Babel Harness · Pre-1.0

A coding agent that shows its work.

Run. Verify. Understand.

A local coding agent with inspectable instructions, explicit permissions, and visible verification results. Start with Chat; use Plan or Deep when you need more structure.

Get started View GitHub

Ordinary coding-loop reliability is under active qualification — seestatus. These are product positioning statements, not proof that every edit has been independently verified.

You
  Fix the failing webhook retry test.

Babel
  ▸ Read src/webhooks/handler.ts
  ▸ Edited src/webhooks/handler.ts
  ▸ npm test -- webhook
  ✓ 18 passed

  Completed · 1 file changed · verified

Reference preview — an illustrative walkthrough, not a live run. See the product tour.

How work happens

Inspect, change, check, and review the evidence

Inspect

Read the repo, the applicable instructions, and the permission boundary before anything moves.

Change

Make the smallest correct change inside the approved scope.

Check

Run the strongest practical check for the touched surface.

Review evidence

See what was run, what passed, and what remains unverified.

Modes

Chat first; more structure when you need it

Daily default

Chat

Everyday interaction with visible tool activity and approvals.

Plan

Planning before mutation. Structured intent, without executing work.

Deep

Additional workflow structure for bigger tasks — not a blanket reliability upgrade.

Current state

What exists today, and how far it is verified

Released

Chat

Chat entrypoint and TUI are present in the public tree as the daily interaction mode.

Ordinary coding-loop reliability is under active qualification.
Provider-backed sessions require local credentials and a working provider; quality varies by model.

Released

Plan

Plan mode exists as a CLI/TUI entrypoint for planning before mutation.

Planning behavior boundaries are documented in the product source.
No live workflow qualification is claimed.

Released

Deep

Deep mode exists as a higher-ceremony path intended for riskier work.

Still maturing; route availability is not a reliability ranking.
Deep is not a blanket reliability upgrade over Chat.

Released

Desktop harness

Windows Desktop bundles the real Babel CLI with provider credentials, enabled/default model selection, project files, approvals, persisted history and verified sandbox setup.

Windows packages are unsigned; checksums verify bytes, not publisher identity.
Automatic Desktop updates are unavailable.
Active model-task shutdown on quit and cross-version profile migration remain unqualified.
Plan and Deep currently require OpenRouter; other Desktop providers offer Chat.

See status for the full capability snapshot with evidence links, and theproduct sourcefor the pinned revision behind this summary. Published Windows downloads, checksums and signing status are listed onGet Babel.