ClaudeMap

pulsar

An MCP server that puts every AI coding agent on your machine on one bus — Claude Code, Codex, OpenCode, pi, omp, Cursor and Gemini CLI can list each other, tag themselves and message one agent, a group or all, waking idle sessions

MCP Serversmcp servermulti-agentagent messagingclaude codecodexopencodenpm

pulsar is an MCP server that turns the coding agents on one machine into a reachable roster: every live session shows up on one list (including Claude Code and OpenCode sessions that never loaded pulsar — the daemon scans for them), annotated with harness, model, project, branch, worktree and PR (found offline from local git refs and session logs), plus tags agents add themselves like role:reviewer or issue:1378. Messages are push, not polling: send to a name, a selector (project=apparish, model=fable, pr=1402 role=reviewer, ~text, or * for a broadcast), and an idle agent is woken; urgent messages land mid-turn where the harness supports it. reply keeps threads, ask blocks until the answer comes, and wait replaces sleep loops. Delivery runs through one small per-machine daemon (started on demand, stops when idle) that keeps the registry and a SQLite message log in ~/.local/state/pulsar and pushes over each harness's own channel — Claude Code's per-session inbox socket (the one SendMessage uses), codex queue --resume, OpenCode's local API, or the pulsar extension for pi and omp. Safety is designed in: messages are stamped as coming from a peer agent and can never approve anything, each agent can accept, hold or refuse, the socket sits in a 0700 folder, limits cap messages at 64 KB, 60 sends and 6 broadcasts a minute and 50 unread per inbox, and there is no telemetry — the only network call is pulsar doctor's npm version check. Tools: list_agents, send, reply, ask, wait, inbox, set_identity, whoami/history, plus a short skill that teaches agents when to use them.

What it does

  • One machine-wide roster: every live agent from any harness, including Claude Code and OpenCode sessions that never loaded pulsar, listed with harness, model, project, branch, worktree and PR
  • Selectors and tags: address one agent, a group (project=x, role:reviewer, ~text) or * for everyone; agents tag themselves and post activity lines like 'reviewing PR #1402'
  • Push delivery that wakes idle agents, with urgent messages landing mid-turn where the harness supports it — Claude Code via its per-session inbox socket, Codex via codex queue --resume, OpenCode via its local API
  • Request/reply primitives: reply keeps threads, ask blocks until the answer comes, wait blocks on a message, a reply or another agent going idle
  • Safe by default: messages are stamped as peer-sent and can never approve anything; each agent accepts, holds or refuses; 0700 socket; limits of 64 KB per message, 60 sends and 6 broadcasts a minute, 50 unread per inbox; no telemetry
  • Install once for all agents: npx pulsar setup configures every harness it finds (backs up files, --uninstall to undo), or per-harness routes like the Claude Code plugin marketplace

When to use it

Use when several coding agents share one machine and you want them to hand work to each other — ping the reviewer agent from your build-fix session, broadcast 'main is frozen' to every session on a project, or have an orchestrator wait for a worker to go idle — without bypassing Claude Code's own guardrails.

Getting started

Everything in one command:

npx pulsar setup

Setup finds your agents, shows exactly which files it will change and asks once; it backs up every file it touches. Useful flags: --dry-run (show the plan, change nothing), --client claude,codex (only these agents), --yes --all-detected (no questions, for dotfiles), --uninstall (undo everything). Prefer a package manager? Claude Code:

/plugin install pulsar --marketplace yocontra/pulsar

The plugin path cannot set Claude's crossSessionInbound approval, which pulsar setup does (claude mcp add-json pulsar --scope user, crossSessionInbound: "accept" in ~/.claude/settings.json, plus the pulsar skill in ~/.claude/skills/pulsar) — run setup once if you want messages to reach sessions that skip permission prompts. No Node on the machine? curl -fsSL https://raw.githubusercontent.com/yocontra/pulsar/main/install.sh | bash installs a private Node and runs the same setup. Restart your agents afterwards.

Things to know

One machine only in v1: cross-machine messaging and remote control over Tailscale are listed as coming soon, not shipped — the README's own comparison table marks pulsar 'one machine (v1)' against Claude Code's local + Remote Control. Setup modifies each harness's config files (with backups) and adds a skill; read its plan before confirming, and prefer --dry-run on scripted machines. Agents behind harnesses pulsar cannot push into get an inbox plus an 'N unread' note on tool results instead of a wake. The trust boundary is your user account: the 0700 socket is reachable by any process running as you — treat it like your files, not a sandbox. The delivered text always says it comes from a peer agent that cannot grant permissions, and Claude's own guardrails still apply on top. MIT. 38 stars as of writing (2026-11-18).

Repository

yocontra/pulsar

Charted

Related on the map

  • Blitz Strike

    MCP Server

    A penetration-testing methodology delivered as one MCP server: BLITZ maps the attack surface, EAGLE-EYE traces source-to-sink, STRIKE live-verifies every finding before it is reported.

    • mcp-server
    • security
    • pentest
    • methodology
  • Ripwire

    MCP Server

    Red Hat's "ripgrep for AI context" — a C++23 zero-dependency CLI that also speaks MCP.

    • mcp
    • code-context
    • cli
    • red-hat
  • Deja Vu

    MCP Server

    One shared session memory for 25+ coding agents — Claude Code, Codex, and Cursor finally remember the same project.

    • mcp
    • memory
    • cross-agent
    • context