Agents now write code faster than anyone can read it. Onus is the layer every coding agent works through. It reads each change for what it means, routes it by risk, and approves it on evidence.
v0.11.0 · Open source, Apache-2.0 · One binary for macOS, Linux and Windows
A careful reviewer covers about 400 lines in one sitting.1 When one company pushed to double merged pull requests, output did double, and review couldn't keep up. Human review fell from 89% to 68% of pull requests.2
Growth after the push to double output.2
Onus reads a change the way a reviewer wishes they could: as a few shifts in meaning, most important first. 1,450 changed lines become four rows. One needs a person.
Four merged pull requests from Twenty, an open-source CRM with 31,000 TypeScript files. These are Onus's reports, row for row, with long file lists shortened. Nothing in the repository was run.
+1,711 −337 in 68 files→13 rows, 3 for a person
STANDARD_OBJECT_FIELDS changes its agentChatThreadParticipant value2-46-workspace-command-…-add-agent-chat-thread-subscriptions.command.tssubscribeToAgentChatThread and unsubscribeFromAgentChatThreadtwenty-client-sdk (3 files, 56 changed lines)twenty-serverand 8 more quiet rows
Before the accuracy work, the migration and the two new mutations were buried in a changed-lines row. Now each one asks for a person.
+291 −417 in 24 files→3 rows, 1 for a person
twenty-server: record-access-policy.service.ts, permissions.utils.ts and 3 moretwenty-server removed with the code they testedtwenty-serverA permission refactor across the ORM used to be one row: “697 changed lines”. Now it goes to a person, and the deleted tests are explained instead of flagged as weakened.
+23,718 −24,519 in 183 files→4 rows, 1 for a person
twenty-server: application-oauth.module.ts, token.module.ts and 12 moretwenty-client-sdk (3 files, 47,359 changed lines)twenty-serverand 1 more quiet row
47,359 regenerated lines fold into one quiet row; they used to raise four false “contract changed” alarms. Reviewers called the auth row borderline: these files only wire modules together.
+562 −56 in 14 files→5 rows, none for a person
isCompositePropertySupportedInGroupBy moved from twenty-server to twenty-sharedtwenty-sharedtwenty-serverand 2 more quiet rows
No row needs a person. The tests that disappeared moved word for word with the function they test; Onus used to call them weakened.
How these were checked, and what is still weak: the report accuracy audit.
Components, contracts, owners and risk labels, rebuilt on every change. Agents ask the map instead of reading 50 files and guessing. This change adds one edge that matters: customer phone numbers now reach an outside SMS provider.
What depends on UserPreferences?notifications, billing
Every task runs with a short-lived token that opens only the doors it needs. To get more, an agent attaches evidence, not a persuasive paragraph. A structured way to escalate cut test-gaming from 23.6% to 5.3% across eight frontier models.3
Token for “Text customers when their order ships”
Each change is routed by the rows of its report. Routine work merges on evidence alone, the judge takes the middle, risky changes go to a person, and some never merge as they are.
You write the rules in onus.yaml. Floors no rule can lower keep risky changes with people, and a learned model can only move a change up a lane, never down.
The judge re-derives the facts instead of giving an opinion. It never reads the author's reasoning, since models favor work that looks like their own.4
Every pull request leaves a record: which agent made it, the lane it took, whether it merged, and whether it was later reverted or caused an incident. CI keeps it on a branch of your repository. An agent setup may auto-merge only after 10 merged changes, and an incident among its last 20 takes that away until its record is clean again.
Agent setups
| Setup | Merged | Auto-merge |
|---|---|---|
codex | 12 | eligible |
claude-code/opus | 9 | recent incident |
devin | 4 | needs 6 more |
person/maria | 31 | eligible |
/onus approve new-external-service:external:acme-sms@notificationsa person approves one row; the lane updates/onus audit miss the opt-out was ignoreda human audit of something that merged on its own/onus incident SMS sent twice involved: notificationswhat broke, and where the next tests should goPeople add what CI cannot see with a comment on the pull request; only people with write access are recorded. The agent is read from the branch name, the bot that opened it, or an Onus-Agent:trailer. Agents can never write their own record.
Onus isn't an agent and doesn't pick one for you. Agents work through it over MCP, or a CLI if they can run a shell command. Each gets the same scopes, lanes and judge, and earns its own track record.
claude mcp add onus -- onus mcp# any MCP client: onus mcpWhat a change would break, who uses it, and which tests cover it.
onus_impactonus_dependentsonus_tests_foronus_invariants
The changes in meaning of its own work, with a checklist of what to verify.
onus_check
Tests run in an environment, a submission with that evidence, the judge's verdict.
onus_env_runonus_submitonus_judge
Rights its token lacks, asked for with evidence and decided by policy or a person.
onus_escalateonus_escalation
Agents cannot approve themselves: they cannot record outcomes, grant escalations, mint tokens or choose the scope they are checked against. Pull requests are attributed to Claude Code, Codex, Copilot, Cursor, Devin, Jules and Aider by their branch or bot, and to anything else by an Onus-Agent: trailer.
Swap every agent overnight and nothing in Onus changes. Only the track records start fresh.
onus ui opens the whole of Onus for the repository you are in. It ships inside the binary and serves on 127.0.0.1 behind a per-session token, so nothing is installed and nothing leaves your machine.
Measured on Twenty, an open-source TypeScript monorepo with 31,000 files, 72,000 symbols and 272,000 dependencies, on a laptop. Each method is published, including the results that went against us.
Is the map right?
100 sampled symbols, 2,584 files. Most of the extra files are barrels that re-export a symbol: they break when it is renamed, though the compiler does not count them. Method of the references benchmark
Are the reports right?
| First audit | After fixes | |
|---|---|---|
| Rows with a wrong fact | 7% | under 1% |
| High-severity misses, in 20 pull requests | 9 | 1 |
| Flag precision, 1 to 5 | 2.05 | 3.5 |
| Usefulness, 1 to 5 | 2.1 | 3.0 |
Every kind of mistake found became a regression case that runs with the test suite. Method of the report audit
Is it fast?
One map server per repository, shared by every agent and worktree; facts are cached by file content. Method of the speed figures
What didn't work
In a blind A/B test on a real Twenty feature, agents with the map tools took 231 s and scored 7.5 of 10; without them, 235 s and 7.8. They explored with grep anyway. What they did use, unprompted, was onus_check before finishing. A second round with a sharper check did no better (6.3 against 8.0): a clean check read as “done”, so every answer now says what it does not verify. Onus puts its effort into checking work, not guiding it.
Claude Code with Sonnet, three runs per condition, graded blind against the shipped change. Method of the agent A/B test
Everything on this page is in one binary. Reports only read files and send nothing anywhere; Onus runs your code only when you ask it to, in a container.
curl -fsSL https://onushq.com/install.sh | shonus report --base main --head HEADonus uimacOS and Linux. Installs to ~/.local/bin after checking the SHA-256, or use Homebrew: brew install onushq/tap/onus. onus ui opens all of it in your browser. The guide is also built in: onus help.
permissions: contents: read pull-requests: write # for the commentsteps: - uses: actions/checkout@v7 with: fetch-depth: 0 - uses: onushq/onus/action@v0.11.0 with: fail-on: secretsThe GitHub Action posts the report as one comment, updated on each push, and fails the job only on what you list in fail-on. Add records: true and it also keeps each agent's track record (permissions and events).
Each archive holds onus and the Svelte plugin, and carries a build attestation: gh attestation verify <archive> --repo onushq/onus.
Onus comes from onus probandi, the burden of proof. Today it sits on a reviewer reading lines. Onus moves it to the agent, which has to show its work.