Question 1

What is Baton?

Accepted Answer

Baton is a local-first coordination hub for running multiple AI coding agents on one repository. It gives each agent an isolated git worktree, streams a live dashboard of who is doing what, keeps a shared evidence-anchored memory, and packages session handoffs as plain markdown files.

Question 2

Which AI coding agents does it work with?

Accepted Answer

Claude Code, Cursor, Codex, and Gemini out of the box, plus local agents like aider and opencode. Anything that can read a markdown file and run a CLI can take a handoff; agents that speak MCP can query the knowledge base directly.

Question 3

How much does it cost?

Accepted Answer

Nothing. Baton is AGPL-3.0 open source. It needs no API key of its own and makes no LLM calls — routing is deterministic scoring, not a model. The optional local tier runs on Ollama, so routine tasks can cost zero even in API terms.

Question 4

How does session handoff work?

Accepted Answer

When an agent stops — limit hit, end of day, or a deliberate switch to a cheaper model — one markdown file captures the whole session: objective, plan, remaining checklist, and cost so far. The next agent resumes from that file. No proprietary state, no export step.

Question 5

What is evidence-anchored memory?

Accepted Answer

Shared facts are pinned to git commits and content hashes. Before a fact is served to an agent, its anchor is checked against the current code; if the code has changed, the fact is withheld rather than served stale. Agents can be told nothing, but they cannot be told something outdated.

Question 6

Won't multiple agents conflict on the same files?

Accepted Answer

Each agent works in its own isolated git worktree, so parallel edits cannot clobber each other's checkouts. On top of that, live edit signals track which files each agent is touching, and overlapping edits trigger a warning before they become a merge conflict.

Question 7

Does my code leave my machine?

Accepted Answer

No. The daemon binds to loopback only and screens both the Origin and Host headers on every request, which blocks cross-site request forgery and DNS-rebinding attacks. There is no cloud component, no telemetry, and no account.

Question 8

What is the dashboard?

Accepted Answer

A realtime web dashboard the daemon serves on localhost:7077. It shows agent activity, file conflicts, live agent terminals (tmux-backed), the shared memory, and the code knowledge graph. Updates stream over plain server-sent events — no websocket library, no polling.

Question 9

How does model routing decide which agent gets a task?

Accepted Answer

Each task is scored by severity and routed into a tier — heavy, standard, light, or local. The scoring is deterministic: no LLM call, instant, and every decision is explainable. The bottom tier runs a local model via Ollama, so low-stakes work is free.

Question 10

What are skills?

Accepted Answer

Ten installable agent playbooks — bug-fix, handoff, routing, and others — installed into each agent's own config format (.claude/skills, .cursor/rules). The bug-fix skill enforces a real gate: no code changes until an independent skeptic corroborates the root cause at 95% confidence and you approve.

Question 11

Do I need a database or any other infrastructure?

Accepted Answer

No. Coordination state lives in git and plain files inside your repo. The daemon is dependency-free — raw node:http, no framework. If you delete the daemon, your repo still contains every handoff and fact as readable markdown.

Question 12

How do I install it?

Accepted Answer

Clone the repository from GitHub, run npm install and npm run build in the root and web workspaces, then start the daemon with node dist/cli.js serve --write. The dashboard comes up on localhost:7077. Full steps are in the docs.

Questions, answered straight.

Still curious?