bun start:local

This stdio server exposes the full Nerve Center toolset (job board, locks, notepad). Running from a checkout is for contributing to Axis itself — users connect to the hosted server instead (next paragraph).

*Address for MCP Clients*: point your agent at the hosted MCP server — no
install, no updates to manage:
```
https://useaxis.dev/api/mcp
```
Authenticate via OAuth or a Bearer API key; see
[agent-instructions/mcp-setup.md](agent-instructions/mcp-setup.md).

Axis derives project identity from the active repository. Supported agent
hosts can provide `AXIS_WORKSPACE_ROOT`, `SUPERSET_WORKSPACE_PATH`, or
`SUPERSET_ROOT_PATH`; these per-session values override a stale positional
root in global MCP config so project souls, jobs, and locks cannot leak
across repository switches. Set `AXIS_PROJECT_NAME` only when a project
name must intentionally remain fixed across workspace changes.

Workspace switching is automatic: every tool call re-resolves the workspace from the runtime hints and from any absolute file path in the call's arguments. When either points at a different repository, the server rebinds itself in-process and notes the switch in the tool response — no restart, no stale "default" board. switch_project remains available for explicit switches.

4. CLI Usage:

   # Add an entry to activity.md
   bun cli add-context "Refactored the API to use Hono"

Parallelism Philosophy

The key to Axis is the Parallel Sprints. You no longer have to manage a single agent sequentially; instead, you orchestrate a swarm.

1. Define the Objective: Tell any agent (the "Manager"): "Build the Authentication System."

2. Autonomous Partitioning: The agent decomposes the objective into jobs (API, UI, Tests) and posts them to the Distributed Job Board.

3. Horizontal Scaling: You open Cursor, Claude Code, and Antigravity. They all instantly "claim" the next available job on the board.

4. Synchronized Execution: While agents work in parallel, they stay in sync via the Live Notepad, ensuring that if one agent changes an API signature, the others adjust their code in real-time.

Connecting Agents (MCP)

Point your IDE (Claude Desktop, Cursor, etc.) at the hosted MCP server:

{
  "mcpServers": {
    "axis": {
      "url": "https://useaxis.dev/api/mcp"
    }
  }
}

Authenticate via OAuth (no key to manage) or a Bearer API key — see agent-instructions/mcp-setup.md. New tools land server-side, so there is nothing to update on your machine.

MCP Tooling

The server exposes 28 tools to agents. src/shared/tool-manifest.ts is the canonical list and is what the test suite checks the server against; this section is a readable summary of it.

Coordination (free, open-core):

• get_project_soul — load project context, goals, and conventions

• update_project_soul — write or refresh the project soul

• post_job — add a job to the distributed Job Board

• list_jobs — inspect status, priority, ownership, and dependencies

• claim_job — atomically claim a specific ticket

• claim_next_job — atomically claim the next available job

• complete_job — report a job outcome and release its file locks

• cancel_job — withdraw a posted job

• propose_file_access — pessimistically lock files before editing; pass filePaths to lock a multi-file batch in one call (all-or-nothing)

• list_locks — inspect active file ownership and intent

• verify_file_lock — confirm a locked file wasn't changed under you before overwriting (tamper check)

• guarded_write — enforced write: the server writes only if you hold the lock and the file is unchanged (rejects clobbers)

• release_file_access — release an owned lock early

• list_agents — see which agents are active or idle on the project (visible before jobs are posted)

• switch_project — rebind a live MCP session to another workspace without reconnecting

• force_unlock — admin override for stale locks from crashed agents

• update_shared_context — append to the Live Notepad. Coordination tool responses also carry an ambient "team activity" trailer: whatever other agents logged since your last call, so nobody has to remember to re-read the notepad. agentId is optional on every tool — it defaults to the session's unique identity.

• release_job — put a claimed job back on the board for someone else

• finalize_session — archive the session and clear all remaining locks

Shared context and search:

• read_context — read the project's shared context

• update_context — write to the project's shared context

• get_shared_context — fetch the current Live Notepad

• search_codebase — semantic search over the indexed repository

• search_docs — search indexed documentation

• index_codebase — build the search index for a repository

• index_file — index a single file, for incremental updates

Account:

• get_subscription_status — current plan and entitlements

• get_usage_stats — request counts and limits

Universal Session History

Axis records every Axis MCP tool call and result at the protocol boundary. This works with any MCP client, including Cursor, Windsurf, GitHub Copilot, Antigravity, Claude Code, Codex, Gemini CLI, Cline, Roo Code, Continue, Aider, and clients Axis has never seen before. These events are always available in the archived session even when the host does not expose its private chat transcript.

Full user/assistant chat is added when the host exposes a transcript. Codex and Claude Code are detected automatically. Any other client can provide a JSON or JSONL export using:

{
  "env": {
    "AXIS_TRANSCRIPT_PATH": "/absolute/path/to/session.jsonl",
    "AXIS_TRANSCRIPT_FORMAT": "generic",
    "AXIS_AGENT_BASE": "github-copilot"
  }
}

The generic adapter accepts common role/content, messages, tool_calls, tool_call, and tool_result shapes. Set AXIS_TRANSCRIPT_FORMAT to codex, claude, or generic. MCP cannot access chat text a host keeps private; in that case Axis still captures the complete Axis tool timeline.

Read-only lock hardening (opt-in). Locks are advisory coordination by design. Set AXIS_ENFORCE_LOCKS=1 to harden them: on grant the server chmods the locked file read-only, so any process — including an agent that ignores Axis — gets EACCES on write. The holder writes through guarded_write (which briefly restores perms); release/complete_job/finalize_session restore the original mode. This changes editing ergonomics (you must write through Axis while a file is locked), so it's off by default. It stops cooperating tools and accidental clobbering; a process running as the same user can still chmod back, which no userspace server can prevent.

Intelligence (hosted, paid):

• index_codebase — build the searchable index for a project

• index_file — index a single file (reads from disk if content is omitted)

• search_codebase — hosted: vector + full-text + trigram retrieval, fused and LLM-reranked, with related files + definitions enrichment (locally this tool answers from ripgrep + keyword ranking)

• search_docs — search indexed documentation

Account:

• get_subscription_status — current plan and entitlements

• get_usage_stats — usage against plan limits

Agent Integration Examples

Claude Desktop (example flow)

1. claim_next_job with your agentId.

2. If claimed, propose_file_access before edits.

3. After completing work, complete_job with outcome notes.

4. Use update_shared_context to summarize decisions.

Cursor (example flow)

1. get_project_soul to load context.

2. claim_next_job or post_job for new work.

3. propose_file_access before editing files.

4. finalize_session at the end of a sprint.

Troubleshooting

• Permissions: Ensure chmod +x src/local/mcp-server.ts or that bun is in your PATH.

• Directories: On first run, the system will auto-create history/ and agent-instructions/. Ensure write permissions.

• Locking Issues: If a file is permanently locked due to a crash, use the force_unlock tool via any agent or delete history/nerve-center-state.json.

Architecture

• Local MCP server (src/local/mcp-server.ts): The canonical stdio server, shipped as @virsanghavi/axis-server. Its tool surface is defined once in src/shared/tool-registry.ts.

• Hosted MCP + API: The paid surface lives in the axis-frontend repo — MCP at https://useaxis.dev/api/mcp, REST at https://useaxis.dev/api/v1.

• Parallel Job Board: Supabase-backed registry for atomic task distribution.

• Distributed Memory: Real-time vector-indexed persistence of agent decisions.

Production & Deployment

Supabase Setup

Run supabase db reset (Docker required) to build the database from the numbered migrations in supabase/migrations — they are the source of truth. supabase/schema.sql is a generated snapshot for reference only; never apply it by hand. See supabase/README.md for the workflow.

RAG API Examples

The RAG endpoints are served by the hosted API (axis-frontend repo, deployed at useaxis.dev).

Embed content

curl -X POST https://useaxis.dev/api/v1/embed \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $AXIS_API_KEY" \
  -d '{
    "items": [
      {
        "content": "This repo uses Bun",
        "metadata": { "filename": "context.md", "source": "agent-instructions" }
      }
    ]
  }'

Search content

curl -X POST https://useaxis.dev/api/v1/search \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $AXIS_API_KEY" \
  -d '{
    "query": "What runtime does this project use?",
    "limit": 5,
    "threshold": 0.5
  }'

Testing

We maintain a suite of Unit and Load tests.

# Run Unit Tests
bun test

# Run Load/Concurrency Verification
bun tests/load-test.ts

Security & Robustness

• Rate Limiting: Enforced by the hosted API (axis-frontend) in front of every /api/v1 route.

• Persistence: State is saved to history/nerve-center-state.json to survive restarts.

• Concurrency: AsyncMutex ensures atomic operations on the Job Board and File Locks.

Feature Status

• RAG / Smart Retrieval: Implemented via /embed and /search in the API.

• Job Board: Implemented in the Nerve Center with optional Supabase persistence.

• File Locking: Implemented with stale-lock cleanup and admin force unlock.