Private Investigator

• The GitHub App

  Install it and pick the repositories: nothing else to set up. Reviews are opt-in: comment /review on a pull request (write access) or press Re-run on its check. Opening or pushing alone does not start a review. Your code is fetched and read only inside an attested OutLayer enclave, running our published build. New installations get a 30-day trial with 40 reviews; each check says how many reviews and days are left. When the trial ends, the check points at upgrade.

  Install on GitHub

• The GitHub Action

  Runs on your own GitHub runner, which already has your code, with your own NEAR AI key, so not even an enclave of ours touches it. About ten minutes to set up the first time.

  Set it up

1. Attest

   Before any code is sent, check NEAR AI Cloud's evidence for a fresh nonce: Intel's signed quote for the model's machine, fully patched, and NVIDIA's signed verdict on its GPUs.

2. Encrypt

   Encrypt every request, tool calls included, to the key that evidence binds. Only the attested enclave can read your code; the gateway in front of it sees ciphertext.

3. Sign

   The enclave signs each reply over the exact request and response bytes, and a reply is used only once its signature checks out.

4. Receipt

   The evidence and signatures form a receipt, saved with the run. It holds hashes, not code, so anyone can check it.

With the Action, these steps run on your runner. With the App, they run inside an attested OutLayer enclave that fetches your code from GitHub itself, and the receipt adds each step's attestation, naming the exact build that read your code. Either way no MultiAgency server sees your code, and the source is public.

1. Step 1

   Get a NEAR AI Cloud API key

   Sign up at cloud.near.ai, add credits under Credits, and create a key under API Keys. Our test reviews cost about a cent each in NEAR AI credits, plus a few minutes of your GitHub Actions time.

2. Step 2

   Add it as the repository secret NEARAI_API_KEY

   Open the repository's secretsor run

   Copy

3. Step 3

   Add the workflow

   Create it on GitHubor save this as .github/workflows/private-review.yml

   Copy

4. Step 4

   Check it, then open a pull request

   On the repository's Actions tab, run private-review by hand: it checks the key, the model and the permissions without reading any code, and says what to fix. Then the review appears on each pull request in a few minutes, with findings on changed lines as inline comments. It follows your REVIEW.md and AGENTS.md from the base branch when you have them. The receipt is saved as the run's private-review-receipt.json artifact, ready to drop into the checker below.

• Event notices

  GitHub sends its notices of pull request events (titles, descriptions, who pushed) to our relay, which forwards only ids to the enclave, and holds those only while a review runs. Its source is public.

• Public receipts

  Each review's receipt is published in OutLayer's public storage: hashes, signatures, attestations, its count of new findings, and a commitment to the pull request that only the review's link opens. Never your repository's name, code or review.

• Counts and timing

  How many reviews each installation runs (for the free tier), and when. OutLayer sees when each step runs and how large its requests are.

Until then, the Action path still runs on your own NEAR AI key with no trial cap. Questions: open an issue.

• What happens on your runner

  The Action runs there. Read its source, pinned by commit in your workflow.

• That a build is its source

  The App's receipt names the build that read your code. Every release is recorded with its commit in app/builds.json, and rebuilds bit for bit from it (CI checks this); rebuilding it yourself is the step that closes the loop.

• The enclaves' platforms

  The proofs rest on Intel TDX and NVIDIA confidential computing, and for the App on OutLayer's keystore and its approved worker builds.

• Which model instance answered

  Instances share one signing key, so a signature proves that an attested enclave answered, not which one.

• What NEAR's images do

  The receipt records the compose hashes, and NEAR publishes the files they measure. Auditing those images is a separate step.

• That the review is right

  The proof covers where your code was read, not the quality of the findings. Measured on a real pull request, the private model missed two regressions in money paths that span files, which a frontier model caught (the evaluation). Use it as a second reviewer, not as the sign-off on payment or permission code.