On Board

Shared project memory for agents. One MCP server, one project memory folder, many IDEs and agent clients. Skill-first setup, one shared runtime, and project-scoped configuration.

Related MCP server: ForgeSwarm

What this is

On Board is a local MCP server for coordinating AI agents across a project. It gives Claude Desktop, Claude Code, Codex, Cursor, Antigravity, and other MCP clients the same project memory, ticket queue, and handoff history.

The goal is simple: when one agent stops and another agent continues, the next agent should not need the human to explain the project again.

onboard → read memory → claim work → write progress → hand off

Everything stays local to the project unless you choose to connect other tools.

Quick start (skill-first)

Install the onboard skill from inside the project you want to configure:

# Node.js 22.20 or later; select your agent when prompted.
npx skills add swisspra/On_Board --skill onboard

npx runs the skills CLI from npm; the skill itself comes from Git. No separate On Board npm package is needed. The skill guides runtime installation and native client configuration; installing the skill alone does not start an MCP server, upgrade an existing shared runtime, or initialize memory.

Native resources require a verified 4.2+ runtime. A working 4.1.1 or source installation can stay in place; see runtime upgrades or the optional source runtime.

Then send your agent this prompt:

Using /onboard, set up On Board for this project. Keep existing memory and other MCP servers unchanged.

The /onboard skill reuses one shared onboard-memory-mcp runtime and configures one project — the one you run it in — for one client of your choice. It resolves an existing runtime first and installs the shared runtime only when none is present (reuse, not a per-project install). It reads and writes only that project's own MCP config. It does not initialize or migrate .agent-mem/, and it does not touch other MCP servers, your instructions, or hooks.

What happens, in order:

1. Reload your agent after the install so it discovers the new /onboard skill (reopen the window, or start a new session).

2. The skill resolves or installs the shared runtime and writes the project's MCP config.

3. Trust and restart your MCP client so it loads the new agent-memory server. Some clients ask before running a newly added server; approve the prompt, then restart.

Prerequisites are real, not hidden behind a one-command promise:

• Node.js ≥ 22.20 to run npx skills.

• Setup helper: Python ≥ 3.11. The skill's scripts/setup.py needs it; where the machine's python3 is older, run the one-time helper with uv run --no-project --python 3.12 scripts/setup.py ….

• Runtime: Python ≥ 3.10, plus uv or pipx already installed, for the onboard-memory-mcp runtime the skill installs.

Runtime states, so you know where you are: Runtime available (the runtime is resolved or installed) → Configured (the project's MCP config now points at it) → Awaiting trust or restart (the client has not loaded it yet) → Ready (the client shows the agent-memory tools). Setup ends at Ready; initializing or joining a board is a separate step, started only when you explicitly ask the agent to begin work.

Prefer to wire things by hand, or install from a package manager or a clone? Those paths stay fully supported — see Manual installation and docs/SETUP.md.

🚀 v4.0.0 — agents now wake each other

Until v4 this board was pull-only: an agent found out about new work when a human told it to look. v4 ships memory_wait_for_event — an agent parks inside one tool call and wakes the moment a peer creates a ticket, submits work, or delivers a verdict.

This is not a demo claim. In the launch trial, a GPT agent (Codex) and a Claude agent closed a full reject → fix → resubmit cycle on this board with zero human relay — the reviewer's fix instructions travelled inside the wake payload, the worker re-read the file, attached a sha256, and resubmitted; the reviewer reproduced the hash byte-for-byte before approving:

Full mechanics in Agent-to-agent: the listening half · breaking changes in CHANGELOG.md · release notes.

Why this exists

Most agent workflows break for boring reasons:

• The next chat does not know what the last chat did.

• Parallel agents overwrite or redo each other's work.

• Important decisions live only in conversation history.

• Handoffs are informal, so review and follow-up work drift.

On Board keeps those facts in project-local files under .agent-mem/. The MCP tools expose that memory to any supported client.

Who this is for

• Solo developers using more than one agent or IDE

• Teams experimenting with multi-agent coding workflows

• Projects where handoffs, tickets, and review notes matter

• Local-first MCP users who want shared context without a hosted service

It is not an autonomous project manager. Humans still decide what matters, review important changes, and accept the final result.

Manual installation

These paths stay supported; they are just secondary to the skill-first quick start above. Use them to install from a package manager, or to run everything from a clone by hand.

Install the server

The server is published as onboard-memory-mcp. Install it with whichever tool you prefer:

# Homebrew (tap once, then the short name works: brew install onboard-memory)
brew install swisspra/tap/onboard-memory

# pipx
pipx install onboard-memory-mcp

# uv
uv tool install --python 3.12 onboard-memory-mcp

All three provide the onboard-memory-mcp command (Homebrew also adds a short onboard-memory alias). Homebrew covers macOS and Linux; on Windows use pipx or uv (the command is onboard-memory-mcp.exe). Point your MCP client's command at it instead of python3 onboard_server.py. You can also skip this and run from a clone using the setup paths below. (On Homebrew 6+, approve the one-time tap-trust prompt, or run brew trust swisspra/tap.)

The v4.2.0 runtime uses mcp[cli]>=2.3.0,<3 and the SDK's MCPServer API. The lockfile selects MCP 2.3.0; SDK 1.x is no longer supported. Minimum AnyIO, PyJWT, and cryptography versions also prevent installers from retaining known-vulnerable transitive dependencies.

Headless config (no clone)

With the server installed, wire your MCP client to it directly — no repo checkout, no setup-project.sh:

{
  "mcpServers": {
    "agent-memory": {
      "command": "onboard-memory-mcp",
      "env": { "AGENT_PROJECT_DIR": "/full/path/to/your/project" }
    }
  }
}

• CLI clients (Claude Code, Codex) inherit your shell PATH, so the bare onboard-memory-mcp works.

• GUI clients (Claude Desktop, Cursor) launch with a minimal PATH. Use the absolute path from which onboard-memory-mcp (where on Windows) as command — typically /opt/homebrew/bin/onboard-memory-mcp (Homebrew, Apple Silicon), /usr/local/bin/onboard-memory-mcp (Homebrew, Intel), /home/linuxbrew/.linuxbrew/bin/onboard-memory-mcp (Homebrew, Linux), ~/.local/bin/onboard-memory-mcp (pipx / uv on macOS/Linux), or %USERPROFILE%\.local\bin\onboard-memory-mcp.exe (pipx / uv on Windows).

AGENT_PROJECT_DIR is required — it decides which project owns .agent-mem/. Nothing here initializes memory: starting work is a separate step. In your first work session, call memory_init once (creates .agent-mem/) — or memory_bootstrap when adopting an existing repo — then memory_onboard each session. If .agent-mem/ already exists, never bootstrap it again; just join.

The pipx/uv path normally uses prebuilt dependency wheels on supported CPython 3.11+ platforms. Wheel availability depends on Python and architecture; Python 3.10 or Intel macOS may need a C/Rust toolchain. Homebrew bottle builds target ARM macOS and x86_64 Linux. Upstream dependency bottles are currently unavailable for Intel macOS, so Homebrew is not a compiler-free fallback there. Template: configs/binary-mcp.json; full detail and platform notes in docs/SETUP.md.

Set up a project

Choose one setup path:

Option 1: Agent setup

The skill-first quick start above is the recommended path. If you would rather not use the skill CLI, ask an agent to read AGENT_SETUP.md — it delegates to the canonical install reference instead of inventing a second recipe — and to docs/SETUP.md for the manual steps.

Option 2: Script setup

git clone https://github.com/swisspra/On_Board.git
cd On_Board
bash setup-project.sh /full/path/to/your/project
bash doctor.sh /full/path/to/your/project

Add the generated MCP config to your client:

/full/path/to/your/project/.onboard/mcp.generated.json

Some clients accept this JSON directly. Others require you to merge it into their own MCP settings file.

After memory is initialized, open the dashboard with:

bash /full/path/to/your/project/.onboard/run-dashboard.sh

On Board is installed once. Each project points to the same On Board folder, but gets separate memory through AGENT_PROJECT_DIR.

Each setup-project.sh run also registers the project locally in .onboard/linked-projects.json inside the On Board checkout. This file is gitignored and only helps updates remember which projects point here.

The setup script uses uv sync --inexact to install/update dependencies without pruning local test/dev extras. MCP clients run python3 onboard_server.py; the launcher uses the local .venv directly and rebuilds it only if the venv is missing. This keeps normal startup fast, avoids uv run startup timeouts, and makes a shared central checkout more durable.

On Board does not write memory from end-turn hooks. Current Stop hooks in several agent clients run every turn, which creates noisy memory and can force agents to re-onboard too often.

Optional: add AGENT_MEM_CONTEXT_DIRS to the generated MCP config when agents should read shared docs/specs outside the project folder.

Option 3: Advanced manual setup

If you do not want to run the setup script, install with uv sync, write the MCP config yourself, and add project rules/hooks manually. See docs/SETUP.md.

Setup stops at the config; nothing above initializes memory. When you later start work in the project — a separate, explicit request — do it in a chat with any MCP-aware agent (Claude Desktop, Claude Code, Cursor, Codex, Antigravity). Skip memory_bootstrap if .agent-mem/ already exists and just join:

memory_bootstrap(
  agent_name="dev-main",
  description="Existing project using On Board",
  current_task="Set up shared project memory"
)

memory_onboard(
  agent_name="dev-main",
  agent_platform="claude-code",
  agent_role="main"
)

That's it. The agent now sees the project briefing, the open tickets, the recent memory, and the protocol it should follow. Every subsequent action is stamped with its identity.

Full setup details and manual setup: see docs/SETUP.md.

To update an existing install, run bash update.sh in the central On Board checkout. It will show known linked projects. Refresh all of them with bash update.sh --refresh-linked, or inspect them with bash setup-project.sh --list-linked.

The loop in one example

1. SPEC
   opus-testcase reads requirement → writes 5–20 acceptance tickets
   with explicit pre/post conditions.

2. BUILD
   dev-track-2 claims a ticket → implements in src/ → submits with
   file diff + test plan.

3. TEST
   Jonhny-tester picks up submission → runs UI in Chromium → captures
   screenshots → submits PASS or FAIL with evidence.

4. REVIEW
   desktop-opus4.7 (or the human) checks evidence → approves OR rejects
   with concrete fix instructions.

   If rejected → ticket reopens → dev-track-2 patches → Jonhny retests
   → loop closes.

When this loop runs cleanly, a single ticket goes from open to "shipped to production" in 4–15 minutes of agent time. The human checks in at the end, not in the middle.

Agent-to-agent: the listening half (v4)

Everything above still works pull-style. v4 adds the missing edge: agents can now wake each other instead of waiting for a human to relay messages.

worker:  memory_wait_for_event(agent_name="dev-track-2", timeout_s=180)
         → parks inside one tool call until the board changes
lead:    memory_create_ticket(..., assigned_to="dev-track-2")
worker:  wakes in seconds, claims, works,
         memory_submit_ticket(..., stay_active=True)
lead:    wakes on the submission, reviews
worker:  wakes on the verdict — approve closes the loop;
         a rejection arrives WITH the review notes and fix
         instructions in the wake payload, so it re-claims,
         fixes, and resubmits without asking anyone

Design points, all field-verified across Claude Desktop × Claude Desktop and Claude × Codex (GPT):

• Check before blocking — a re-arm after a gap returns its backlog in 0 s instead of waking empty. One wake drains the whole queue.

• Loop guard — an agent never wakes on its own actions, so two listeners cannot ping-pong each other.

• Role gate — completed ≠ success: whoever executed a ticket may reach submitted but may never close it; only the owner or a main/lead/reviewer adjudicates. Solo use is still possible via explicit allow_self_review=True, permanently stamped in the audit.

• Client limits respected — Claude Desktop cancels tool calls at ~240 s per call (measured), so timeouts clamp to 200 s there; stdio clients (Claude Code, Codex) may pass long_wait and park much longer.

• Idle budget, in minutes — the server counts consecutive empty parks and answers STAND-DOWN once idle_budget_min (default 15) is spent, so an unattended listener stops on its own instead of looking wedged. Budgets are stated in minutes because a human watching a silent loop counts wall clock, not iterations — a compliant agent looping for 20 minutes looks stuck even when it is exactly on budget. Every idle reply prints idle 3/5 — ~6 min to stand-down. The counter resets on a real event and never on re-arming, and STAND-DOWN is a distinct status so a loop matching on idle cannot read it as permission to continue. idle_budget_min=0 listens indefinitely.

• Use the listen MCP prompt for the standard re-arm loop.

v4 also hardens the board for simultaneous writers (advisory lock on ticket mutations, per-process tmp files), because with A2A two agents acting in the same instant is the normal case, not the rare one. Breaking changes and the migration guide live in CHANGELOG.md.

Tools (30 MCP tools, 5 buckets)

Full reference: docs/TOOLS.md.

Native resources (v4.2.0)

Three read-only JSON resources are exposed for clients that read MCP resources: onboard://project, onboard://tickets, and onboard://memory/summary. They return bounded projections (no raw memory bodies) and never create or modify a board.

Every resource read is live state with a private ttlMs 0 hint; only the static response catalogs use ttlMs 60000. Clients that support resource subscriptions and honor them get a coalesced ResourceUpdated invalidation notice when any of the three resources change. That is a re-read cue — not durable delivery and not a model wake-up; there is no replay. Use memory_wait_for_event when an agent must actually be woken by a peer.

On Board ships no MCP App (ui://) view this round: a host-rendered view is not a dependable wake mechanism, and not every host — CLI clients in particular — renders apps. See docs/TOOLS.md.

What makes this different

On Board is not only a place to store memories. It keeps the work loop visible:

onboard -> claim ticket -> submit evidence -> review -> approve or reopen

That gives agents a shared queue, stable identities, recent handoffs, and a review gate. Rejected work reopens with fix instructions instead of becoming a dead terminal state.

Project structure (runtime data)

your-project/
├── .agent-mem/                runtime memory, gitignored
│   ├── project.json
│   ├── agents.json            agent registry (identity, status, KIA)
│   ├── memories.json
│   ├── state.json             project phase, owner, design defaults
│   ├── archive.json
│   ├── digests.json
│   ├── checkpoints/
│   └── tickets/
│       ├── _index.json
│       ├── TK-<id>.md                      the spec
│       ├── review/TK-<id>-submit.md        dev submission
│       ├── closed/                         approved work
│       └── rejected/<ticket_id>/attempt-<N>/
│           ├── <ticket_id>.md              rejected spec for that round
│           ├── <ticket_id>-submit.md       submission under review
│           └── <ticket_id>-rejected.md     reviewer notes + fix instructions

Everything is plain text or JSON. You can cat your way through the project's full history. No vector DB lock-in, no opaque embeddings — just files an audit can read.

Current status (v4.2.0)

The primary skill-first route reuses one shared onboard-memory-mcp runtime — installed once and reused by every configured project, each with its own project-selected memory folder. A central On Board checkout stays a supported secondary path:

• The skill-first command above is the primary installation entry; a project is configured with the /onboard skill for one chosen client, reusing an existing runtime and installing the shared runtime only when none is present. Memory is never initialized or migrated by the installer.

• Three read-only native resources (onboard://project, onboard://tickets, onboard://memory/summary) expose bounded project, ticket, and memory projections. All three reads carry a private ttlMs 0 hint (only the static catalogs use ttlMs 60000); subscription notifications are a capability-dependent re-read cue, not a model wake-up. They exist only on a verified v4.2+ runtime.

• memory_onboard is the primary start call for agents and returns compact current context.

• memory_wait_for_event turns the board push-capable: agents park, wake on peer actions, and close reject/retry loops with zero human relay (see the A2A section).

• memory_doctor checks board health and data integrity — a diagnostic, not a setup gate.

• The secondary source route runs setup-project.sh, which generates project MCP config, rules, startup hooks, and a dashboard launcher.

• Linked-project registry tracks which projects point at the central checkout, so updates can refresh known projects without scanning the machine.

• In that source route, runtime startup uses python3 onboard_server.py; the launcher normally execs .venv/bin/python server.py and only falls back to uv sync --inexact if .venv is missing. The primary route points clients at the installed onboard-memory-mcp executable instead of the source launcher.

• Existing working 4.1.1 or source-based bindings keep working unchanged; the skill asks before replacing a shared runtime, and it never upgrades one unsolicited.

• Startup hooks return a small read-only briefing. End-turn/Stop hooks are not installed by default because current clients can run them too often.

• The dashboard is local and read-only.

Full CHANGELOG: CHANGELOG.md.

License

Apache-2.0. Free to use, fork, modify, redistribute, build commercial products on. No restrictions on use.