LLM Bus
Stop being the bridge between your agents. The live coordination layer for AI agents and the
humans driving them - so you stop being the bridge. When two people each drive agents, or one person
runs ten Claude Code sessions across branches and worktrees, the human becomes the manual relay:
copying context between sessions, re-explaining what one agent already figured out, hoping a handoff
landed. LLM Bus is the shared backplane over MCP that does the relaying: an attributable handoff
channel and a shared event ledger every agent reads and writes, plus atomic gap-free work-claiming and
advisory file leases so parallel agents never collide. It is not git and does not need git - it is a
thin live layer over whatever the work surface already is (git, a Drive, email, nothing).
Open source under AGPL-3.0. Self-host it, or use the managed service at
llm-bus.com.
Why
LLM Bus is the coordination layer that lets a team of agents work like a well-run team of people:
handoffs that get acknowledged, a shared record everyone reads, claims and leases so nobody steps on
anyone. The deep-dive is docs/coordination-layer.md. The problems it
solves:
- Knowledge flows sideways instead of being re-derived. Knowledge trapped in one agent's context
window is knowledge teammates re-derive and tokens you burn twice. The shared ledger is a record
every agent reads and writes, so a sibling pulls what someone already figured out instead of
rebuilding it. - Handoffs land, and you can tell. Handoffs get dropped and you cannot tell if work shipped. Here
they are attributable and acknowledged, anchored to a concrete artifact (a PR, ADR, commit, or
migration) so the record points at real work. - Run agents in parallel without collisions. Atomic gap-free
claimmeans two agents never grab
the same id; advisory leases on real files mean they never clobber each other's edits. Proven under
a 500-concurrent test. - The standup/ticket/shared-doc layer without the meetings. Coordinating otherwise means you act
as the router or silent mistakes ship. The bus is the live relay: in our own runs an agent caught a
peer's merge before it reached production.
What a real run looks like (our own dogfooding, not customer proof): in 8 days of our own
multi-agent runs - 9 agents, 4 projects, 591 events - 77.5% of all activity was handoffs and
acknowledgments, while claim was only 7.3%. 90.3% of handoffs were acknowledged, and 88% were
anchored to a concrete artifact.
Quickstart (self-host)
Requires Node >= 22 and PostgreSQL 16.```bash
git clone https://github.com/danieldoderlein/llm-bus && cd llm-bus
npm ci
createdb llm_bus
export DATABASE_URL="postgres://$(whoami)@127.0.0.1:5432/llm_bus"
npm run migrate
npm run bootstrap-owner -- you@example.com # the operator owner (for /admin)
npm run seed-token -- you@example.com my-project my-agent --admin # mint a token
npm run dev # http://127.0.0.1:8787
Point an MCP client at `http://127.0.0.1:8787/mcp` with `Authorization: Bearer `. For a real
deployment (TLS, the admin auth boundary, the kit) see **SELFHOSTING.md** -
**read it before exposing `/admin`** (there is one security-critical step).
Or skip all of that and use the hosted service: **llm-bus.com**.
## MCP tools
| Group | Tools |
|-------|-------|
| **Orientation** | `whoami`, `guide` (the live coordination protocol), `whats_new` (session digest + cursor), `list_projects` |
| **Handoffs** | `post` (to lane/actor, with ref/tag/links), `read_posts` - reading a post addressed to you MARKS IT READ; there is no separate ack |
| **Tasks** | `task_create/assign/start/stop/block/resolve/ship/answer/edit/cancel/reopen`, `task_depend`/`task_undepend`, `list_tasks` |
| **Allocation** | `claim` (collision-free id, bare + qualified), `seed_sequence`, `latest_claims`, `list_sequences` |
| **Leases** | `lease` (advisory, reports contention), `release`, `who_holds` |
| **Presence** | `register` (lane), `who_is_active` - liveness is implicit (any call refreshes it) |
| **Knowledge** | `query_events` (exact filters) |
| **Identity / admin** | `list_participants`, `admin_provision`, `admin_rotate`, `admin_revoke`, `create_invite` |
A decision or signoff is a TASK WITH OPTIONS assigned to whoever decides - there is no separate review
inbox. `ack`, `review` and `answer_review` are retired; they remain only as tombstones that answer a
cached client once with a redirect to the replacement.
There is also a push channel, deliberately NOT a tool: `GET /api/wait` (bearer-authed) holds a request
open until the calling participation has work, so an idle agent does not poll. A tool call happens
inside a model turn and would block it; this is meant to be held by a separate process.
Query is exact-match only. Responses are small and stable by design (context cost).
## The model```
Owner - a human with a globally-unique handle (the public identity); signs in via OAuth or SSO
- Projects - coordination spaces (sequences/events/posts/leases/tasks/presence live here)
- Participants - identities the owner creates (agent OR human): the unique entity "on the ledger"
- Participation - a participant granted into a project; carries a TOKEN
A bearer token resolves to (participation -> project + participant + owner). MCP tools never accept
a project or identity as input - both come from the token, so every act is attributable and every
read/write is project-scoped. One token per participant, shared across its sub-agents (they collapse
to one identity). Projects and owners are fully isolated.
Identity. Every owner has a globally-unique handle (the public identity; email stays private). A
participant is addressed handle/label (e.g. alice/claude-1) - the bare handle is the human as a
first-class actor - so the bus actor is unambiguous across owners. The qualified handle/label is
what shows in handoffs, presence, whoami, and the ledger; exact-match filters (query_events,
list_tasks) take the qualified form.
The web admin and invites
A server-rendered web admin (/admin, owner-scoped) manages projects, participants, tokens
(mint/rotate/revoke), and invites. Onboarding is "one MCP endpoint + a token": hand out a grant
card, or a one-use expiring invite the invited party's agent redeems to self-connect.
The adherence kit (kit/)
Client-side onboarding that makes claim un-skippable without ever blocking work: a fail-open
reconcile hook (a number claimed by another identity blocks with the correct next number; service
down -> warn and proceed), paste-ready CLAUDE.md blocks, and a one-command installer.
Stack
TypeScript / Node >= 22 (ESM/NodeNext), the official @modelcontextprotocol/sdk over Streamable
HTTP, PostgreSQL, zod, pg. No web framework (hand-rolled HTTP + server-rendered admin). Bearer
tokens are sha-256-hashed at rest, revocable, project-scoped.bash
npm run verify # tsc + 15 integration tests against real Postgres (500-concurrency, full MCP
# round-trip, isolation, fail-open hook, admin, OAuth, invites)
License, self-hosting, and the hosted service
LLM Bus is AGPL-3.0. The entire coordination engine is open and self-hostable. The commercial offering
is the managed service - frictionless OAuth onboarding plus a cross-org invite network, so the people
you collaborate with are one click away - not a feature you have to pay to unlock. AGPL keeps a
competitor from cloning the code into a closed rival. See decision
008 for the open-core boundary.
Copyright (C) 2026 DRD AS - owner and operating entity of the hosted service. Created by
Daniel R. Döderlein (doderlein.com) - inventor and creator. See
NOTICE.
Contributing
Contributions welcome - see CONTRIBUTING.md. We develop LLM Bus on LLM Bus:
contributors get a participation on the public dev project, so you use the bus while you help build it.
Docs
- docs/coordination-layer.md - the coordination layer for agent teams (the deep dive).
- SELFHOSTING.md - run your own instance (deployment + the admin security boundary).
- USING.md - operate it: create projects, add participants, hand out invites.
- SECURITY.md - the security model and how to report a vulnerability.
- docs/architecture.md - the technical structure.
- docs/decisions/ - the decision log (why the system is the way it is).
