--- name: chutes-ai description: "Chutes.ai hub skill — the entry point for decentralized open-source AI inference on Chutes. Use this skill when a user mentions Chutes, chutes.ai, decentralized AI, DeepSeek/Qwen/Kimi/GLM API access, TEE or confidential-compute inference, serverless GPU inference, or wants to register, manage API keys, discover models, or make OpenAI-compatible inference calls. This is the hub: it handles the session credential check, account lifecycle, API key lifecycle, basic inference, and routes to sibling skills (chutes-sign-in, chutes-deploy, chutes-mcp-portability, chutes-routing, chutes-usage-and-billing, chutes-platform-ops, chutes-agent-registration) for everything else." --- # Chutes.ai — Hub Skill (four product lanes) This skill is the **hub** for Chutes.ai integration. It covers the "Use Chutes" lane end-to-end (account → API keys → models → inference) and routes to sibling skills for the other three lanes. Chutes provides OpenAI-compatible inference for top open-source models (DeepSeek, Qwen, Kimi, GLM, Gemma, MiniMax, Nemotron) on decentralized GPU infrastructure, very cheaply. As of 2026-06-11 the entire hosted LLM catalog runs in TEEs (confidential compute) — every model on `/v1/models` has `confidential_compute: true`. There are no Llama models on the platform. Any tool that talks to OpenAI can point at Chutes instead. ## Four product lanes | Lane | What | Skill | |---|---|---| | **Use Chutes** | Account, API keys, models, inference, routing basics | `chutes-ai` (this skill) | | **Build on Chutes** | Sign in with Chutes, OAuth apps, framework adapters | `chutes-sign-in` **[BETA]** | | **Operate on Chutes** | Model aliases, usage/quota/billing, secret rotation, token lifecycle | `chutes-routing` / `chutes-usage-and-billing` / `chutes-platform-ops` (wave 2 stubs) | | **Run agents with Chutes** | Claude / Hermes / Cursor / Aider via MCP + drop-in configs | `chutes-mcp-portability` **[BETA]** | Plus: deploying your own chutes (vLLM / diffusion / custom CDK / TEE via `tee=True`) lives in `chutes-deploy` **[BETA]**, TEE attestation verification lives in `chutes-tee`, and agent-native self-onboarding lives in `chutes-agent-registration` (wave 2 stub). ## Skill router — when to hand off Hand off to the matching sibling skill immediately when the user says anything in the right column: | Hand off to | User intent | |---|---| | `chutes-sign-in` | "Add Sign in with Chutes to my app", OAuth app, `/idp/apps`, PKCE, `cid_`/`csc_`, rotate client secret, scopes, Next.js auth | | `chutes-deploy` | Deploy a model on Chutes, vLLM chute, diffusion chute, build a chute image, TEE deploy (`tee=True`), rolling updates, `POST /chutes/`, `POST /images/` | | `chutes-mcp-portability` | Use Chutes from Cursor / Cline / Aider / Hermes / Claude Desktop, MCP server, drop-in config, make Chutes available to another agent | | `chutes-routing` (stub) | `default:latency`, `default:throughput`, inline model pools, stable model aliases, routing recipes, cost-aware routing | | `chutes-usage-and-billing` (stub) | Chutes spend, quotas, discounts, subscription usage, payment history, invocation stats | | `chutes-platform-ops` (stub) | Rotate OAuth client secret (outside SIWC setup), introspect/revoke Chutes tokens, list authorizations, model alias CRUD | | `chutes-agent-registration` (stub) | Agent self-onboarding without human signup, `POST /users/agent_registration`, hotkey-signed registration | Wave 2 stubs describe scope and endpoints but do not yet have walkthroughs. When the user hits a stub, explain the stub status and point at the read-only MCP tools in `chutes-mcp-portability` for interim data. --- ## Session Initialization When this skill is first invoked in a session, **check for stored credentials before doing anything else**: ```bash python /manage_credentials.py check ``` - If credentials exist → retrieve the API key silently for use in API calls: ```bash python /manage_credentials.py get --field api_key ``` - If no credentials exist → proceed with Account Creation (below) and save immediately after. - **Never paste raw `cpk_` / `cid_` / `csc_` values into the conversation.** Use them only in request headers and subprocess stdin. Replace `` with `plugins/chutes-ai/skills/chutes-ai/scripts/` in this repo. --- ## Quick Orientation Two base URLs: | Purpose | URL | |---|---| | Account management (keys, billing, chutes, `/idp/*`) | `https://api.chutes.ai` | | Inference (OpenAI-like request/response shape) | `https://llm.chutes.ai/v1` | Auth (re-verified live 2026-06-11 — this **reverses** the April 2026 finding): - **`Authorization: Bearer cpk_...` is the universal header.** Verified 200 on `GET llm.chutes.ai/v1/models`, on a live paid `POST llm.chutes.ai/v1/chat/completions` (real completion returned, 2026-06-11), AND on management GETs (`/users/me`, `/model_aliases/`) at `api.chutes.ai`. It is also the header the platform's own `ai-plugin.json` and `llms.txt` document. - `X-API-Key` returns **401 on `api.chutes.ai`** management endpoints. On the inference surface it is **confirmed silently ignored** (live, 2026-06-11): a `POST /v1/chat/completions` sent with `X-API-Key: cpk_...` got the anonymous nginx 429, byte-identical to a fully unauthenticated POST, while Bearer succeeded in the same minute — the request falls through to the anonymous path. **Do not use `X-API-Key`.** - `GET /v1/models` requires no auth (verified 2026-06-11). - The fingerprint-login JWT is no longer needed for `GET /users/me` — a plain Bearer `cpk_` key works. - CLI CRUD endpoints like `GET /api_keys/` used hotkey-signed headers in April 2026 testing (not re-verified as of 2026-06-11). --- ## Step 1: Account Creation Chutes uses a **32-character alphanumeric fingerprint** as the primary credential. It is shown **only once** during creation and cannot be recovered without a linked Bittensor wallet. **Via API (recommended for agents):** ``` POST https://api.chutes.ai/users/register Body: { "username": "desired-username" } ``` Usernames: 3–20 characters, alphanumeric. Wave-3 live registration finding (verified 2026-04-15): the practical agent flow has extra prerequisites not captured in the bare endpoint sketch above. - Registration requires a human-obtained one-time token from `https://rtok.chutes.ai/users/registration_token`. - That token is protected by Cloudflare / browser verification and may be IP-bound, so a headless agent may need the user to fetch and paste it. - The registering coldkey must have at least `0.25 TAO`, otherwise the API returns: `You must have at least 0.25 tao on your coldkey to register an account.` - The token can expire or mismatch the caller IP; if registration starts failing with `Invalid registration token, or registration token does not match expected IP address`, fetch a fresh token and retry immediately. For agent-led onboarding, treat this as a human-in-the-loop step rather than a fully autonomous one. **Via web browser:** direct users to `https://chutes.ai/auth/start` (the "Create Account" button on `https://chutes.ai/auth` opens a support widget, not the form). **Via Google / GitHub:** `https://chutes.ai/auth` → OAuth sign-in → link account. After registration: 1. Show the fingerprint to the user clearly. 2. Save it to the credential store immediately (Step 2b below). 3. Mention recovery: if they link a Bittensor wallet, they can reset via `https://chutes.ai/auth/reset` or `POST /users/change_fingerprint`. **Login:** `POST /users/login` with the fingerprint returns a session/JWT for further operations. --- ## Step 2: API Key Creation ``` POST https://api.chutes.ai/api_keys/ Body: { "name": "my-agent-key", "admin": false } ``` Response includes `secret_key` (prefixed `cpk_`) — **shown only once**. Save immediately. Other API key operations: - `GET /api_keys/` — list (no `secret_key` returned) - `DELETE /api_keys/{api_key_id}` — revoke The `user_id` is embedded in the key itself (`cpk_..`), or just call `GET /users/me`. --- ## Step 2b: Credential Store After creating an account and API key, **immediately save credentials** to the secure credential store (OS keychain on macOS / Linux Secret Service, AES-256-GCM encrypted file fallback). ```bash python /manage_credentials.py set-profile \ --username \ --user-id \ --fingerprint \ --api-key ``` Read in a new session: ```bash python /manage_credentials.py get --field api_key python /manage_credentials.py get # all fields as JSON ``` Multiple profiles: `--profile ` on any command. `list-profiles` to enumerate. `check` to inspect backend / permissions without revealing secrets. **OAuth app credentials** (for Sign in with Chutes) use the same store under distinct fields: ```bash python /manage_credentials.py set-profile \ --profile oauth.my-app \ --client-id \ --client-secret ``` Env var overrides (highest priority, for CI/CD): - `CHUTES_API_KEY`, `CHUTES_FINGERPRINT` - `CHUTES_OAUTH_CLIENT_ID` / `CHUTES_CLIENT_ID` (either is accepted) - `CHUTES_OAUTH_CLIENT_SECRET` / `CHUTES_CLIENT_SECRET` (either is accepted) - `CHUTES_PROFILE` (profile name) Security details live in `references/api-reference.md` and `docs/credential-store.md`. --- ## Step 3: Model Discovery ``` GET https://llm.chutes.ai/v1/models Authorization: Bearer cpk_... # optional — endpoint is public (verified 2026-06-11) ``` As of 2026-06-11 the catalog is **13 LLMs, all TEE** (`confidential_compute: true`, `-TEE` suffixed IDs). Each model exposes `id`, `root`, `chute_id`, `confidential_compute` (boolean — **use this, not the `-TEE` suffix**, as source of truth), `owned_by` (`sglang` / `vllm`), `pricing.{prompt,completion,input_cache_read}` (USD per 1M tokens), `context_length`, `max_output_length`, `supported_features` (`tools`, `json_mode`, `structured_outputs`, `reasoning`), `supported_sampling_parameters`, `input_modalities`, `output_modalities`, `quantization`. Chutes hosts more than LLMs: **image, video, TTS (54 voices / 9 languages), STT, music, moderation, and custom inference.** Quick static reference: `references/known-models.md`. Always query the live endpoint for authoritative data. When helping users choose (live catalog as of 2026-06-11): - **Frontier coding/agentic** → `moonshotai/Kimi-K2.6-TEE` or `zai-org/GLM-5.1-TEE`; budget → `MiniMaxAI/MiniMax-M2.5-TEE`. - **Reasoning** → `zai-org/GLM-5.1-TEE`, `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-TEE`; budget → `Qwen/Qwen3-235B-A22B-Thinking-2507-TEE`. - **Cheap-fast chat** → `google/gemma-4-31B-turbo-TEE`; absolute cheapest → `unsloth/Mistral-Nemo-Instruct-2407-TEE`. - **Vision/multimodal** → `moonshotai/Kimi-K2.6-TEE` (text+image+video), `Qwen/Qwen3.5-397B-A17B-TEE`, `google/gemma-4-31B-turbo-TEE`. - **Privacy** → everything: all hosted LLMs are `confidential_compute: true` now. - **Cost** → `pricing.prompt`, `pricing.completion`, cache hits via `pricing.input_cache_read`. - **Features** → `supported_features` (not every model has tools / structured output). - **Context** → match `context_length`. --- ## Step 4: Making Inference Calls Chutes uses OpenAI-like request/response shapes on the inference surface, and auth now matches OpenAI SDK defaults. Verified live 2026-06-11 (reverses the April 2026 finding): - `Authorization: Bearer cpk_...` returned 200 on `/v1/models` **and on a real `POST /v1/chat/completions`** (live completion returned, model `unsloth/Mistral-Nemo-Instruct-2407-TEE`); it is also the header the platform documents (ai-plugin.json, llms.txt) for all of `llm.chutes.ai/v1` - `X-API-Key` is confirmed silently ignored on the inference surface — the same completion POST sent with `X-API-Key` got the anonymous nginx 429, byte-identical to a fully unauthenticated POST, while Bearer succeeded in the same minute So use standard `Authorization: Bearer` everywhere. Generic OpenAI SDKs that hardcode Bearer work as-is — just set `base_url="https://llm.chutes.ai/v1"` and `api_key="cpk_..."`. Successful completions carry `x-chutes-invocationid` plus `x-chutes-quota-total` / `-used` / `-remaining` and `x-chutes-rl-user` response headers, and the `usage` block includes `prompt_tokens_details.cached_tokens` — prompt caching is active on inference (all verified 2026-06-11 on a direct-model-id call). **Python:** ```python import requests response = requests.post( "https://llm.chutes.ai/v1/chat/completions", headers={"Authorization": "Bearer cpk_...", "Content-Type": "application/json"}, json={ "model": "deepseek-ai/DeepSeek-V3.2-TEE", "messages": [{"role": "user", "content": "Hello!"}], }, timeout=60, ) response.raise_for_status() print(response.json()["choices"][0]["message"]["content"]) ``` **cURL:** ```bash curl https://llm.chutes.ai/v1/chat/completions \ -H "Authorization: Bearer cpk_..." \ -H "Content-Type: application/json" \ -d '{"model":"deepseek-ai/DeepSeek-V3.2-TEE","messages":[{"role":"user","content":"Hello!"}]}' ``` **Node:** ```javascript const response = await fetch('https://llm.chutes.ai/v1/chat/completions', { method: 'POST', headers: { 'Authorization': 'Bearer cpk_...', 'Content-Type': 'application/json', }, body: JSON.stringify({ model: 'deepseek-ai/DeepSeek-V3.2-TEE', messages: [{ role: 'user', content: 'Hello!' }], }), }); ``` **Vercel AI SDK:** `npm install @chutes-ai/ai-sdk-provider`. Different engines (`sglang` vs `vllm`) accept different sampling params — check `supported_sampling_parameters` before passing non-standard options. ### Inline routing (one-liner; full recipes in `chutes-routing`) Pass a comma-separated model list as `model` for sequential failover, append `:latency` or `:throughput` to rank by live metrics: ``` model="zai-org/GLM-5.1-TEE,moonshotai/Kimi-K2.6-TEE,Qwen/Qwen3-32B-TEE:latency" ``` Saved pools via the dashboard use `default`, `default:latency`, `default:throughput`. For routing recipes, stable aliases, and cost-aware routing, hand off to the `chutes-routing` skill (wave 2 stub — use the MCP tools meanwhile). --- ## Step 5: Account Info & Basic Balance ``` GET https://api.chutes.ai/users/me Authorization: Bearer cpk_... ``` Live finding (verified 2026-06-11): a plain Bearer `cpk_...` key now works on `/users/me` — the fingerprint-login JWT workaround is no longer needed for this read. `X-API-Key` returns 401 here. Returns `username`, `user_id`, `balance` (USD), `payment_address` (Bittensor SS58), `hotkey`/`coldkey`, `quotas`, `permissions`. For anything beyond a balance check — discounts, quotas by chute, subscription usage, invocation stats, payment history — hand off to `chutes-usage-and-billing` (wave 2 stub; use MCP read tools in the meantime). --- ## Special features (one-line summaries, deep content in siblings) - **TEE (confidential compute).** As of 2026-06-11 **every** hosted LLM runs in Intel TDX (`confidential_compute: true`). Attestation evidence: `GET /chutes/{chute_id}/evidence?nonce=<64 hex chars>` — the nonce query param is required and must be exactly 64 hex characters (32 bytes). Golden TEE measurements: `GET /servers/tee/measurements` (unverified as of 2026-06-11). See the `chutes-tee` skill for the verification flow. - **Harvard research endpoint (25% off).** Drop-in replacement base URL `https://research-data-opt-in-proxy.chutes.ai/v1`. Trade-off: prompts/responses are recorded for research. Do not send sensitive data. - **Cache hit pricing.** Repeated prompts transparently hit the cache and pay `pricing.input_cache_read`. - **Sign in with Chutes.** OAuth 2.0 + PKCE. `POST /idp/apps`, `cid_` / `csc_` returned. Full integrator flow lives in `chutes-sign-in`. - **Model aliases.** Stable semantic handles via `/model_aliases/`. Recommended packs (`interactive-fast`, `private-reasoning`, `cheap-background`, `agent-coder`, `tee-chat`) live in `chutes-routing` stub; create one with `chutes-deploy` → `alias_deploy.py` **[BETA]**. - **Custom deployment.** `chutes-deploy` **[BETA]** covers vLLM / diffusion / TEE (`tee=True` template kwarg — there is no "teeify" command in the SDK) / rolling updates. The pricing page now advertises self-serve private TEE deployments (RTX Pro 6000 96GB Blackwell, $1.80/hr + one-time deploy fee of 3x the hourly rate) **[BETA]** (unverified as of 2026-06-11). --- ## Credential Security Checklist 1. **Fingerprint is shown once.** Losing it without a linked wallet means losing access. 2. **Always save credentials to the keychain store.** Never write them to plaintext files. 3. **API key is shown once.** If lost, delete the key and create a new one. 4. **Never echo secrets in conversation.** Use `manage_credentials.py get --field ` and pipe the output. 5. **Fingerprint recovery** works if a Bittensor wallet is linked (`https://chutes.ai/auth/reset` or `POST /users/change_fingerprint`). --- ## Quick Reference: Key Endpoints | What | Method | URL | |---|---|---| | Register | POST | `https://api.chutes.ai/users/register` | | Login | POST | `https://api.chutes.ai/users/login` | | My account info | GET | `https://api.chutes.ai/users/me` | | Create / list / delete API key | POST / GET / DELETE | `https://api.chutes.ai/api_keys/[{id}]` | | List models | GET | `https://llm.chutes.ai/v1/models` | | Chat completions | POST | `https://llm.chutes.ai/v1/chat/completions` | | Payment history | GET | `https://api.chutes.ai/payments` | | Quota limits | GET | `https://api.chutes.ai/users/me/quotas` | | Discounts | GET | `https://api.chutes.ai/users/me/discounts` | | List / create model alias | GET / POST | `https://api.chutes.ai/model_aliases/` | | OAuth app create | POST | `https://api.chutes.ai/idp/apps` | | Chute TEE evidence | GET | `https://api.chutes.ai/chutes/{id}/evidence?nonce=<64-hex>` | | Platform pricing (public, incl. live TAO/USD rate) | GET | `https://api.chutes.ai/pricing` | | Subscription usage | GET | `https://api.chutes.ai/users/me/subscription_usage` | Everything else: see `references/api-reference.md` or `https://api.chutes.ai/openapi.json`. --- ## Additional Resources - `references/api-reference.md` — extended endpoint details - `references/known-models.md` — static snapshot of popular models - `references/model-aliases.md` — alias packs + lifecycle (NEW) - `docs/sign-in-with-chutes.md` — OAuth / app builder overview - `docs/model-aliases.md` — why aliases beat hardcoded IDs - Full Swagger docs: `https://api.chutes.ai/docs` - Plugin manifest: `https://chutes.ai/.well-known/ai-plugin.json` - OpenAPI schema: `https://api.chutes.ai/openapi.json` - Knowledge base: `https://chutesai.zohodesk.com/portal/en/kb/chutes-ai`