magi

magi is an open-source terminal coding agent built around one hard question: when is a turn actually finished? Most agent loops end the turn the moment the model stops calling tools. That is usually wrong in both directions, producing agents that stop three quarters of the way through a task and agents that never stop at all. magi changes the ending into an act that has to be justified. When the agent declares it is done, a council of three auditors each reads the full record of what actually ran, what it really exited, and what changed on disk, then votes done, reject, or abstain. Only when the evidence backs the claim does the turn end; otherwise the feedback becomes the next instruction and the agent keeps working.

Key Features

• Council verification loop: three default auditors (named after the MAGI) judge the same record with different lenses and an order of search.
• Evidence-based requirements walk: each member marks each requirement SATISFIED, UNSATISFIED, or NO-EVIDENCE, settled by a verbatim fragment of a tool result rather than the agent's own account.
• Replayable record: every step is captured, so an inspected loop can be replayed and understood rather than assumed.
• Explorable agents: magi runs and supervises several agents at once, both from the terminal and a companion web console.
• Single binary: written in Go, CGO-free, with no heavy runtime dependency.

Use Cases

Who Should Use This Tool?

• Developers who distrust 'looks done': anyone whose agents routinely stop just short of the goal, or never stop.
• Teams operating autonomous agents at scale: where a half-finished or runaway run costs real money and time.
• Research and evaluation: the council gives a structured, inspectable answer to "did the task actually complete".
• People building agent loops who want a principled termination condition rather than a heuristic.

Problems It Solves

1. Premature finish: agents stop mid-task because the model ran out of tool calls. magi forces evidence before the turn ends.
2. Loops that never end: agents keep going because nothing asks them to justify stopping. magi makes completion an explicit, contested claim.
3. Untrustworthy self-reporting: a model saying "I did it" proves nothing. magi settles each requirement against recorded tool output.

Pricing

Advantages & Unique Selling Points

1. Principled termination: the finish is decided by evidence, not by the model's own stop signal.
2. Three independent lenses: correctness, verification, and completeness are each walked before anyone can vote, so no single missed detail slips through a unanimous verdict.
3. Open, inspectable, replayable: the record is a first-class artifact you can audit.

What Makes It Stand Out: the "council that refuses a finish" is a genuinely different answer to the agent loop's end-of-turn problem, and it is public in Apache-2.0 so you can read exactly what it does.

Getting Started

Quick Start Guide

1. Build or download the magi binary (Go 1.26+, CGO-free single binary).
2. Run it in a terminal and give it a task.
3. Watch the loop: agent takes steps, then declares "council: complete".
4. Let the auditors vote: the record is walked line by line against the requirements.
5. Accept or continue: if the council rejects the finish, the feedback becomes the next instruction and the agent keeps going.

Integration

Integrates with:

• Terminals on macOS and Linux
• Any model/CLI that can be driven from a Go binary
• A companion web console for monitoring several agents at once

Frequently Asked Questions

Does magi replace models?

No. magi is an agent harness that changes how a run is declared complete for whatever model you drive through it.

Is it only for serious agent users?

Not necessarily, but the value shows most where a wrong "done" is costly, such as autonomous runs or one-shot benchmarks.

Is it free?

Yes. It is open source under Apache-2.0 and runs entirely on your machine.

Alternatives

If magi is not the right fit, consider these alternatives:

• Claude Code: more polished for everyday development, but with the conventional implicit end-of-turn.
• Codex / Cursor: strong for interactive coding, less focused on proving completion.
• OpenHands: good if you want a full agent platform rather than a termination-focused harness.

Tips & Best Practices

1. Write requirements as explicit lines: the auditors walk "one line per thing the task asked for", so vague requests produce weak checks.
2. Treat NO-EVIDENCE as an answer: a missing result is still information; do not pad the record with the agent's self-report.
3. Use the console for supervision: running several agents at once is where the council's consistency saves the most time.

Conclusion

magi attacks the most frequently ignored part of an agent loop: knowing when the work is truly done. By making completion a contested, evidence-based decision rather than a silent stop, it gives agent users a termination condition they can trust and inspect.