Branchwork

Branchwork

**See what your agents are doing — against the plan.** Agents orchestrate themselves; Branchwork is the layer they report to: shared plan state, scoped project practices served at the moment a task starts, and a live dashboard that shows declared progress next to observed ground truth (commits, PRs, CI). The agent drives. Branchwork knows, advises and shows.


Plans live as YAML in `~/.claude/plans/`. Any MCP-speaking agent — a Claude Code session and its sub-agents, a runner, Bob Shell — connects to the Branchwork MCP server and:

• declares where it is (update_task_status with its agent label + artifact links — PR, commit, CI run);
• asks for task context (get_task_context: the task's spec, prior learnings, and every project practice whose scope matches the files it is about to touch — your org's hard-won rules, injected exactly when they're needed);
• teaches (practice_add: promote a recurring learning into a standing, scoped rule).

Mark a plan `mode: observe` and Branchwork tracks a foreign orchestrator without driving anything — no spawn, no gates on your worktree, pure visibility. Practices are advisory by design: CI stays the enforcement layer.

**Managed mode still exists** for when you want Branchwork to drive: every task has a Start button that spins up a Claude agent on a dedicated git branch — you watch it work, type to it, review the diff and merge from the browser (Claude Code today, Aider/Codex/Gemini as drivers). Ships as a single ~15 MB Rust binary. No Node, no Docker, no daemon to install separately.

Screenshots

Plan board — collapsible phases, live status


Agent terminal — full Claude Code session in the browser


New plan — describe, pick a folder, an agent creates the plan


Sidebar — projects, driver auth status, effort lev