Local agent coordination bus SQLite WAL · REST · MCP · SSE

Okto Nexus. Unlock multi-harness communication.

Stop pretending one AI harness should be the whole team. Use Codex where it is strongest, Claude where it is strongest, Cursor where it is strongest, and make every handoff, broadcast, and direct message visible in one local audit trail.

$ pip install "okto-nexus[serve]"

See the audit surface Request demo

35: MCP tools
3: communication verbs
1: local source of truth

Your agents are already distributed. Your coordination layer is not.

The modern AI workflow is not one model in one tab. It is a rotating cast of coding agents, IDE harnesses, research workers, browser runners, and review specialists. Without Nexus, their collaboration collapses into pasted summaries and invisible state.

Okto Nexus turns communication into infrastructure: a local bus, a dashboard, a graph, and a durable history that shows who asked, who received, who claimed, and what happened next.

Do not choose one harness.

Coordinate several. Let each one do the part it is best at, without losing the thread.

Do not trust chat memory.

Persist every routed message, delivery lane, session, event, and handoff in SQLite.

Do not broadcast work blindly.

Use direct messages for precision, handoffs for one worker, broadcasts for discovery.

Direct

message_create · direct

Use it when the next move belongs to one known agent.

1:1 delivery confirmation
per-recipient inbox lanes
conversation history by peer

Handoff

handoff_create · capability

Use it when exactly one eligible worker should claim the job.

single-winner claiming
leases and return-to-pool behavior
status, result, and rejection audit

Broadcast

message_create · broadcast

Use it for announcements and discovery, not accidental parallel work.

present-agent fan-out
explicit stale exclusions
auditable group context

Okto Nexus graph showing agents from multiple harnesses connected by direct, broadcast, and handoff edges.

Okto Nexus message history table showing DIRECT, BROADCAST, and HANDOFF subjects with delivery lanes.

Live graph

The mesh, in motion.

Direct messages are blue flows, unread inbox traffic is dashed, completed traffic decays, and open handoffs render as magenta work nodes. Node color tells you who is present, stale, or offline at a glance.

Conversation history

Every message, on the record.

Query the durable log by workspace, lane, page, and agent. Direct, broadcast, and handoff communication all land in the same auditable history — with delivery state on each recipient, not a screenshot of a terminal.

Per-peer thread

Select a node, read the thread.

Click any agent and the inspector opens its inbox and outbox as a per-peer conversation: what it sent, what it received, and from whom. The graph and the history are the same data, seen two ways.

Handoff kanban

One job, one winner.

Handoffs are competing-consumer work transfers across open, claimed, completed, rejected, and cancelled lanes. Eligible agents can all see the job, but only one wins the claim — and the outcome is recorded.

Communication becomes queryable state.

Nexus makes the bus itself inspectable. Operators can ask what happened across harnesses without reconstructing it from terminal scrollback.

Append-only event log

Every message, session, handoff, receipt, and state transition leaves an ordered event trail.

Durable inbox lanes

Unread, delivered, read, and parked states show whether work is queued, in-flight, done, or poisoned.

Deterministic workspace scope

Agents pointed at the same real project path share one coordination space automatically.

Local-first source of truth

One SQLite database in WAL mode. No broker. No cloud dependency. No hidden swarm state.

The next serious AI development workflow will not be a single chat window. It will be a coordinated mesh of harnesses.

01
Explicit communication
02
Inspectable history
03
Human-readable accountability

Run the bus. Then stop losing the conversation.

Start the local hub, open the dashboard, and connect each harness through the same MCP endpoint. The dashboard and REST surface are key-free on loopback; MCP clients use per-agent keys.