AI assistant
AI changes land on their own branch, for review
The assistant reads your repository and proposes changes. Its writes go to a dedicated branch and open as a draft review request; a person approves every merge, your default branch stays untouched, and secret values never reach the model. Every action is metered and logged.
The eight workflows we ship
Deliberately scoped: each one solves a real task, and each one is audit-logged.
Ask a repository question
Natural-language Q&A over the repository's contents.
Explain a file
A plain-language explanation of what a file does.
Explain an error
Paste an error, get a contextual explanation grounded in your code.
Generate or update a README
Drafted from the actual repository contents, not boilerplate.
Propose a change
The assistant drafts a constrained edit plan for you to review first.
Create a change on a branch
The edit plan is committed to a dedicated AI branch and opened as a draft review request; a human merges.
Review-request summary
A plain-language summary of what a review request changes.
Usage metering
Every AI action is recorded, counted against plan credits, and visible to the whole workspace.
The safety model
AI in the loop is a careful teammate: it drafts, and a person decides. These constraints are enforced server-side.
Branch-only writes
The assistant can never write to your default branch. Its commits land on dedicated keelson-ai/* branches and always open as draft review requests, subject to the same branch protection as everyone else.
Human approval, always
No AI-created change reaches your main version without a person reviewing and merging it. This is an invariant, not a setting.
Secret values never reach the model
Every provider call passes through redaction middleware. Vault values stay encrypted and are never included in model context.
Bring your own key
Use your own Anthropic or OpenAI-compatible API key on any plan, including Free. BYOK usage is not metered against platform actions.
Usage caps and visibility
Platform-funded actions are metered per plan with warnings before limits. Hitting a cap blocks the action with a clear message — never a surprise charge.
An off switch
The assistant sits behind a global kill switch, and workspaces can disable it entirely. Autonomous multi-step agents and direct-to-production writes are deliberately not built.
Every plan includes AI — Free gets platform actions, and BYOK works everywhere.
Anthropic and OpenAI-compatible providers are supported.