Talk
Work with a planning agent that can inspect the repository but cannot edit it.
Benevolent Delegator for LLMs
BDFL is a terminal supervisor for AI agents. Create versioned plans, review diffs, approve whole plans or individual sections, delegate work to isolated workers, review every result, and stay in control.
One supervisor. Your choice of agents. More coming soon.
01 / THE IDEA
Use Codex, Claude Code, or Ollama independently for planning and worker roles. Work with a planning agent, compare and approve versioned plans or individual sections, then let isolated worker agents implement the approved work. 02 / IN THE TERMINAL
Manage running sessions or resume closed ones. Navigate to any worker from the bottom rail to follow progress or respond when it needs attention.
This demo is live. Click the top actions and agent rail.
03 / WORKFLOW
Role-specific bdfl
MCP tools and the bdfl-plan
skill keep planning and verification read-only while workers edit only isolated worktrees.
Work with a planning agent that can inspect the repository but cannot edit it.
Define shared decisions, owned paths, dependencies, locks, local checks, and global validation.
Review plan versions, compare diffs, and request revisions where needed.
Approve exact sections. Execution stays blocked until every section is approved.
Build eligible chunks in isolated branches and worktrees where the dependency graph allows.
Review each worker’s actual diff and checks. Accept it or send feedback.
Verify the consolidated result with global checks and a fresh non-implementing agent.
Integrate only after final review while the frozen target remains unchanged and clean.
04 / CONTROL SURFACE
Role-specific bdfl
MCP tools constrain each agent: planners and verifiers do not edit; workers edit only isolated worktrees.
Browse durable plan versions, compare adjacent versions, inspect exact changes, and approve individual sections while preserving unchanged approvals.
Use Codex, Claude Code, or Ollama independently for planning and worker roles, with separate models, effort levels, and options. Each worker receives isolated, focused context. Independent chunks run in parallel while prerequisites and named locks keep conflicting work in order.
Inspect each worker’s summary, diff, changed paths, checks, and commit metadata. Accept the result or send feedback to the same worker.
Each repository owns its ignored .bdfl/
runtime state. BDFL does not publish local state, metrics, analytics, or logs; use Ollama with a local model for a fully local setup.
05 / GET STARTED
Install globally, start BDFL in a Git repository, then choose separate planning and worker agents, models, effort levels, options, and worker capacity.
$ npm install --global @thisisnsh/bdfl
Requires macOS or Linux, Node.js 20+, Git, and at least one supported agent CLI installed and authenticated.