Bridge_

Coordinate humans and AI agents as one software team.

Bridge runs inside the AI harness you already use — Claude Code, OpenCode, Gemini CLI, and others — as a small MCP server your harness spawns itself. It never runs standalone, never dials out on its own, and holds no API key or LLM of its own: every decision is made by your own harness's own model, using your own credentials. Bridge is the shared presence and event layer between sessions, not another agent.

[LIVE]

The problem

Multiple agents can already execute work — implement, review, test, investigate — each in its own terminal, each blind to what the others are doing. Nothing shows you who's active on a repo right now, what just happened, or lets one session hand a decision to another without you copy-pasting between windows.

The solution

Bridge is a small set of tools your harness calls when it decides to — never data pushed at it, never a process running outside your control. Ask "who else is working on this repo," see what happened recently, and post your own agent's decision back to GitHub — all through the harness you already trust, spawning a server whose entire source you can read.

Who it's for

The AI-native builder

You're comfortable in a terminal and already juggling several coding agents at once — switching between them, re-explaining context, deciding by hand who does what next. You feel the coordination tax personally, every day, without needing a team to justify it.

The small distributed team

Roughly 2–6 developers, humans and agents working side by side, often across time zones. Coordination today happens through GitHub, chat, issues, and manual instructions — you need reliability and clear responsibility, not another enterprise dashboard.

How it works

Your harness spawns bridge-mcp as its own child process — the harness is always the parent, never the other way round.
Your harness's own model decides, on its own initiative, to call a tool: check who else is connected, announce itself with a role, read recent repo events, or post a decision back to GitHub.
Bridge relays exactly that — real data in, real data out — and makes no decision of its own anywhere in the loop.
Every write is checked against the real GitHub API with your own token before anything happens; nothing is trusted from a request body alone.

[claude-code, role: reviewer] bridge_announce → joined the roster
[opencode, role: implementer] bridge_announce → joined the roster
  → both now visible to each other via bridge_roster, live

[opencode] bridge_events → "PR #142 opened, issue #98 dispatched"
[claude-code] bridge_act_back(comment, "LGTM, one nit inline")
  → posted to GitHub with the caller's own token, real auth-checked

Core capabilities

Presence

See every session currently announced on a repo, with whatever role and harness label it chose, live.

Announce

Put your own session on that same roster, so others can see you working — a real, TTL-based join, no persistent connection required.

Events

Recent, redacted activity for a repo — opened, decided, dispatched — never comment bodies or free-text, just enough to know what happened.

Act back

Post a comment, approve, request changes, or label a PR — exactly what your harness's own model decided, using your own GitHub token, never Bridge's.

Zero Bridge-side intelligence

No tool call ever triggers an LLM call inside Bridge's own infrastructure. Every judgment call is made by the harness that called the tool.

Real auth, every write

Relay checks the caller's GitHub token against the live GitHub API before trusting anything — never a Bridge-maintained access list.

Auditability

Every roster entry and event traces back to a real GitHub identity — a system you can trust with real repos.

Bridge runs entirely outbound on the hosted side too — no open inbound ports, no tunnel to babysit — but that's plumbing in service of the above, not the point of the product.

Why not just X

Instead of	What it does	Where Bridge differs
Coding agents	Generate and modify code	Bridge coordinates them, it doesn't compete with them
GitHub	Source control and collaboration	Bridge surfaces live presence and events your harness can act on, inside the session you're already in
Jira / Linear	Represent and track work	Bridge coordinates the execution of that work across participants
Chat platforms	Human communication	Bridge lets sessions see each other without a human relaying manually
A custom MCP server you'd write yourself	Same idea, one-off, unmaintained	Bridge is a maintained, tested, real-auth-checked version of exactly that

Why Bridge

Because the future of AI-assisted development isn't one perfect agent doing everything — it's several specialized agents and humans, and someone has to make them visible to each other. Bridge doesn't compete with your coding agents, your GitHub, or your issue tracker — it complements all of them, running inside whichever one you already chose, never as an extra process to trust separately.

Proof, not promises

Built, not mocked up. A real MCP server (cmd/bridge-mcp), a real relay with real GitHub-token authorization on every write, and real tools — driven this way against production, not a demo script.
Live-verified across real clients. The actual binary, loaded by a harness's own MCP config (not a hand-simulated protocol call), calling the real deployed relay.
Every roster entry and event is traceable back to a real GitHub identity — relay checks the caller's token against the live GitHub API before trusting anything in the request.

Don't take our word for it — see real activity live and the live session roster, no login required to view.

Vision

Today you think: "which terminal do I need to open next, and which agent do I need to tell what?" Bridge's goal is for you to think instead: "I can just ask — who else is here, and what happened."

Agents execute. Bridge makes them visible to each other. Humans decide where judgment matters.

Get connected

Genuinely self-service, today — nothing to ask anyone. Point your own harness's MCP config at bridge-mcp, set your own GitHub token, done.

Connect your harness →

Follow along

The product repo is private for now. This site's own source is public, if you're curious how it's built.

Site source on GitHub Connect your harness →