Nothing on this page is a certification badge.

Named actions, named models, named retention windows. If a claim here costs us nothing to make, it would not buy you anything either.

Logs and control

Pro only

What is written down, and who can read it back.

Tool, arguments, result and elapsed time for each step. Arguments are never truncated — truncation is always where the interesting part was.

Named actions suspend the run and wait. There is no timeout that approves for you, and no configuration that removes the gate.

Every permission change, approval and refusal, as NDJSON, with the actor attached. Ships to your SIEM or sits in object storage.

Run the executor inside your own network. The control plane never sees the repository, only the run metadata.

Ninety days by default, anything from seven upward on Enterprise. Deletion is a hard delete, including from backups within thirty days.

Point it at your own model provider accounts. The permission layer does not change with the model.

Models

Pro only

Named per stage rather than described in the aggregate. Bring your own keys and none of it changes.

Stage	Model	Provider	Your keys	Retention	Sees
Planning	Claude Opus 5	Anthropic	supported	Zero retention	Sees the issue text and the file tree, not file contents.
Editing	Claude Sonnet 5	Anthropic	supported	Zero retention	Sees only the files the plan named.
Classification	Local rule set	Runs in your runner	not applicable	Never leaves	Regex and heuristics. No model involved.
Embeddings	text-embedding-3-large	OpenAI	supported	30 days at provider	Optional. Disabled by default.

Twenty runs a month, free, with the same permissions and the same transcripts as every other plan.