CamusTrust the work, not the model that made it.
Camus is the local control plane for AI work you need to verify. Choose one model or coding agent to make it and another to challenge it against your contract. Camus runs the checks you configure, keeps control local, and binds the evidence to the exact result before you decide to ship.
Models are multiplying. Accountability isn’t.
Claude, GPT, Grok, Qwen, SWE, open weights, and native agent harnesses improve on different curves. Switching models is easy. Proving what ran, under which budget, against which exact artifact, is not.
Better models still miss their own mistakes.
Capability improves the work. It does not turn confidence into evidence.
A harness changes more than the model.
Context, tools, retries, and side calls can change the result and its cost.
Review can detach from the final artifact.
A clean verdict means little if the work changes after the reviewer saw it.
Change the models without changing what “trusted” means.
Camus keeps the contract, artifact, identities, evidence, verdict, and human decision bound together while the model layer changes.
Camus records the fix—and refuses to launder the failure around it.
In one bounded dogfood run documented in the public model-setup report, Qwen3.8 Max through Camus file actions made the exact fix, the host verifier passed, and GPT-5.6 Luna approved it. The same fixture through native Qwen Code retained the right-looking change but never closed the harness session, so its sealed standing remained failed.
Qwen3.8 Max made the canonical one-line fix in five responses and six actions.
The frozen host check passed against the exact candidate.
A separately selected Luna reviewer approved with no findings.
The native path kept a correct-looking edit but no definitive terminal, so Camus did not upgrade it.
The useful result remained advisory.
task_class simple_bounded_code
maker qwen/qwen3.8-max
executor file_actions
reviewer codex:gpt-5.6-luna
verification passed
review approved_no_findings
human_accept required
maker_time 83.8s
maker_tokens 7,445The fixture and report are public; the underlying run receipt remains private, so this is a documented case rather than independently replayable evidence. It is not a model ranking. Inspect the fixture ↗ and the dogfood report ↗.
Four parts. One honest result.
Camus is not another agent. Models handle semantic work; a deterministic local kernel owns state, budgets, Git custody, recovery, and evidence. No model gets to award itself trusted standing.
Set the contract
State the goal, what must be true, the permitted knowledge, and the budget.
Let the maker work
A qualified model or native coding harness owns the semantic work. The local kernel handles the plumbing.
Challenge it separately
The selected reviewer tries to break the artifact. Recorded identities determine whether that review is independent; configured tests and sources arbitrate what they can.
Decide and seal
Ambiguity goes to you. Every verdict binds to the exact artifact it certified.
You are not removed from the loop. You are removed from babysitting it.
What is trusted today—and what remains evidence-gated.
Camus applies versioned contracts across supported paths. Each exact path earns only the standing its checks, identity, and evidence support.
Trusted path
Claude → Codex proof gate
Fixed cross-vendor roles, isolated Git custody, HEAD-bound verification, bounded recovery, and a review bound to the exact candidate.
Advisory path
Flexible Build
Choose an available maker and reviewer separately, then use Camus file actions or an eligible native Codex, Qwen Code, Grok Build, or Devin harness. The candidate stays advisory and never lands without human acceptance.
Bring your Grok login or choose SWE-2 High through Devin, with a separate reviewer. SWE supports bounded coding work on Apple Silicon; internal token spend is unknown and requires explicit consent. Check SWE setup and limits ↗
Evaluation infrastructure
Automatic routing stays off
Matched evals and blinded human calibration can accumulate task-class evidence. No model is promoted and no universal “best model” is claimed until that evidence earns it.
The quality floor comes first. Only then does Camus optimize tokens, time, and cost. Cheap failure never wins.
One control plane. Two ways to work.
Camus starts with code, where trusted tests can arbitrate; the same control plane extends to evidence-heavy research and writing. Use the CLI for repositories or the browser for documents and investigations.
For developers
Camus CLI
Give an agent a real repository without giving up custody. Work stays isolated, reviews bind to exact candidates, and tests arbitrate what they can. Flexible Build remains advisory until you accept it.
- isolated worktrees and controlled merge
- deterministic feature state and bounded budgets
- maker and reviewer selectable independently
- HEAD-bound verification
- provider-free receipt inspection
For research and marketing
Loop Studio
Write a memo, investigate competitors, or turn Hivemind knowledge into a grounded deliverable. Use plain language, inspect every objection, and step in only for real judgment calls.
- browser interface, no JSON noise
- the full acceptance contract in your own words
- separate maker/reviewer choices for code, research, and marketing
- Claude, GPT, Grok, Qwen, compatible APIs, and SWE via Devin for code
- built-in Claude/Codex seats use versioned qualification contracts; configurable seats require exact local qualification
- native harness artifacts separately prove readiness and policy compatibility
- Hivemind grounding through your Claude MCP
- explicit publication consent and sealed receipts
The control plane stays on your machine.
Orchestration, budgets, run state, and receipts stay local. Configurable API credentials are held by the local service and sent only to the selected provider endpoint; they are never sent to camus.sh or exposed to native workers. Providers still receive the context you choose to send. Native Qwen and configured API-backed Grok workers can reach only the selected model through a host-owned one-model gateway. The built-in Grok subscription seat keeps Grok Build's own login and inference route, strips API keys, pins the reviewed artifact, and applies Camus's bounded tool policy. Devin similarly uses its saved login for SWE coding work, with no API-key fallback. Its internal inference totals are unavailable: Camus limits time and observed actions, not an unobservable token bill.
$ npm i -g camus-cli@latest
$ camus models
$ camus build --maker <backend>:<model> \
--reviewer <backend>:<model> --task "..." --contract "..." \
--verify "pnpm test"
✓ selected seats authorized for launch (built-in or qualified)
✓ candidate isolated from your branch
✓ trusted verifier passed
! clean advisory review; human acceptance still required
$ camus build --inspect <runId>
✓ receipt inspected without a provider callBring one real task. Leave with a receipt—or an honest refusal.
Camus is in public alpha. Run it yourself, or bring one bounded code, research, or launch-critical content workflow to a design-partner session. We will record what worked, what stopped safely, and where human judgment was actually needed.
Do not post credentials, private source, raw diagnostics, or unreviewed receipts.
No intelligence should be the only judge of its own work.
Models will change. The need for independent judgment, inspectable evidence, and a meaningful human decision will not.