independent · provider-neutral · open source

Your agent wrote it. DiffOwl reviews it.

The independent verification layer for agent-written code. Reviews every commit on the model you already use — whichever agent (or human) wrote it. No account, no dashboard, no per-seat bill.

npm i -g diffowl See how it works
~/your-repo
$ git commit -m "feat: agent patch"
commit complete — review started
$ diffowl review
2 findings stored — .diffowl/state.db
The case for it
The author shouldn't be the only reviewer.

Four reasons — each provable today, not a roadmap promise.

01

Independent eyes

The agent that wrote the patch is biased toward approving it. DiffOwl is a separate review loop — implementer and reviewer stay distinct roles.

02

Nothing to sign up for

No account, no dashboard, no third party in the loop. The CLI runs on your machine and talks straight to the model provider you already use — DiffOwl adds nothing in between.

03

Any model, no new bill

Rides on your existing OpenCode providers — Copilot, OpenAI, Ollama. No separate keys, no per-seat SaaS. You own the cost-vs-quality dial.

04

Findings that remember

Stable fnd_* IDs in SQLite, deduped across commits, with a real lifecycle: fix, dismiss, defer, reopen on regression.

The closed loop
Most tools leave a comment. DiffOwl remembers the outcome.

Reviewer and implementer stay separate — and every finding has a lifecycle.

01

Agent writes

A person or any coding agent produces a patch and commits.

→
02

DiffOwl reviews

Bounded local context, structured findings with confidence.

→
03

Agent resolves

Confirmed issues fixed; dismissals recorded with reasons.

→
04

DiffOwl remembers

Dismissed findings stay quiet on re-review; regressions reopen the original finding automatically.

Sixty seconds to first review
Install. Init. Review.
bash
# install globally
$ npm install -g diffowl
# pick a model, write config
$ diffowl init
.diffowl.yml created
# review the latest change
$ diffowl review
report → .diffowl/reviews/latest.md

  1. Bring your own provider

    DiffOwl uses OpenCode's existing credentials — opencode connects Copilot, OpenAI, or local Ollama.

  2. Hook it into git

    Reviews fire automatically on every commit and stay out of your way — read the report whenever you're ready.

  3. Hand off to an agent

    Install the portable diffowl-resolve skill, then say "Resolve the latest DiffOwl review."

Honest comparison
Different from hosted PR review.
Capability DiffOwl Hosted bots Agent self-review
No SaaS sign-up required Yes No (Their platform) Yes
Choose any model Yes Vendor-locked Same model
Independent of author Yes Yes No
Persistent findings Yes Partial No
Account required No Yes No
Rate limits Your provider's Free tier capped Your provider's
Subscription model None (Your keys) Per-seat SaaS None (Agent keys)

DiffOwl is built around practical control: local execution, model choice, and findings that stay traceable after the first review.

Where it's going
What's next?
01
Shipped

Reliable runner

Stable hooks, context, execution.

02
Shipped

Durable findings

Stable IDs, SQLite, JSON, lifecycle.

03
Shipped

Measured quality

Replayable eval — precision, recall, cost.

04
In progress

Branch review

Review the branch, not the commit — plus a pre-PR gate.

05
Planned

Agent & CI loop

Impact graph, MCP, SARIF, GitHub Action.

Stop merging code nobody reviewed.

An independent reviewer for the code your agent wrote — local, on the model you choose, with nothing to sign up for.

$ npm install -g diffowl
Star on GitHub Quickstart