Leopold
Brief it like a teammate. It conducts Claude Code in your seat.
Leopold is an autonomous orchestration harness for Claude Code. You sit down and debate the work with it the way you already debate with Claude Code: goals, constraints, taste, what "done" means. That conversation becomes a durable brief. Then Leopold takes the seat and drives Claude Code continuously, deciding the way you would, instead of stopping to ask you at every fork.
It is built to keep going: think, research, wait on long tasks, pick the next item, decide, log the decision, repeat, until the plan is done or a stop condition fires. Your git stays locked the whole time.
The name is a tip of the hat to Bugs Bunny. In Long-Haired Hare (1949), Bugs takes the podium disguised as the great conductor Leopold and runs the whole orchestra with a wave of the baton. That is the job: you are the composer, Leopold is the conductor, Claude Code is the orchestra.
Status: alpha. The in-session engine (skills + hooks) works today. The external SDK driver (packages/driver/) is built and typechecks against the Agent SDK (alpha). See Roadmap.
What is a harness, and why does this one matter
In current usage, an agent harness is everything wrapped around the model except the model itself: tool execution, memory and state, orchestration, guardrails, and observability. Put simply, Agent = Model + Harness. A good harness is what turns a general LLM into a system that can run long, complex workflows reliably, and it can be the difference between a model that quits after one turn and one that finishes the job.
Claude Code is already a strong harness for a single interactive turn. Leopold adds the layer it is missing for unattended, long-running work:
- a decider (your charter) so the agent has the authority to choose instead of asking, and
- continuity (a stop hook) so a finished turn rolls into the next item instead of halting,
- behind guardrails (a tool-call gate) so autonomy never touches your git.
The problem
A normal Claude Code session is a conversation. It pauses at every decision: "approach A or B?", "should I commit?", "do the next item or stop?". That is the right default when a human is watching. It is the wrong default when you want a session to run for an hour while you do something else.
The pauses come from three different places, and each needs a different fix:
- Safety defaults — commit, push, and destructive actions ask first. Correct, and Leopold keeps them locked.
- No designated decider — "approach A or B" is a product call. Without your judgment encoded, the agent has no authority to choose, so it asks.
- No continuity — when a batch finishes, stopping is the default. Nothing tells the session to pick up the next thing.
Leopold flips levers 2 and 3 while respecting lever 1.
How it works
Two phases.
Phase 1 — Brief (/leopold-brief)
A structured debate, not a form. You talk through the mission; Leopold pushes back, asks the sharp questions, and writes four durable artifacts:
MISSION.md — what we are building, why, and the definition of done.CHARTER.md — your decision charter: priorities, taste, hard never / always rules, risk tolerance. This is the part that "becomes you."GUARDRAILS.md — what is autonomous vs gated, stop conditions, budgets, the kill switch.PLAN.md — the backlog the autonomous run will burn down.
You can debate, revise, and re-run this until the brief reflects how you actually think. The quality of the autonomous run is capped by the quality of the brief, so this phase matters.
Phase 2 — Run (/leopold-run)
Leopold enters autonomous mode and conducts Claude Code in a loop:
- Read
PLAN.md, pick the next unchecked item.
- Do the work, reaching for the right gstack skill for the situation (
/spec before building, /code-review after, /verify to confirm, /investigate when stuck).
- Hit a fork? Consult
CHARTER.md. If the call is reversible and the charter is clear, decide, log it to DECISIONS.md, and keep going. Only a genuinely irreversible and ambiguous fork stops the run.
- Mark the item done, pick the next.
A Stop hook is what makes it continuous: when Claude finishes a turn, the hook checks the plan and the stop conditions, and if work remains it re-injects "continue". A PreToolUse hook keeps commit, push, and destructive commands locked even while autonomous.
Every decision Leopold makes on your behalf lands in DECISIONS.md, so you can review what "you" decided while you were away.
Why it is built on gstack
