Baton

Baton orchestrates AI agents through configurable pipelines. Like a relay baton, it passes work and context from one step to the next, coordinating agents and human decisions along the way. It combines a Rust API and CLI with a SvelteKit interface for configuring agents, managing tasks and reviewing runs. Baton controls the workflow; agent runtimes execute individual steps.

An example workflow; pipeline steps and transitions are configurable.

Install the CLI

On macOS or glibc-based Linux (x86-64 or ARM64), install the latest GitHub Release:

curl -fsSL https://raw.githubusercontent.com/klab365/baton/main/install.sh | sh

The script installs baton to ~/.local/bin; make sure that directory is on your PATH. Set BATON_INSTALL_DIR to change the destination or BATON_VERSION to select a release, for example:

curl -fsSL https://raw.githubusercontent.com/klab365/baton/main/install.sh | BATON_VERSION=0.1.0 sh

You can instead download an archive from GitHub Releases and manage the binary with your preferred tooling, including mise. Linux builds target glibc, not musl. The release binary includes the API and web UI; Node and Bun are not required at runtime.

Get started

In a Git project, start the installed API:

cd /path/to/project
baton

On first launch Baton creates .baton/ and initializes the project configuration automatically. Open http://127.0.0.1:5173 for the embedded UI. BATON_PORT changes the starting port. If it is occupied, Baton uses the next available port and prints the address. Configure credentials and agents, create a pipeline, then create and start a task. Follow the Software Factory getting started guide for a complete example. The fake runtime is available for offline testing. Agent execution requires Git; real model execution also requires a compatible provider credential.

A worktree is not a sandbox: agents may run commands with the permissions of the Baton process. Use a container, VM or dedicated OS user for stronger isolation. Task prompts and run data are local and can be sensitive.

Pipeline shell steps run their configured sh -c command in the task worktree, with the Baton process's permissions (not in a sandbox). They stop on cancellation and time out after 300 seconds by default; set timeoutSeconds per shell step (1–86400 seconds) to change the limit. Exit code 0 produces completed; all other exit statuses and timeouts produce failed. Configure transitions for those outcomes to route results; without a matching route, failures stop the run. The exit status and bounded stdout/stderr are passed to later steps. Shell steps do not create commits and are not supported in external baton work sessions.

Agents can be configured with an explicit allowed tool list in the Agents UI. An omitted list preserves the default tools; an empty list disables all tools. Steps with transitions require report_outcome. Disabling write and edit does not prevent changes if shell is enabled: shell commands can modify files and access resources outside the worktree.

Set up with a coding agent

Copy this prompt into your coding agent from the Git repository where you want to use Baton:

Help me set up Baton for this repository. First read the current agent setup guide:
https://github.com/klab365/baton/blob/main/docs/getting-started-for-agents.md
Follow it, inspect my environment and existing configuration, and ask me before
making changes or taking actions that need my approval. Do not overwrite existing
configuration or ask me to paste secrets into this conversation.

See the agent setup guide for what the agent will do and for links to the full documentation.

Development

mise install
bun install --frozen-lockfile
mise run dev

This starts the API and Vite development UI using examples/demo. The API is on port 5173 and Vite on port 5174, proxying /api to the backend. To build a standalone binary from source, run bun run build followed by BATON_WEB_DIR="$PWD/build" cargo build --release -p baton-cli. Without BATON_WEB_DIR, Rust builds show a placeholder instead of the embedded UI. To run checks:

mise run check-format
mise run lint
mise run check
mise run test
mise run test-embedded-web
mise run build

Documentation

• User guide — features, workflows, CLI and troubleshooting
• Getting started: build a Software Factory, setup for coding agents, and the example project
• Architecture and request lifecycle
• Project configuration
• Pipelines and workflow transitions
• Tasks and runtime orchestration
• Work in an existing Pi, Claude Code or Codex session
• Skills and provider credentials
• Release process

Baton currently runs agents locally; remote execution targets are not implemented.