Meet Roster

Roster is an open-source local tool router for AI agents. It brings command-backed MCP servers and approved skills into one place on your machine, with local outcome learning and reversible client setup.

It routes capabilities, not AI models. Your agent stays responsible for deciding what to do and verifying the result.

One local entry point

Instead of managing separate tool-server entries across clients, connect the client to Roster. The client launches a local stdio process; Roster connects to your configured command-backed servers and forwards tool calls.

Transparent passthrough is the default. Your agent sees the available backend tools under stable, namespaced IDs. You do not need to narrow your tool collection or enable a special routing mode to start using Roster.

Configuration sync and eject are supported for Claude Code, Cursor, Codex, and OpenClaw. Roster can discover more configuration formats, but discovery is not equivalent to automatic write support or a verified live integration.

Local by design

The configuration, capability index, derived outcome log, ratings, and routing maintenance live on your machine. There is no Roster account, Roster API key, or hosted routing service.

The Coach records useful call-level evidence and uses later maintenance to inform routing. The Playbook indexes approved skill instructions alongside tools. In five mode, a skill invocation returns instructions and resource paths; Roster does not run its bundled scripts.

Lexical retrieval needs no model download. Semantic retrieval is optional and also runs locally once its runtime and model are available.

Your backend tools can still contact their normal services. “Local router” is not a promise that every downstream API call is offline. Read the privacy boundary.

The starting five, when you want it

Five mode is an optional way to use the same router, not the whole product.

1. The agent calls draft with its immediate need.
2. Roster returns compact tool or skill cards, up to five by default.
3. The agent chooses a capability and invokes it through call.
4. The Coach records the derived call outcome locally.

The client sees two stable tools, draft and call. Request k from 1 to 10; the default is 5. A returned result is not proof that the user’s whole task is finished.

The “200 tools” example illustrates a large toolkit. It is not a measurement of your installation or a guaranteed savings figure. Your receipt reports the inventory Roster actually discovers.

The product vocabulary

Name	What it does
The Rotation	The local MCP router and backend connections.
The Coach	Local outcome storage, retrieval, ratings, and maintenance.
The Playbook	Trust-gated skill instructions in the capability index.
The Combine	Reproducible task suites and declared checks.
The League	A separate, deferred public rankings experience with explicit evidence rules.

You do not need the Combine or League to use the local router.

Available now

The core is implemented and the repository is public. The scoped npm package is published as version 0.0.4. The public League remains deferred.

These guides describe the revision shown above, not the older roadmap’s planned HTTP support, dashboard, or automatic failover. The current backend boundary is command-backed stdio.

Next: Install and run your first receipt, or hand the setup to your agent.

Router implementationProduct statusRecorded core verification