Braid

Coding agents that check each other

Give several agents one repo and they trip over each other's branches, and a green task can still break the tree once it merges.

Break a goal into a graph of tasks and run each one with a coding agent in its own git worktree. A task only merges when its checks pass, a reviewer agrees, and the checks pass again on the merged tree.

The Braid dashboard in the middle of a demo run: seven agents on a floor, one strand already tied into the integration branch. — Fig. 1 · A real run, screenshotDemo run, scripted agents

run · 4 agents

⑂schemamerged

⑂authretry

⑂errorsresolving

⑂routesreview

braid/…/integration1 merged, checks pass

Fig. 2 · Braid, drawn in codeIllustration, not a screenshot

Opening it the first time

On a Mac

This build is ad-hoc signed by me, not notarised by Apple, so macOS will refuse to open it on a double-click. That is Gatekeeper doing its job with an app it has never seen, and you only have my word for it, so the source is there if you would rather read it or build it yourself.

1Open the .dmg and drag Braid to Applications. Worth doing rather than running it from Downloads: macOS moves a quarantined app to a random read-only path, which breaks things in ways that look like bugs.
2Right-click it, choose Open, then confirm at the prompt.

Or from a terminal, once:

xattr -dr com.apple.quarantine /Applications/Braid.app

On Windows

The installer is not code-signed yet, so SmartScreen may say Windows protected your PC. Choose More info, then Run anyway.

What it does

Every task in its own worktree: Each task gets a worktree and branch started from everything already merged. Your branches and working tree are never written to; work lands on braid/<runId>/integration for you to review.
Checked three times before it lands: A task merges only when your own checks pass, an optional reviewer agent approves, and the checks pass again on the merged tree. A merge that breaks the tree is rolled back.
Conflicts get a narrow agent: A merge conflict goes to a resolver that keeps both sides instead of redoing the task, then the checks run again.
Bring the agent you already use: Claude Code, Codex, Gemini CLI, or a model on NVIDIA NIM. A reviewer can be a different agent from the one that wrote the code.
Stop any time, pick up later: Every agent call has a time limit and every task a token allowance. Rate limits and outages are retried without costing the task an attempt, and braid resume keeps merged work and retries the rest.
Watch it happen: A live dashboard shows every agent, check and merge as it happens. Notifications tell you when a run finishes or a task fails. braid pr opens a pull request for the result.

Specifics

Platform: macOS 14+ (Apple silicon) · Windows · CLI
Built with: TypeScript · Tauri
Agents: Claude Code · Codex · Gemini · NIM
Account: None
Telemetry: None
Network: Your agent's provider only

Shortcuts

⌘O: Open a repository
⌘N: New window
⌘,: Settings