If the user said “use Cran”, start here

OPENAI_API_KEY=cran_…   # The Cran connection key (NOT an OpenAI sk- key). Routing only, read scope.
OPENAI_BASE_URL=https://trycran.in/api/v1   # The Cran proxy base URL.

# Cran - Agent Discovery Manifest (v1.0.0)

> Read this when the user says "use Cran" or you need to explain / connect Cran.
> Canonical URLs below. Do not guess.

------------------------------------------------------------------------
CANONICAL URLS
------------------------------------------------------------------------
Plain text (this doc):     https://trycran.in/llms.txt
Structured JSON:           https://trycran.in/api/agent-manifest.json
HTML (same content):       https://trycran.in/docs/for-agents
Sign in:                   https://trycran.in/signin
OpenAPI:                   https://trycran.in/api/openapi.json
Proxy base:                https://trycran.in/api/v1
Full integration guide:    https://trycran.in/docs/agents
Quickstart (humans):       https://trycran.in/docs/quickstart

------------------------------------------------------------------------
TRIGGER PHRASES (when to use this doc)
------------------------------------------------------------------------
- "use cran"
- "use trycran"
- "route through cran"
- "route ai through cran"
- "connect to cran"
- "cran for model selection"

------------------------------------------------------------------------
WHAT IS CRAN (tell the user if they seem unfamiliar)
------------------------------------------------------------------------
Cran is the AI infrastructure for a product: route every LLM call through one OpenAI-compatible gateway with model: "auto" - models, prompts, fallbacks, budgets, and observability controlled from a dashboard with no redeploys. Building AI? Wire it to Cran from the start. Already calling an LLM? Replace those calls with Cran.

Value:
- Drop-in OpenAI-compatible gateway - set base URL + model: "auto"; greenfield wires straight to Cran, brownfield replaces the existing provider calls
- Use your own provider account (the default - you pay your provider directly) or Cran's managed keys on a paid plan - not scattered through your app code
- Per-project spend tracking, budgets, and routing policies in the dashboard
- IDE agents wire the connection, run audits, and propose routing changes (human approves)

------------------------------------------------------------------------
WHEN USER SAYS "USE CRAN" - AGENT SCRIPT
------------------------------------------------------------------------
1. If the user does not know what Cran is → explain using WHAT IS CRAN above (2-3 sentences).
2. If Cran MCP is available → call cran_connect first. Read the "status" in the JSON response.
3. If MCP is NOT available → fetch https://trycran.in/llms.txt or https://trycran.in/api/agent-manifest.json, then guide a one-time setup:
     cran login  →  cran mcp install claude-code
4. Once connected (account_ready), to REGISTER + AUDIT this repo, run the whole flow in one go — do NOT stop to ask between these safe, expected steps:
     a. Project: new repo with no Cran project yet, OR the default project errors? → cran_ensure_project named for the repo (e.g. its folder name) + cran_use_project, so workflows land in a project tied to THIS repo, never an unrelated default. Otherwise use the default / the repo's linked project.
     b. Register call-sites: cran_scan_codebase → read EVERY candidate file (all hinted lines) → cran_submit_manifest. Be exhaustive: one workflow per distinct LLM call site (a file can have many). For each, also record: downstream_consumers (copy the candidate's "consumed by" list — these become the pipeline edges), is_helper (true for candidates the scan marked HELPER — shared utilities, not stages), endpoints (the routes/surfaces that invoke it, from the scan's "served by"), and feature (a short label like "Post generation"). Scan-then-submit is ONE continuous flow; submitting is safe + idempotent (upsert by slug), so never ask "should I submit?".
     c. Audit: cran_run_architecture_audit — THE default for "run an audit" / "audit my project". It covers the WHOLE project (every workflow) — do not ask which workflow. Summarize the findings and link https://trycran.in/dashboard/architecture.
     d. Only for a SPECIFIC workflow's model comparison (usually to act on an architecture finding) → cran_run_audit.
5. GO LIVE (deploy) — the consumer's goal once workflows are audited + a model is published:
     1. Publish a model per workflow: after an audit, set the winner live (dashboard 'Make it live', or cran_propose_change -> human approves -> cran_apply_proposal).
     2. Wire the app (a code change you CAN make): point the OpenAI/SDK client at OPENAI_BASE_URL + OPENAI_API_KEY read from env, and use model: "auto" at every call site - never hardcode the key.
     3. Get the connection key: cran_link_repo (writes .env) or the dashboard Deploy page. Read scope, routing only.
     4. Hand the deploy to the USER: tell them to set OPENAI_API_KEY + OPENAI_BASE_URL in their host (e.g. Vercel -> Settings -> Environment Variables, Production) and redeploy. Do not set their production secrets yourself.
     5. Verify: open the dashboard Deploy page and send one request - it flips to 'you're live' and the response carries an x-cran-model header. After this, change models/prompts from the dashboard with NO redeploy.
   Security: only the connection key (cran_…) goes in the app (as OPENAI_API_KEY); the cran login session is agent-only — never in app env, never committed. Change code to READ the key from env; tell the USER to set the production secret in their host.

------------------------------------------------------------------------
ONBOARDING DECISION TREE (cran_connect / cran_onboard statuses)
------------------------------------------------------------------------
[needs_login]
  Agent: Ask the user to run `cran login` once in a terminal and approve in the browser. One account login connects Cran for EVERY repo — there is no per-project linking.
  Tell user: Cran needs a one-time browser login. Run `cran login`, approve in the browser, then ask me to use Cran again.

[account_ready]
  Agent: Connected via account login — no per-repo link needed. Project tools (workflows, audits, proposals, limits) target the user's DEFAULT project. Call cran_list_projects to see all and cran_use_project to switch. To register THIS repo's call sites: cran_scan_codebase → cran_submit_manifest. cran_link_repo is OPTIONAL — only to route the app's live traffic through the proxy.
  Tell user: You're connected to Cran. I'll wire this product's AI calls through Cran and register them - connect a provider key in the dashboard (Settings → AI providers) so it can run.

[already_linked]
  Agent: This repo already has proxy routing env (OPENAI_API_KEY + OPENAI_BASE_URL in .env). Scaffold LLM clients with model: "auto". Do NOT re-link.
  Tell user: This repo is already wired to route traffic through Cran. I'll use model: "auto".

------------------------------------------------------------------------
MCP TOOLS (in order of onboarding)
------------------------------------------------------------------------
- cran_connect  →  User says use Cran - entry point; returns onboarding status (account_ready when logged in).
- cran_onboard  →  Same assessment as cran_connect.
- cran_list_projects  →  List the user's Cran projects.
- cran_use_project  →  Switch the active project for this session (id or slug). Omit to use the default. No files written.
- cran_ensure_project  →  Create or find a project. Use for a NEW repo (name it after the repo) so its workflows don't land in an unrelated default project.
- cran_scan_codebase  →  STEP 1 — discover ALL LLM call sites locally (source never uploaded); per workflow record feeds_into (data-flow) + endpoints/feature (which app surface it serves). Be exhaustive. Always follow immediately with cran_submit_manifest.
- cran_submit_manifest  →  STEP 2 — register the scanned workflows. Safe + idempotent (upsert by slug); run it right after scanning, never ask to confirm.
- cran_run_architecture_audit  →  THE default audit — run for 'run an audit' / 'audit my project'. Reviews the WHOLE project's AI architecture (every workflow); don't ask which workflow.
- cran_run_audit  →  Per-workflow MODEL benchmark for ONE workflow (NOT the project audit) — use to act on an architecture finding.
- cran_link_repo  →  OPTIONAL: route live traffic — mint a connection key + write .env. Not needed for scan/audit/manage.

Install MCP: npm i -g cran-cli && cran login && cran mcp install claude-code

------------------------------------------------------------------------
CLI COMMANDS (fallback when MCP unavailable)
------------------------------------------------------------------------
- cran login  →  One-time browser auth → ~/.cran/session.json (connects Cran for every repo)
- cran mcp install claude-code  →  Add Cran MCP to your IDE (claude-code | cursor | claude | windsurf | codex) after login. Works on your default project.
- cran projects list  →  List your Cran products
- cran link [--team SLUG] [--project NAME]  →  OPTIONAL: mint a connection key + write .env to route live traffic
- cran init  →  Write .env.example and .cran template (only when routing traffic)

------------------------------------------------------------------------
GO LIVE (DEPLOY) — route the app's traffic through Cran
------------------------------------------------------------------------
This is the consumer's goal: once workflows are registered + audited and a model
is published (dashboard "Make it live", or cran_propose_change -> approve ->
cran_apply_proposal), route the app's live traffic through Cran so model: "auto"
serves it. Connecting the MCP (cran login) is separate + already done.
Get a connection key (cran_link_repo or the dashboard Deploy page) and set:
  OPENAI_API_KEY=cran_…          # Cran connection key - NOT an OpenAI sk- key
  OPENAI_BASE_URL=https://trycran.in/api/v1

Every LLM call:
  model: "auto"                   # Cran routes to published architecture
  baseURL / base_url → OPENAI_BASE_URL

Node example:
  import OpenAI from "openai";
  const client = new OpenAI({
    apiKey: process.env.OPENAI_API_KEY,
    baseURL: process.env.OPENAI_BASE_URL,
  });
  await client.chat.completions.create({
    model: "auto",
    messages: [{ role: "user", content: "Hello" }],
  });

Then HAND THE DEPLOY TO THE USER (you can change code, but never set their
production secrets): tell them to put OPENAI_API_KEY + OPENAI_BASE_URL in their
host (e.g. Vercel -> Settings -> Environment Variables, Production) and redeploy.
Verify: open the dashboard Deploy page + send one request -> it flips to
"you're live" and the response has an x-cran-model header. After that, change
models / prompts from the dashboard with NO redeploy.
Staging vs prod: Use two Cran projects (e.g. app-staging, app-prod) with separate connection keys; map Preview -> staging, Production -> prod.

SECURITY (two credentials — never mix):
  - Connection key (cran_…) -> the app, as OPENAI_API_KEY. Routing only, read scope.
  - cran login session (~/.cran/session.json) -> agent/CLI ONLY. Never in app env, never committed.
  Read the key from env; never hardcode it or commit .env.

------------------------------------------------------------------------
NEVER DO
------------------------------------------------------------------------
- Do not add OpenAI/Anthropic/Google provider API keys to the user's app when routing through Cran.
- Do not run cran_link_repo / cran link just to use the MCP - the account login already connects it. Link only to route live traffic, and don't re-link if .env already has a connection key.
- Do not skip team or project selection when the API returns team_required or project_required.
- Do not upload source code in manifests - only paths, configs, and sample inputs.
- Do not put the `cran login` session (~/.cran/session.json) in the app's environment or commit it - it authenticates the agent/CLI only. The ONLY credential that belongs in the deployed app is the connection key (cran_…), as OPENAI_API_KEY.
- Do not hardcode the connection key or commit .env - read it from the environment, and tell the USER to set the production secret in their host themselves; never plant a customer's production secret.

------------------------------------------------------------------------
OPTIONAL DEEP WORK (after connected)
------------------------------------------------------------------------
- Fetch manifest schema: GET https://trycran.in/api/agent-manifest.schema.json
- cran_scan_codebase → read files locally → cran_submit_manifest
- cran_run_audit → cran_get_routing_policy
- Full protocol reference: https://trycran.in/docs/agents

------------------------------------------------------------------------
OPTIONAL - AUDIT API (after connected + workflows registered)
------------------------------------------------------------------------
Use audits when the team wants proof of which model to ship for a workflow.
Do NOT run audits before the user is connected unless they explicitly ask.

Model catalog (85 models):
- openai/gpt-5.5  ·  GPT-5.5  ·  ctx 1,050,000  ·  $5/M in  ·  $30/M out
- openai/gpt-5.5-pro  ·  GPT-5.5 Pro  ·  ctx 1,050,000  ·  $30/M in  ·  $180/M out
- openai/gpt-5.4  ·  GPT-5.4  ·  ctx 1,050,000  ·  $2.5/M in  ·  $15/M out
- openai/gpt-5.4-pro  ·  GPT-5.4 Pro  ·  ctx 1,050,000  ·  $30/M in  ·  $180/M out
- openai/gpt-5.4-mini  ·  GPT-5.4 mini  ·  ctx 272,000  ·  $0.75/M in  ·  $4.5/M out
- openai/gpt-5.4-nano  ·  GPT-5.4 nano  ·  ctx 272,000  ·  $0.2/M in  ·  $1.25/M out
- openai/gpt-5.2  ·  GPT-5.2  ·  ctx 272,000  ·  $1.75/M in  ·  $14/M out
- openai/gpt-5.2-pro  ·  GPT-5.2 Pro  ·  ctx 272,000  ·  $21/M in  ·  $168/M out
- openai/gpt-5.1  ·  GPT-5.1  ·  ctx 272,000  ·  $1.25/M in  ·  $10/M out
- openai/gpt-5  ·  GPT-5  ·  ctx 272,000  ·  $1.25/M in  ·  $10/M out
- openai/gpt-5-mini  ·  GPT-5 mini  ·  ctx 272,000  ·  $0.25/M in  ·  $2/M out
- openai/gpt-5-nano  ·  GPT-5 nano  ·  ctx 272,000  ·  $0.05/M in  ·  $0.4/M out
- openai/gpt-5-pro  ·  GPT-5 Pro  ·  ctx 128,000  ·  $15/M in  ·  $120/M out
- openai/gpt-4.1  ·  GPT-4.1  ·  ctx 1,047,576  ·  $2/M in  ·  $8/M out
- openai/gpt-4.1-mini  ·  GPT-4.1 mini  ·  ctx 1,047,576  ·  $0.4/M in  ·  $1.6/M out
- openai/gpt-4.1-nano  ·  GPT-4.1 nano  ·  ctx 1,047,576  ·  $0.1/M in  ·  $0.4/M out
- openai/gpt-4o  ·  GPT-4o  ·  ctx 128,000  ·  $2.5/M in  ·  $10/M out
- openai/gpt-4o-mini  ·  GPT-4o mini  ·  ctx 128,000  ·  $0.15/M in  ·  $0.6/M out
- openai/o3  ·  o3  ·  ctx 200,000  ·  $2/M in  ·  $8/M out
- openai/o3-pro  ·  o3-pro  ·  ctx 200,000  ·  $20/M in  ·  $80/M out
- openai/o3-mini  ·  o3-mini  ·  ctx 200,000  ·  $1.1/M in  ·  $4.4/M out
- openai/o4-mini  ·  o4-mini  ·  ctx 200,000  ·  $1.1/M in  ·  $4.4/M out
- openai/o1  ·  o1  ·  ctx 200,000  ·  $15/M in  ·  $60/M out
- anthropic/claude-fable-5  ·  Claude Fable 5  ·  ctx 1,000,000  ·  $10/M in  ·  $50/M out
- anthropic/claude-opus-4-8  ·  Claude Opus 4.8  ·  ctx 1,000,000  ·  $5/M in  ·  $25/M out
- anthropic/claude-sonnet-5  ·  Claude Sonnet 5  ·  ctx 1,000,000  ·  $2/M in  ·  $10/M out
- anthropic/claude-opus-4-7  ·  Claude Opus 4.7  ·  ctx 1,000,000  ·  $5/M in  ·  $25/M out
- anthropic/claude-opus-4-6  ·  Claude Opus 4.6  ·  ctx 1,000,000  ·  $5/M in  ·  $25/M out
- anthropic/claude-opus-4-5  ·  Claude Opus 4.5  ·  ctx 200,000  ·  $5/M in  ·  $25/M out
- anthropic/claude-opus-4-1  ·  Claude Opus 4.1  ·  ctx 200,000  ·  $15/M in  ·  $75/M out
- anthropic/claude-sonnet-4-6  ·  Claude Sonnet 4.6  ·  ctx 1,000,000  ·  $3/M in  ·  $15/M out
- anthropic/claude-sonnet-4-5  ·  Claude Sonnet 4.5  ·  ctx 200,000  ·  $3/M in  ·  $15/M out
- anthropic/claude-sonnet-4  ·  Claude Sonnet 4  ·  ctx 1,000,000  ·  $3/M in  ·  $15/M out
- anthropic/claude-3-7-sonnet  ·  Claude 3.7 Sonnet  ·  ctx 200,000  ·  $3/M in  ·  $15/M out
- anthropic/claude-haiku-4-5  ·  Claude Haiku 4.5  ·  ctx 200,000  ·  $1/M in  ·  $5/M out
- anthropic/claude-3-haiku  ·  Claude 3 Haiku  ·  ctx 200,000  ·  $0.25/M in  ·  $1.25/M out
- google/gemini-3.1-pro  ·  Gemini 3.1 Pro  ·  ctx 1,048,576  ·  $2/M in  ·  $12/M out  (preview)
- google/gemini-3-pro  ·  Gemini 3 Pro  ·  ctx 1,048,576  ·  $2/M in  ·  $12/M out  (preview)
- google/gemini-3.5-flash  ·  Gemini 3.5 Flash  ·  ctx 1,048,576  ·  $1.5/M in  ·  $9/M out
- google/gemini-3-flash  ·  Gemini 3 Flash  ·  ctx 1,048,576  ·  $0.5/M in  ·  $3/M out  (preview)
- google/gemini-3.1-flash-lite  ·  Gemini 3.1 Flash Lite  ·  ctx 1,048,576  ·  $0.25/M in  ·  $1.5/M out  (preview)
- google/gemini-2.5-pro  ·  Gemini 2.5 Pro  ·  ctx 1,048,576  ·  $1.25/M in  ·  $10/M out
- google/gemini-2.5-flash  ·  Gemini 2.5 Flash  ·  ctx 1,048,576  ·  $0.3/M in  ·  $2.5/M out
- google/gemini-2.5-flash-lite  ·  Gemini 2.5 Flash Lite  ·  ctx 1,048,576  ·  $0.1/M in  ·  $0.4/M out
- google/gemini-2.0-flash  ·  Gemini 2.0 Flash  ·  ctx 1,048,576  ·  $0.1/M in  ·  $0.4/M out
- google/gemini-2.0-flash-lite  ·  Gemini 2.0 Flash Lite  ·  ctx 1,048,576  ·  $0.075/M in  ·  $0.3/M out
- xai/grok-4  ·  Grok 4  ·  ctx 256,000  ·  $3/M in  ·  $15/M out
- xai/grok-3-mini  ·  Grok 3 Mini  ·  ctx 131,072  ·  $0.3/M in  ·  $0.5/M out
- groq/llama-3.3-70b-versatile  ·  Llama 3.3 70B (Groq)  ·  ctx 131,072  ·  $0.59/M in  ·  $0.79/M out
- groq/llama-3.1-8b-instant  ·  Llama 3.1 8B (Groq)  ·  ctx 131,072  ·  $0.05/M in  ·  $0.08/M out
- mistral/mistral-large-latest  ·  Mistral Large  ·  ctx 131,072  ·  $2/M in  ·  $6/M out
- mistral/mistral-small-latest  ·  Mistral Small  ·  ctx 131,072  ·  $0.1/M in  ·  $0.3/M out
- deepseek/deepseek-chat  ·  DeepSeek V3  ·  ctx 128,000  ·  $0.27/M in  ·  $1.1/M out
- deepseek/deepseek-reasoner  ·  DeepSeek R1  ·  ctx 128,000  ·  $0.55/M in  ·  $2.19/M out
- together/llama-3.3-70b-instruct-turbo  ·  Llama 3.3 70B (Together)  ·  ctx 131,072  ·  $0.88/M in  ·  $0.88/M out
- together/qwen2.5-72b-instruct-turbo  ·  Qwen 2.5 72B (Together)  ·  ctx 32,768  ·  $1.2/M in  ·  $1.2/M out
- fireworks/llama-v3p3-70b-instruct  ·  Llama 3.3 70B (Fireworks)  ·  ctx 131,072  ·  $0.9/M in  ·  $0.9/M out
- fireworks/kimi-k2-instruct  ·  Kimi K2 (Fireworks)  ·  ctx 131,072  ·  $0.6/M in  ·  $2.5/M out
- cohere/command-a  ·  Command A  ·  ctx 256,000  ·  $2.5/M in  ·  $10/M out
- azure/gpt-4o  ·  GPT-4o (Azure)  ·  ctx 128,000  ·  $2.5/M in  ·  $10/M out
- azure/gpt-4o-mini  ·  GPT-4o mini (Azure)  ·  ctx 128,000  ·  $0.15/M in  ·  $0.6/M out
- azure/gpt-4.1  ·  GPT-4.1 (Azure)  ·  ctx 1,047,576  ·  $2/M in  ·  $8/M out
- openrouter/auto  ·  OpenRouter Auto  ·  ctx 128,000  ·  $5/M in  ·  $15/M out
- openrouter/llama-3.3-70b-instruct  ·  Llama 3.3 70B (OpenRouter)  ·  ctx 131,072  ·  $0.3/M in  ·  $0.4/M out
- perplexity/sonar-pro  ·  Sonar Pro  ·  ctx 200,000  ·  $3/M in  ·  $15/M out
- perplexity/sonar  ·  Sonar  ·  ctx 128,000  ·  $1/M in  ·  $1/M out
- cerebras/llama-3.3-70b  ·  Llama 3.3 70B (Cerebras)  ·  ctx 128,000  ·  $0.85/M in  ·  $1.2/M out
- cerebras/gpt-oss-120b  ·  GPT-OSS 120B (Cerebras)  ·  ctx 128,000  ·  $0.35/M in  ·  $0.75/M out
- sambanova/llama-3.3-70b  ·  Llama 3.3 70B (SambaNova)  ·  ctx 128,000  ·  $0.6/M in  ·  $1.2/M out
- nvidia/llama-3.3-70b-instruct  ·  Llama 3.3 70B (NVIDIA NIM)  ·  ctx 128,000  ·  $0.3/M in  ·  $0.3/M out
- huggingface/llama-3.3-70b-instruct  ·  Llama 3.3 70B (Hugging Face)  ·  ctx 128,000  ·  $0.4/M in  ·  $0.4/M out
- hyperbolic/llama-3.3-70b-instruct  ·  Llama 3.3 70B (Hyperbolic)  ·  ctx 131,072  ·  $0.4/M in  ·  $0.4/M out
- novita/llama-3.3-70b-instruct  ·  Llama 3.3 70B (Novita)  ·  ctx 131,072  ·  $0.39/M in  ·  $0.39/M out
- moonshot/kimi-k2-turbo  ·  Kimi K2 Turbo  ·  ctx 256,000  ·  $0.6/M in  ·  $2.5/M out
- moonshot/kimi-k2-thinking  ·  Kimi K2 Thinking  ·  ctx 256,000  ·  $0.6/M in  ·  $2.5/M out
- qwen/qwen-max  ·  Qwen Max  ·  ctx 131,072  ·  $1.6/M in  ·  $6.4/M out
- qwen/qwen-plus  ·  Qwen Plus  ·  ctx 131,072  ·  $0.4/M in  ·  $1.2/M out
- zhipu/glm-4.6  ·  GLM-4.6  ·  ctx 200,000  ·  $0.6/M in  ·  $2.2/M out
- openai/text-embedding-3-small  ·  Embedding 3 Small  ·  ctx 8,191  ·  $0.02/M in  ·  $0/M out
- openai/text-embedding-3-large  ·  Embedding 3 Large  ·  ctx 8,191  ·  $0.13/M in  ·  $0/M out
- google/gemini-embedding  ·  Gemini Embedding  ·  ctx 2,048  ·  $0.15/M in  ·  $0/M out
- cohere/embed-v4  ·  Cohere Embed v4  ·  ctx 128,000  ·  $0.12/M in  ·  $0/M out
- openai/gpt-image-1  ·  GPT Image 1  ·  ctx 0  ·  $0/M in  ·  $0/M out
- openai/dall-e-3  ·  DALL·E 3  ·  ctx 0  ·  $0/M in  ·  $0/M out
- google/imagen-4  ·  Imagen 4  ·  ctx 0  ·  $0/M in  ·  $0/M out  (preview)

Available on this deployment:
  - openai/gpt-5.5
  - openai/gpt-5.5-pro
  - openai/gpt-5.4
  - openai/gpt-5.4-pro
  - openai/gpt-5.4-mini
  - openai/gpt-5.4-nano
  - openai/gpt-5.2
  - openai/gpt-5.2-pro
  - openai/gpt-5.1
  - openai/gpt-5
  - openai/gpt-5-mini
  - openai/gpt-5-nano
  - openai/gpt-5-pro
  - openai/gpt-4.1
  - openai/gpt-4.1-mini
  - openai/gpt-4.1-nano
  - openai/gpt-4o
  - openai/gpt-4o-mini
  - openai/o3
  - openai/o3-pro
  - openai/o3-mini
  - openai/o4-mini
  - openai/o1

Run audit:
  POST https://trycran.in/api/audits
  Authorization: Bearer cran_<connection_or_agent_token>

Fetch audit:
  GET https://trycran.in/api/audits/<audit_id>

List audits:
  GET https://trycran.in/api/audits?limit=20

Manifest schema:
  GET https://trycran.in/api/agent-manifest.schema.json

OpenAPI:
  GET https://trycran.in/api/openapi.json

Full agent protocol:
  https://trycran.in/docs/agents

------------------------------------------------------------------------
END
------------------------------------------------------------------------