Agent-readable run protocol v1

Check the model in this Agent session.

Give your coding Agent one short instruction. It will download the skill, select Codex or Claude Code automatically, make 20 fresh model requests through its host-managed login, and return the complete fingerprint report.

Shortest prompt

Run https://verifyllmapi.com/run to verify this AI Agent's model.

This starts the default complete check: 20 fresh model requests (4 probe cells × 5 independent requests). The calls can use your Codex or Claude Code allowance, or your configured provider balance, and may take several minutes. No API key belongs in the prompt.

Request budget

ModeProbe designFresh model requestsUse
Default complete (quick)4 cells × 520Normal verification
standard8 cells × 540Larger evidence budget
paper8 cells × 15120Paper-sized budget

One request produces one short answer in one fresh host session. The scanner aggregates all answers locally. VerifyLLMAPI receives no prompts, answers, credentials, route metadata, or reports.

Instructions for the AI Agent

  1. Download the stable latest package at https://verifyllmapi.com/verify-llm-api.skill into a temporary directory and extract it.
  2. Read its bundled VERSION. If a local verify-llm-api Skill exists, run the downloaded scripts/ensure-latest.mjs --skill-dir LOCAL_SKILL_DIR --archive DOWNLOADED_ARCHIVE. It replaces the local copy only when the downloaded semantic version is newer. If no local copy exists, install the downloaded package.
  3. Read the confirmed current SKILL.md completely. Run its mandatory update check and proceed only after it reports CURRENT or LOCAL_NEWER. If it reports UPDATED, reload the replaced Skill and check again before scanning.
  4. Let the scanner select Codex or Claude Code automatically. Do not select the host from credentials, provider variables, or model variables. Use an explicit --host only for troubleshooting.
  5. Do not inspect environment API keys, auth files, cookies, browser storage, process memory, or the parent conversation. Authentication must remain managed by the Agent host.
  6. Pass --model only if an exact provider-supported slug is directly available. A family label such as GPT-5 is not exact and must not be passed. Never infer it from environment variables; otherwise leave it unpinned and preserve the report caveat.
  7. If the update check fails, stop without scanning. Otherwise run the default complete check—the quick profile with 20 fresh model requests—unless the user requests a larger evidence budget. Return the complete bordered terminal card in a fenced text block.

Machine clients can read the same contract as JSON at /run/protocol.json.

What the scan measures

The scanner automatically selects the current Agent host. Codex uses independent codex exec --ephemeral --ignore-user-config sessions; Claude Code uses independent claude --safe-mode -p --no-session-persistence sessions. It forms answer distributions and compares them with public references using Jensen–Shannon divergence.

A result is behavioral consistency evidence, not vendor attestation. A model absent from the published reference set correctly returns INCONCLUSIVE.

Download the latest skill