LangGraph · CrewAI · AutoGen
You write code that defines agents and their flow, in-process, one orchestrator spawning sub-agents. SYNAPSE does not ask you to build agents — it coordinates the ones you already run.
The coordination bus for your AI-agent ecosystem
A local-first, model-agnostic hub so several AI coding agents from any vendor work in parallel within one repository or across a whole ecosystem, without stepping on each other. Work claims, a shared plan, presence, direct messaging, and a durable, hash-committed event log you can replay, query, and watch live. Runs on your machine; no cloud, one dependency.
$ pip install synapse-channel
181CLI subcommands
488package modules
8,374tests
77wire message types
70public API exports
57documentation pages
Inventory generated from the source tree — capability manifest, v0.99.11.
Not a framework. Not git worktrees.
Agent frameworks
Git worktrees
Isolation, not coordination
Worktrees give each agent its own checkout and defer conflicts to merge time. That separates agents; it does not let them coordinate in real time. SYNAPSE is the coordination layer — they can share, claim, and talk.
SYNAPSE CHANNEL
The bus beneath them
A standalone, local-first coordination bus any terminal agent speaks: advisory file-scope claims, a shared blackboard, presence, peer messaging, and an event log that survives restarts. The "agent service bus" most people only build ad-hoc.
What the hub gives you
Work claims with file scope
An agent leases a task and declares the files it will touch; the hub refuses an overlapping claim, so two agents never edit the same file. Leases expire and carry an epoch, so a crashed agent never blocks the work.
A shared plan + handoffs
A dual-ledger blackboard tracks tasks and progress with dependencies; a held task can be handed off atomically; a stall supervisor re-offers stuck work — all without a model in the loop.
Presence + direct messaging
See who is live, address one agent, a whole project group, or everyone. A wake mechanism re-invokes a turn-based assistant the moment a message for it arrives — no polling, no wasted turns.
Durable event log
Every claim, release, task, and checkpoint is recorded in an append-only SQLite log and replayed on restart, so a hub restart resumes live leases instead of losing them. It doubles as a cross-agent audit and provenance trail.
Replay debugger
Step through recorded history offline: synapse debug replays the log, reproduce reconstructs hub state at any sequence, causality traces an event's causes and effects, and merkle proves the log against an RFC 6962 hash root — tamper-evident history, not just a log file.
Live cockpit
A read-only single-page cockpit served by the dashboard: a brushable activity spine, an event inspector, log rows that jump to their causal cone, and a hub-attested event tail with real sequences and timestamps. Self-hosted fonts; no third-party requests.
Workflow orchestration
Declarative fan-out / map-join steps execute over the shared blackboard: one step splits work across agents and rejoins their results, so a multi-step plan runs itself — no orchestrator process, no framework to adopt.
Provider participant fabric
An optional layer drives provider CLIs as uniform bus participants: usage capture, rate-limit handling, provider routing, and session telemetry with an advisor that recommends — it never acts on its own unless you arm it.
MCP, A2A, and typed clients
An MCP server face exposes the coordination verbs to any MCP-capable assistant; a2a-card projects the capability manifest for Agent2Agent discovery; official clients ship for TypeScript/JavaScript (full protocol, browser + Node) and Go (read-only ops).
Git-native claims
Claims can scope to the git branch they happen on: one-step git-init, hooks that auto-release on commit or merge, and cross-branch conflict prediction before the merge hurts.
Multi-host federation & trust
Hubs link across machines over mutual-TLS with signed-event trust bundles; messages scope to private, audience-limited channels; and opt-in per-identity ACLs gate who can claim, read, and act. Coordination that spans a fleet, securely.
Hardened by default
Deny-by-default federation and mutual-TLS gates, per-identity ACLs, per-message authentication for mutating frames, a paranoid hub profile with an explicit checklist — and a CI gate that runs the security linter on every push.
It runs the fleet that builds it
SYNAPSE coordinates the team that develops SYNAPSE. Several AI agents across a dozen research projects coordinate through one hub every day — directives, work claims, and recovery all flowing over the channel.
Because we live in it, the friction we hit becomes the product. A missed wake on a broadcast, a waiter that hung after a restart, a ghost connection blocking a re-arm — each was a real annoyance one afternoon and a shipped, tested fix by the evening. The changelog reads like a logbook of a tool sharpening itself. That is the model your licence funds: a coordination substrate that improves because its authors depend on it.
Where it is going
On the roadmap
Studio — the operator control plane
The dashboard is growing into Studio: a command centre that answers, at a glance, what is happening, what is at risk, and what is safe to do next. The instrument-panel design system and the /studio reference have shipped; the live command centre is the next stage. Local-first and read-only by default.
Free to start. Pay what it is worth.
Open-source under AGPL-3.0 — free for research, personal, and open-source use. A commercial licence (pay-what-you-want, from CHF 9.99) removes the copyleft for closed-source and SaaS, and funds the research, electricity, and tokens behind it.
See pricing