FAQ

One model has one model's blind spots. atcr runs a panel of different models and providers, then a deterministic Go reconciler merges their findings. The value is the merge: a finding two independent models both caught scores higher than either model's opinion alone. You get one report ranked by agreement, not N walls of prose.

Those are single-model reviewers — most are hosted SaaS apps that read your PR and post comments. atcr is a local binary that runs a panel of independent models and scores their agreement, writing one auditable report to disk. The differences are architectural, not feature-by-feature. There's a full side-by-side on the Compare page, including where one of the others is the better call.

The models aren't; the reconciler is. Clustering, deduplication, confidence scoring, and report rendering are pure Go — same inputs, same output. The prompts orchestrate; the binary does everything that must be reproducible. Every cross-stage handoff is a machine-parseable file on disk you can diff and audit.

Only the model calls you configure, plus three disclosed surfaces. atcr is local-first and bring-your-own-keys: it has no account and no analytics backend. The review tree — payloads, per-agent findings, and the reconciled report — is written to .atcr/ in your repo. Three things can leave, and the privacy page discloses each in full: a usage ping carrying four allowlisted fields, on by default (opt out with ATCR_TELEMETRY=0); a --sync-cloud scorecard, uploaded only when you pass the flag and pseudonymous rather than secret; and an off-by-default community quality signal. At setup, atcr quickstart also downloads the community personas from GitHub unless you pass --offline, which scaffolds from the built-in personas instead. Run it offline against local models with the ping disabled, and nothing leaves at all.

Anything behind an OpenAI-compatible endpoint — OpenRouter, OpenAI, local Ollama, vLLM, and others. You bind each persona to a model in registry.yaml and set per-agent fallbacks. atcr doctor probes every configured endpoint before you spend a review.

Yes. The companion Agent Skill contributes the host model's own review as a +1 source, so even a single key yields two or more independent sources and a working confidence signal. Add more providers later and the panel grows.

atcr review --fail-on high runs review, reconcile, and a severity gate in one command, exiting non-zero on a confirmed high-severity finding. Point --output-dir at an artifacts path so the run never disturbs the interactive .atcr/latest pointer.

They're preserved, not flattened. When one panelist objects to another's finding, the reconciled report marks it disputed, records the objection, and lowers the confidence — instead of silently dropping it or pretending it was confirmed. The disagreement radar surfaces exactly where the panel split.

The nine default personas are editable markdown files under .atcr/personas/. Change a focus, swap a model, remove one, or add a specialist for a single PR. See Personas and Reconciler.

It runs the patch before it trusts it. An executor model generates a patch per eligible finding, atcr applies it to the working tree, then validates it with your validation command — a build by default, tests only if you configure them — running in an ephemeral, network-isolated sandbox against a writable copy, so no host file is ever mutated by unverified model output. A patch that fails validation is reverted from the working tree automatically. What survives is opened as a GitHub pull request: atcr never auto-merges and never commits autonomously, so a human still approves every fix. Two flags opt out of the safety net explicitly — --no-sandbox validates on the host, and --allow-config-edits lets patches touch CI workflows and other blocked config paths. See the auto-fix docs.

Findings are grouped by structure, not wording. Each one is mapped to the smallest AST block covering its line, and findings whose blocks match land in the same cluster — so the same bug survives the line-number drift between runs. Near-identical findings in a cluster then collapse into one canonical row, and the result is scored by how many independent reviewers agreed. The full mechanism — clustering, dedupe, and trust-aware confidence — is on the Reconciler page.