synapse-channel 0.27.1

Local-first multi-agent coordination bus: work claims, presence, chat, and resource sharing over a WebSocket hub

Local-first coordination bus for multiple agents working in parallel on one repository

A local-first coordination bus for multiple agents working in parallel. A single WebSocket hub is the source of truth for presence, work claims, chat, task status, and resource offers, so concurrent workers do not step on each other or duplicate effort.

The bus is transport-light (one dependency, websockets), hub-centric by design (one place owns presence, leases, and history), and runs entirely on the local machine. Model workers reply on-channel through any OpenAI-compatible endpoint, including a local Ollama server, with a deterministic rule-based fallback for offline use.

At a glance

graph LR
    A1["Agent"] --> H
    A2["Agent"] --> H
    A3["Worker"] --> H
    SUP["Supervisor"] --> H
    H["SynapseHub<br/>single source of truth"] --> CL["Claims & leases<br/>scope · epoch · checkpoint"]
    H --> BB["Blackboard<br/>plan + progress"]
    H --> CAP["Capabilities<br/>cards + routing"]
    H --> LOG["Event log (SQLite WAL)<br/>durable · replayed on restart"]

A claim leases a unit of work with a file scope, so two agents never edit the same files; the plan, handoffs, checkpoints, and a stall supervisor keep the work moving; and the durable event log means a hub restart resumes live leases rather than losing them.

Install

python -m pip install synapse-channel       # the release from PyPI
python -m pip install -e ".[dev]"           # or an editable dev checkout

This installs the synapse command. To run the hub as an always-on local service or a container, see the deployment guide (a systemd user unit and docker compose are both included).

Quick start

Launch a hub plus one or two local model workers in one command:

synapse team

Then, from another terminal, watch the channel or send a message:

synapse listen --name USER
synapse send --name USER --target FAST "what is the status of TASK-1?"

Running pieces individually

synapse hub --port 8876
synapse hub --port 8876 --db ./synapse.db            # crash-safe: resumes leases + history on restart
synapse hub --port 8876 --relay-log ./feed.ndjson    # mirror the channel to a compact file for observers
synapse worker --name FAST --provider ollama --model gemma3:4b
synapse worker --name OFFLINE --provider rule        # no network, canned replies
synapse worker --name TIER --provider tiered --model small --heavy-model big  # route trivial→rule, hard→heavy
synapse relay ./feed.ndjson                          # decode and print that file as readable lines
synapse board                                        # print the shared task/progress blackboard
synapse task declare BUILD --title "compile"         # declare/update the shared plan from the CLI
synapse task update BUILD --status done              # mark a plan task done so dependents unblock
synapse supervisor --idle-seconds 300                # LLM-free: re-offer tasks that stall
synapse manifest                                     # print the capability cards agents advertised
synapse hub --host 0.0.0.0 --token s3cret            # require a shared secret when binding off-loopback
synapse send --token s3cret --name USER "hello"      # agents present the token to a secured hub

Durability

Passing --db backs the hub with an append-only SQLite event log (standard library, WAL mode). Every claim, release, task update, resource offer, and chat message is recorded, and the hub rebuilds its state by replaying the log on start-up. The guarantee is split honestly by workload: the lease/claim path commits at synchronous=FULL (durable across an OS crash); the high-volume chat/history path commits at synchronous=NORMAL (durable across an application crash, may lose the last commit on power loss).

Token-thrifty observation

--relay-log mirrors every broadcast to a newline-delimited file in a compact short-key form (encode_lite), so a token-budgeted agent can watch the channel by tailing a file instead of holding a socket. synapse relay <file> decodes it back to readable lines and can resume from a saved --cursor. The lite form keeps the seven core envelope fields and drops auxiliary ones; the file is bounded by --relay-max-lines. A committed benchmark measures the saving honestly — see benchmarks/.

Exposure

By default the hub binds to loopback and runs with no authentication — the right posture for one operator on one machine. When that is not enough (a worker with tool-use, or a hub bound off-loopback), --token requires a shared secret that connecting agents present with --token; the hub warns if it is bound to a non-loopback host without one. This is a proportionate gate, not a cryptographic identity system.

Coordination model

1. Claim before you work: an agent leases a task by id; a live lease blocks other agents from claiming the same task.
2. Declare a file scope on the claim (a worktree and paths); the hub refuses a claim whose files overlap another agent's live claim — this is how two agents are kept off the same files. Agents in different worktrees never contend.
3. Leases auto-expire, so a crashed agent never holds a claim forever, and each lease carries an epoch so a superseded agent cannot act on a dead claim. An owner can save a durable checkpoint on the task; if its lease lapses, the next agent to claim the task inherits that checkpoint and resumes rather than restarting.
4. Release on completion; status and an optional artefact reference can be attached while the task is in progress. A held task can also be handed off atomically to another online agent — keeping its scope, status, and context, with no window for a third agent to grab it mid-transfer.
5. Presence, who, full state snapshots, and chat history are queryable at any time. After a reconnect, an agent resumes by idem_key (retried claims are not applied twice) and a resume cursor (fetch exactly the messages it missed).

Alongside the lease registry, a shared blackboard holds the team's plan: a task ledger of declared work with dependencies (the hub refuses dependency cycles, so ready tasks are well-defined) and an append-only progress ledger a supervisor can read to spot stalls. A declared LedgerTask is the plan; a claim is the lease on doing it — the two share a task id but stay independent, so the simple claim flow keeps working. View it with synapse board.

See TEAM_PROTOCOL.md for the working agreement and message reference.

Library use

import asyncio
from synapse_channel import SynapseHub, SynapseAgent

async def main() -> None:
    hub = SynapseHub()
    asyncio.create_task(hub.serve("localhost", 8876))
    agent = SynapseAgent("ALPHA", uri="ws://localhost:8876")
    # ... drive the agent: claim, chat, request state ...

Two self-contained, runnable demos live in examples/: coordination_demo.py narrates a full task through the bus (declare, block, claim, refuse an overlap, unblock, hand off), and llm_team_demo.py asks an on-channel model worker a question. Each starts its own in-process hub, so python examples/coordination_demo.py runs with nothing else set up.

Architecture

Module	Responsibility
state	Presence, scoped task-claim leases, epochs/versions, and resource offers (transport-agnostic).
ledger	Shared blackboard: the declared task plan (with dependencies) and an append-only progress stream.
scoping	Worktree- and path-overlap detection that keeps two agents off the same files.
lifecycle	Typed task-status states and the legal transitions the hub enforces.
deadlock	Wait-for cycle detection so circular hold-and-wait claims are refused.
protocol	The on-wire message envelope and message-type constants.
relay	Lite/heavy codec (encode_lite/decode_lite) and append-only NDJSON log helpers for file-based observers.
hub	The routing core: connections, names, history, broadcast.
client	The reusable async agent connection and coordination helpers.
persistence	Append-only SQLite event store (WAL) giving the hub a crash-durable spine.
journal	Records mutations as events and replays them to rebuild state on restart.
ratelimit	Per-agent token-bucket limiter so one runaway agent cannot swamp the hub.
auth	Optional shared-secret connect token (proportionate, not a cryptographic identity).
chat_backends	Pluggable reply backends (OpenAI-compatible HTTP, rule-based).
routing	Classify a request into a task class and route it to a tiered backend.
llm_worker	An on-channel agent that answers addressed messages via a backend.
supervisor	LLM-free watcher that spots stalled plan tasks and re-offers them.
capability	Agent capability cards (A2A-shaped) and the hub-aggregated manifest.
launcher	One-command local hub + worker startup.
cli	The unified synapse command.

Capability inventory

Module and surface inventory — counts kept in sync with the source tree by CI.

SYNAPSE CHANNEL capability inventory

Surface	Current inventory
Package version	0.27.1
Public API exports	41
Package modules	21
Classes	28
Wire message types	47
CLI subcommands	17
Test functions	545
Benchmark harnesses	2
Documentation pages	12
GitHub Actions workflows	9
Optional-dependency groups	3

This snapshot is a static inventory generated from the source tree. Performance and coverage claims have their own committed evidence — see VALIDATION.md and benchmarks/.

Documentation and project

• ARCHITECTURE.md — the module map and coordination model.
• TEAM_PROTOCOL.md — the working agreement and wire reference.
• VALIDATION.md — how it is tested and the gates a change clears.
• CONTRIBUTING.md · SECURITY.md · GOVERNANCE.md · ROADMAP.md
• Full documentation site: https://anulum.github.io/synapse-channel

Known limitations

• Single hub, single machine. There is no built-in failover or horizontal scale; the hub is one process and the design is deliberately local-first. A hub restart resumes from the durable log, but it is not a high-availability cluster.
• Connect authentication is a proportionate shared secret, not a cryptographic identity system — no key exchange, signatures, or per-message authentication. Do not expose the hub on an untrusted network and rely on the token alone.
• Agents are trusted. The bus coordinates agents; it does not sandbox them. An agent is trusted to the extent the operator trusts the process it runs in.
• Task-class routing is heuristic. The classifier sorts a request by length and a keyword set; tune the thresholds for your workload. Per-tier model latency is not benchmarked offline (it needs a live model server).

How to cite

If you use SYNAPSE CHANNEL in your work, please cite it. Metadata is in CITATION.cff; a BibTeX entry:

@software{sotek_synapse_channel,
  author  = {Šotek, Miroslav},
  title   = {SYNAPSE CHANNEL: Local-first multi-agent coordination bus},
  url      = {https://github.com/anulum/synapse-channel},
  version = {0.17.0},
  year     = {2026}
}

Licence

Dual-licensed: AGPL-3.0-or-later, with a commercial licence available. See LICENSE for the full text and NOTICE.md for the licensing boundary; contact [e-mail] for commercial terms. The repository is REUSE 3.x compliant.

---

   

© 1998–2026 Miroslav Šotek · anulum.li · [e-mail]