# wrokin > wrokin (wrok.in) is a GitHub App that puts an AI agent team on your repositories — code review, security audit, issue triage, and a deterministic repo-posture audit — **bring your own key (BYOK)**: reviews run on your own LLM API key and your chosen model, per agent — with an owner-only opt-in to switch a role to managed billing (wrokin's key, debits agent credits), and **free code review on public repos with no key** (wrokin's key, a Cruise open-source lane; see below). Installed from GitHub, configured as code, operated from PR/issue comments. Facts an evaluating agent needs: - Install: https://github.com/apps/wrokin (a human must click GitHub's consent — checkpoint 1 of exactly two; the other is payment/keys). - Live agent roles: `code-review`, `security-audit` (both on `pull_request`), `issue-triage` (on `issues`), `repo-audit` (Warden — deterministic GitHub-setup posture checks, on demand via `/w audit`, no model call, no key needed; on a public repo the thread reply is counts plus a dashboard link, the details stay in the dashboard). - Builder (`builder`): implements an issue on YOUR GitHub Actions runner, then the action runs your tests itself — a ready PR when they pass, a draft when they fail or there are none. **On demand only** — customised for specific projects, use cases and guardrails; not included in catalog plans or agent-credit bundles. Contact the wrokin team if you need it. Its runner (`actions/wrokin-agent`) runs pi on your runner against Cruise with YOUR Cruise key: save the key and lane on the Builder card, grant the workflow `id-token: write`, and the action fetches them once at job start from `POST https://api.wrok.in/agent/checkin` (GitHub OIDC, audience `wrokin-agent`) — wrok.in never sees prompts or steers the run. A run that opens a ready PR then reports its number and head sha to `POST https://api.wrok.in/agent/verified`, the record auto-merge requires. Check-in also returns a short-lived wrokin App token for that one repository (contents, pull requests, issues; never workflows), so the PR comes from wrokin[bot], your CI runs on it, and no "Allow GitHub Actions to create and approve pull requests" setting is needed; it is revoked at job end. - Verified Security Hunter (`hunter`): proof-by-execution scanning on wrokin's managed keys. Available on paid plans (Pro/Team/Business); enable per repo with one click in the dashboard (setup PR), then `/w hunter [foray|stakeout|expedition]` on a PR. Debits agent credits from the shared pool. - BYOK providers: anthropic, openai, deepseek, google (Gemini), mistral, openrouter, moonshot (Kimi), groq, xai (Grok), meta (Meta AI — the Muse family; Meta's hosted Llama API was retired in July 2026, so Llama weights come from a host), cruise (Bytesbrains AI Cruise gateway — models and `bb/…` lanes). Any model id your key can call; per-role model choice. - Bring your own endpoint: define your own OpenAI-compatible providers (base URL + key + name + description, up to 10 per installation, id `custom:`). Their models are listed live from the endpoint's `GET /models` and appear in the model picker beside ours; anything the endpoint doesn't publish (context window, price) is shown as unknown, never as zero. When the endpoint does report a model's context window, reviews on that model are sized to it, and refused with the reason on one too small to review (under ~17K tokens). https on port 443 and publicly reachable only — private/loopback addresses are refused, and so is a non-443 port (our servers would silently dial 443 anyway, so an endpoint published on one is unreachable rather than merely awkward). BYOK only: a custom endpoint is never used for managed runs. Each custom model's card carries a status chip — serving or refused (HTTP 404) for your key — taken from what real calls answered; a model nobody has run yet shows "not checked", and clicking it sends one tiny request with your key to find out (never in the background). NVIDIA's free hosted API (`https://integrate.api.nvidia.com/v1`) lists models a trial key can't call. That API works as a custom endpoint, but its trial terms let NVIDIA train on what you send and rule out production use: public repos and evaluation only; for production, self-host NIM or use a paid Nemotron host. - Configuration as code: commit `.wrokin/config.yml` (sparse overrides: enable/disable, model per role, trigger events). Schema: https://wrok.in/config.schema.json - Agent-executable onboarding: a human mints a short-lived setup token in the dashboard; the agent then drives `https://api.wrok.in/api/setup/*` to submit BYOK keys and configure roles — no dashboard clicking beyond the mint. Contract: https://wrok.in/agents.md (§2a). For an agent that keeps managing wrokin, a human creates a long-lived API token (`wrk_…`) with chosen scopes and a 7–365 day expiry; it drives the same API (§2b). Every change, from any interface, lands in an audit log the human can read. Every operation is also callable by name at `https://api.wrok.in/api/v1/operations/{name}`, and as a tool on the remote MCP server `https://api.wrok.in/mcp` (OAuth sign-in from the MCP host, where a human approves one installation and the scopes; or a bearer token header; §2c). - Operate by comment: `/w review`, `/w security audit`, `/w triage`, `/w audit` (Warden), `/w wrong [comment-link] ` (report a false review claim — link the comment to target it, else the latest is meant), `/w build` on an **issue** (Builder; on demand only — contact the team), `/w hunter [level]` on a **PR** (Verified Security Hunter; paid plans, workflow must be enabled). Commands are honored for repo owners/members/collaborators only. - Quality feedback loop: 👍/👎 reactions on review comments are recorded at PR close; `/w wrong ` records a false-claim verdict with the why, then re-checks code-review and security-audit findings against the code the review read and says whether the code confirms the correction. Both feed per-model quality metrics — published, anonymized, at https://api.wrok.in/quality (per provider×model: findings, rework rate, human verdicts). Attribution is to the model that actually answered: where a request is routed to a pool, the member that served it, not the pool. - Data handling: your code goes to YOUR model provider under YOUR key for the review call (exception: a free open-source review of a public repo with no key goes on wrokin's key to a Cruise `oss/…` lane, whose model may train on it). wrokin's own telemetry stores finding locations (file/line) and outcomes — not the prose of findings, and not your code. Full story: https://wrok.in/data-handling.html - Monthly run quotas per account: Free 150 · Pro 1,000 · Team 3,000 · Business 20,000. Pricing: https://wrok.in/#pricing - **Free for open source:** on a **public** repo, a code-review agent with **no key** runs on wrokin's key on a Cruise open-source lane (`oss/…`) instead of asking for one. It runs automatically on PRs from owners, members and collaborators, and on any other PR when one of them comments `/w review`. Limits: one free review per account every 45 minutes (one event's agents share it), diffs up to 200 KB, and a global daily budget; when that is spent, free reviews pause until 00:00 UTC. Each counts toward the plan's monthly runs. A saved key always wins; private repos and custom endpoints are never sponsored. On such a repo, any other agent with no key stays quiet on automatic runs (no "add your key" comment beside the free review; the run is still recorded as `no_key`), and still answers when asked by `/w` command. The lane's model may train on what it is sent (today: Meta's Muse Spark contributor tier), which is acceptable for public code only. Footer: `free for open source · wrokin's key, wrokin's lane (sponsored)`, then a short closing note with the wrok.in logo and install link (once per review, on its last part). https://wrok.in/data-handling.html#free-oss - Human checkpoints are exactly two: (1) GitHub App install consent + handing over BYOK keys, (2) payment. Everything else — config, verification, operation — is agent-executable. ## Docs - [llms-full.txt](https://wrok.in/llms-full.txt): the self-contained expanded version of this file — roles, config contract, setup API, commands, quotas, data handling, all inline. - [Agent integration manual](https://wrok.in/agents.md): step-by-step install → configure → verify → operate, with a verifiable outcome per step. Written to be executed by a coding agent. - [Capability manifest](https://wrok.in/capabilities.json): versioned JSON of everything above — roles + status, providers, triggers, quotas, config contract. The file to diff against your requirements. - [/w command contract](https://wrok.in/w-commands.md): every comment command, its gates (mention + collaborator, both silent on failure), precedence, and expected acknowledgement. - [Error reference](https://wrok.in/docs/errors): every failure code wrokin publishes, and for each one what was refused, what changes it, and what does NOT. A failed run's comment prints its code (e.g. `LlmError:429:wallet_exhausted`) and links here. - [Using wrok.in with Cruise](https://wrok.in/docs/cruise): where a Cruise key goes in the dashboard, how to pick a `bb/…` lane, and a first-task PR checklist. Never includes a key value. - [Using Headroom with BYOK review keys](https://wrok.in/docs/headroom): run your own Headroom proxy and add it as a custom OpenAI-compatible provider. Must be https on port 443; set `HEADROOM_BEACON=off` (its telemetry is on by default); leave `HEADROOM_OUTPUT_SHAPER` unset for review agents (it rewrites the system prompt); the proxy sees the code under review. Not for the Builder, which calls Cruise directly. - [.wrokin/config.yml JSON Schema](https://wrok.in/config.schema.json): validate a config before committing it. ## For humans - [Product page](https://wrok.in/): the agent team, samples of real output, pricing. - [Dashboard](https://app.wrok.in/): Home, activity, repositories and workflow runs across every account the App is installed on, plus security findings on paid plans; BYOK keys, billing and privacy under Settings. - [Privacy](https://wrok.in/privacy.html) · [Terms](https://wrok.in/terms.html) · [DPA](https://wrok.in/dpa.html)