Model APIs and inference APIs and MCP servers for AI agents

LLM, embedding and multimodal model endpoints an agent can delegate sub-tasks to.

18 model apis and inference APIs an AI agent can call: auth method and header, official MCP server (7), free tier (9), docs links. Machine-readable JSON included. Agents: fetch catalog/model-inference.json instead of parsing this page. Last checked 2026-09-24.

Ratings and usage (★) fetched: 2026-09-24 16:54 UTC – 2026-09-24 17:02 UTC.

Service	What an agent can do	Auth	MCP server	Free tier	Details (with sources)	Notes	Proof (ratings, usage)
Anthropic Claude API	Send messages to Claude models for text and vision generation, tool use, batch jobs and document analysis.	API key `Authorization: Bearer <ANTHROPIC_API_KEY> (x-api-key still supported as legacy fallback), plus anthropic-version header`	—	paid only	Base URL: https://api.anthropic.com Rate limits: Per usage tier and model (RPM, ITPM, OTPM); Start tier e.g. 1,000 RPM for Opus/Sonnet/Haiku; 429 on exceed ↗ Data policy: Retained data never used for training without express permission; prompts/outputs not retained by default except 30-day rule for Covered Models; ZDR available ↗ Async jobs: Message Batches API: POST /v1/messages/batches, poll GET /v1/messages/batches/{id}, then fetch results ↗	docs.anthropic.com and docs.claude.com redirect to platform.claude.com. Rate limits depend on usage tier. The API's MCP connector lets Claude call remote MCP servers; there is no MCP server that wraps the API itself.	—
Cohere API	Call chat, embed, rerank and classify endpoints for text generation, semantic search and RAG.	API key `Authorization: Bearer <COHERE_API_KEY>`	—	free trial key: 1,000 calls/mo, rate-limited	Base URL: https://api.cohere.com/v2 Free plan: 1,000 API calls (trial/evaluation key), per month ↗ Rate limits: Trial: Chat 20 req/min, Rerank 10/min, Embed 2,000 inputs/min; Production: Chat 500/min, Rerank 1,000/min ↗ Example call: chat / text generation: POST https://api.cohere.com/v2/chat ↗ Unit price: Command R: $0.15 input / $0.60 output per 1M tokens; Embed 4: $0.12 per 1M tokens; Rerank 4 Fast: $2 per 1K searches ↗ MCP: docs-only: the server searches documentation, it cannot call the API ↗	Vendor MCP at https://docs.cohere.com/_mcp/server only searches the docs. Trial keys are not for production use. v2 API is current; v1 is legacy. Flagship chat model: command-a-plus-05-2026	SourceForge 4.0★/1 review (2026-09-24 17:02 UTC)
Comet Opik Cloud API	Lets an agent log and read LLM traces and spans, score them with evaluations and manage datasets and prompts in Opik.	API key / OAuth `Headers 'authorization: <OPIK_API_KEY>' (no Bearer prefix) and 'Comet-Workspace: <WORKSPACE>'; remote MCP uses browser sign-in`	official	Free cloud plan: 25k spans/month, 60-day data retention, up to 10 team members	Base URL: https://www.comet.com/opik/api/v1/private Free plan: 25k spans, 60-day data retention, up to 10 team members, per month ↗ Data policy: Free cloud plan keeps data 60 days ↗ Example call: list traces of a project: GET https://www.comet.com/opik/api/v1/private/traces?project_name={name} ↗ Unit price: Opik Cloud Pro: $19/mo incl. 100k spans, then $5 per extra 100k spans ↗ MCP tools: 5: read, list, write, schema, read_skill ↗	REST base https://www.comet.com/opik/api/v1/private/. Remote MCP needs OAuth browser sign-in; local: uvx opik mcp configure.	—
DeepSeek API	Send chat and reasoning requests to DeepSeek models via an OpenAI- and Anthropic-compatible REST API.	API key `Authorization: Bearer <DEEPSEEK_API_KEY>`	—	paid only	Base URL: https://api.deepseek.com Free plan: none ↗ Rate limits: Account-level concurrency: deepseek-flash 2,500, deepseek-v4-pro 500 concurrent requests; 429 when exceeded ↗	No native SDK: use the OpenAI SDK (base_url https://api.deepseek.com) or the Anthropic SDK (https://api.deepseek.com/anthropic). Off-peak is 50% off; peak is 01:00-04:00 and 06:00-10:00 UTC Mon-Fri. Fees come from a topped-up or granted balance.	SourceForge 5.0★/1 review (2026-09-24 17:02 UTC)
fal.ai	Generate images, video, audio and 3D with 1,000+ hosted models via queued REST or real-time WebSocket.	API key / OAuth `Authorization: Key <FAL_KEY>`	official	Paid only; prepaid credits, priced per model (e.g. FLUX.1 schnell $0.003/megapixel)	Base URL: https://queue.fal.run Free plan: none ↗ Rate limits: New accounts: 2 concurrent requests; limit rises with credit purchases (40+ via sales); excess requests wait in queue ↗ Data policy: Request payloads stored 30 days by default (opt out with X-Fal-Store-IO: 0); generated media served as public CDN URLs ↗ Async jobs: POST https://queue.fal.run/{model} returns request_id, status_url, response_url; poll status or use webhooks ↗ Example call: generate an image from a text prompt via the queue (FLUX.1 schnell): POST https://queue.fal.run/{model_id} ↗ Unit price: FLUX.1 [schnell]: $0.003 per megapixel (rounded up); price varies per model ↗ MCP tools: 11: search_models, get_model_schema, get_pricing, recommend_model, run_model, submit_job … ↗	REST auth header prefix is 'Key', not 'Bearer'; the MCP server takes 'Authorization: Bearer <FAL_KEY>' or OAuth. Long jobs use the queue API (submit, status, result).	—
Fireworks AI	Run serverless or dedicated inference and fine-tuning for open models via OpenAI-compatible APIs.	API key `Authorization: Bearer <FIREWORKS_API_KEY>`	—	trial credits: $1 on signup	Base URL: https://api.fireworks.ai/inference/v1 Free plan: $1 in free credits, one-time ↗ Rate limits: Adaptive serverless limits per account and model (TPM ceilings by model size); upper bound grows with spending tier; 429/503 ↗ Data policy: Does not log or store prompt or generation data for open models without explicit opt-in (zero data retention) ↗ Async jobs: Batch API: create batch job, it waits in pending state, poll job status ↗ Example call: chat completion: POST https://api.fireworks.ai/inference/v1/chat/completions ↗ Unit price: Serverless per 1M tokens, e.g. GLM 5.3 Flash $0.15 input / $0.50 output; unlisted models over 16B params $0.90 per 1M tokens ↗ MCP: docs-only: the server searches documentation, it cannot call the API ↗	Vendor MCP at https://docs.fireworks.ai/mcp only searches the docs. The Responses API can call external MCP servers as tools.	—
Google Gemini API	Generate text and images, analyze multimodal input (video, audio, PDFs), and run code execution with Gemini models.	API key `x-goog-api-key: <GEMINI_API_KEY> (or ?key= query param)`	—	free tier: rate-limited access to Flash/Pro models	Base URL: https://generativelanguage.googleapis.com/v1beta Free plan: free input and output tokens on supported models, rate-limited per project ↗ Rate limits: Per project (not per key): RPM, TPM, RPD by model and tier (Free, Tier 1-3); view active limits in AI Studio ↗ Data policy: Free tier content used to improve Google products; paid tier content not used ↗ Async jobs: Batch API at 50% cost: create batch job, poll job state until JOB_STATE_SUCCEEDED ↗ Example call: generate text: POST https://generativelanguage.googleapis.com/v1beta/interactions ↗ Unit price: gemini-3.5-flash paid tier: $1.50 per 1M input tokens, $9.00 per 1M output tokens (free tier available) ↗ MCP: docs-only: the server searches documentation, it cannot call the API ↗	Vendor MCP at https://gemini-api-docs-mcp.dev only searches the docs. Free-tier content may be used to improve Google products. Unrestricted standard API keys are rejected in 2026; use a current key type. Google publishes a Discovery document, not OpenAPI. The quickstart now uses POST /v1beta/interactions (header Api-Revision: 2026-05-20) rather than generateContent	—
Groq API	Run fast LLM chat, tool use and speech-to-text on hosted open models via an OpenAI-compatible API.	API key `Authorization: Bearer <GROQ_API_KEY>`	—	free tier: rate-limited per model, no card	Base URL: https://api.groq.com/openai/v1 Free plan: per-model limits, e.g. openai/gpt-oss-120b 30 RPM, 1K RPD, 8K TPM, 200K TPD, per day ↗ Rate limits: Per model RPM/RPD/TPM/TPD; Free plan e.g. gpt-oss-120b 30 RPM, 1K RPD; Developer plan higher ↗ Data policy: Inference data not retained by default; retained up to 30 days only for batch/fine-tuning features or reliability needs ↗ Async jobs: Batch API (Developer plan): POST /openai/v1/batches, poll GET /openai/v1/batches/{batch_id} ↗ Example call: chat completion: POST https://api.groq.com/openai/v1/chat/completions ↗ Unit price: openai/gpt-oss-120b: $0.15 per 1M input tokens, $0.60 per 1M output tokens (Llama 3.x models now listed as contact sales) ↗	Free-tier limits vary by model (e.g. 30 RPM / 1K RPD); see console.groq.com/settings/limits. Not to be confused with xAI Grok. Llama 3.x models (llama-3.3-70b-versatile, llama-3.1-8b-instant) are now contact-sales, not self-serve per-token	—
Hugging Face Inference Providers	Call hundreds of open-weight chat, vision, image, video and audio models across many providers with one HF token.	API key `Authorization: Bearer <HF_TOKEN>`	official	Free tier: $0.10/month in Inference Providers credits (PRO $2.00/month)	Base URL: https://router.huggingface.co/v1 Free plan: $0.10 in credits for free users (PRO $2.00), per month ↗ Data policy: HF does not store request/response bodies or use data for training; debug logs kept up to 30 days without user data ↗ Example call: chat completion on an open model: POST https://router.huggingface.co/v1/chat/completions ↗ MCP tools: 4: hf_whoami, hub_repo_search, hub_repo_details, hf_fs ↗	The OpenAI-compatible endpoint (router.huggingface.co/v1) handles chat only; use InferenceClient for other tasks. The token needs the 'Make calls to Inference Providers' scope. Pick a provider with a model suffix such as :fastest or :cheapest.	—
Mistral AI API	Run chat completions, embeddings, OCR and document processing with Mistral models via REST.	API key `Authorization: Bearer <MISTRAL_API_KEY>`	official	free tier: Studio free mode, rate-limited	Base URL: https://api.mistral.ai/v1 Free plan: included monthly usage within Free mode limits, no card ↗ Rate limits: Per-model tokens-per-minute and other limits shown per organization in Admin Panel > API > Limits; Free mode has lower limits ↗ Data policy: Zero data retention available on paid plans for supported stateless API calls; training opt-out handled separately ↗ Example call: create a chat completion: POST https://api.mistral.ai/v1/chat/completions ↗ Unit price: Mistral Large: $0.5 per 1M input tokens, $1.5 per 1M output tokens; Batch API 50% off ↗	Official MCP (same API key) currently exposes only Studio Skills management tools, not chat inference; call the REST API for inference.	SourceForge 5.0★/1 review (2026-09-24 17:02 UTC)
OpenAI API	Generate text, images, audio and embeddings, and run tool-calling via the Responses and Chat Completions APIs.	API key `Authorization: Bearer <OPENAI_API_KEY>`	official	paid only	Base URL: https://api.openai.com/v1 Rate limits: RPM/RPD/TPM/TPD/IPM per model, set by usage tier (Free, Tier 1-5); 429 slow_down / rate_limit_error ↗ Data policy: API data not used for training unless you opt in; abuse-monitoring logs kept 30 days; ZDR available for approved customers ↗ Async jobs: Batch API: upload JSONL, create batch, poll batch status (completes within 24h) ↗ MCP tools: 5: search_openai_docs, list_openai_docs, fetch_openai_doc, list_api_endpoints, get_openapi_spec ↗	Docs moved from platform.openai.com to developers.openai.com (old URLs redirect). Usage-based billing with a payment method; rate limits scale with usage tier. The official MCP server is documentation-only and does not call the API.	SourceForge 5.0★/3 reviews (2026-09-24 17:02 UTC)
OpenRouter	Route chat requests to hundreds of LLMs from many providers through one OpenAI-compatible endpoint with fallback.	API key / OAuth `Authorization: Bearer <OPENROUTER_API_KEY>`	official	free models: 50 req/day (1,000/day after $10 purchase)	Base URL: https://openrouter.ai/api/v1 Free plan: free models: 50 requests/day (1,000/day once $10+ credits purchased), 20 RPM, no card ↗ Rate limits: Free models 20 RPM and 50 RPD (1,000 RPD with >= $10 credits); paid models limited by provider/DDoS protection; X-RateLimit-* headers ↗ Data policy: OpenRouter does not store prompts or responses unless you opt in; provider training/logging policies documented per provider ↗ Example call: send a chat completion to a routed model: POST https://openrouter.ai/api/v1/chat/completions ↗ Unit price: Tokens billed at the provider's list price per model (no markup); 5.5% fee on credit purchases (Standard), 8% (Business) ↗ MCP tools: 17: send-message, generate-image, list-models, get-model, list-model-endpoints, list-providers … ↗	The official remote MCP server uses OAuth PKCE and mints a 7-day key with a $10 default spend limit. It includes send-message and generate-image tools. OAuth PKCE is also available for apps.	SourceForge 5.0★/1 review (2026-09-24 17:02 UTC)
PiAPI	Generate images, video and music through one API proxying Midjourney, Flux, Kling and other models.	API key `X-API-Key: <key>`	community	trial credits: $0.50 on signup	Base URL: https://api.piapi.ai Free plan: $0.50 in free credits, one-time ↗ Async jobs: POST https://api.piapi.ai/api/v1/task creates a task; get task endpoint for results, or webhook notifications ↗ Example call: generate an image from text (Flux): POST https://api.piapi.ai/api/v1/task ↗ Unit price: Flux: flux1-schnell $0.0015 per image, flux1-dev $0.015 per image, flux1-dev-advanced $0.02 per image ↗	Third-party reseller, not the model owner. The Midjourney access is unofficial, so it can break or violate Midjourney's terms of service. Tasks are async (create task, then poll or webhook).	GitHub 75★ apinetwork/piapi-mcp-server (2026-09-24 16:54 UTC)
Replicate	Run thousands of hosted image, video, audio and text models via HTTP predictions; also fine-tune and deploy.	API key `Authorization: Bearer <REPLICATE_API_TOKEN>`	official	paid only	Base URL: https://api.replicate.com/v1 Rate limits: Create prediction 600 req/min; other endpoints 3,000 req/min; accounts with granted credit but no payment method: 6 req/min ↗ Data policy: API prediction inputs, outputs, files and logs deleted automatically after one hour by default ↗ Async jobs: POST /v1/predictions returns a prediction to poll (GET /v1/predictions/{id}) or use webhooks; 'Prefer: wait' blocks up to 60s ↗ Example call: run an official model and wait for the output (FLUX schnell text-to-image): POST https://api.replicate.com/v1/models/{owner}/{name}/predictions ↗ Unit price: black-forest-labs/flux-schnell: $3.00 per 1,000 output images ($0.003/image); other models billed per output or per second of compute ↗	Billed per second of compute or per output. The MCP server is also available locally via npx replicate-mcp.	—
Together AI	Run serverless inference, batch jobs, fine-tuning and dedicated GPU endpoints for open-source models.	API key `Authorization: Bearer <TOGETHER_API_KEY>`	—	Paid only: no free trial; minimum $5 prepaid credit purchase	Base URL: https://api.together.ai/v1 Free plan: none ↗ Rate limits: Dynamic per-model limits that scale with sustained traffic; 429 includes x-ratelimit-reset ↗ Data policy: By default stores prompts/responses and may use them for product improvement (admins can disable); training on org data is opt-in ↗ Async jobs: Batch API: upload JSONL, start batch job, poll until finished, retrieve results (up to 50% cheaper) ↗ MCP: docs-only: the server searches documentation, it cannot call the API ↗	Vendor MCP at https://docs.together.ai/mcp only searches the docs. The API is OpenAI-compatible.	—
Vercel AI Gateway	Send chat, image, embedding and rerank requests to 100+ models through one gateway with failover and spend tracking.	API key / OAuth `Authorization: Bearer <AI_GATEWAY_API_KEY>`	—	free tier: monthly credits on subset of models	Base URL: https://ai-gateway.vercel.sh/v1 Free plan: free AI Gateway credits on a subset of models, rate-limited per model ↗ Rate limits: Free tier: lower per-model limits (429); paid tier: no AI Gateway limits, only upstream provider limits ↗ Data policy: Per-request Zero Data Retention routing to providers that do not retain or train on prompts (Pro and Enterprise) ↗ Example call: create a chat completion: POST https://ai-gateway.vercel.sh/v1/chat/completions ↗	The base URL https://ai-gateway.vercel.sh works with OpenAI and Anthropic SDKs. Vercel's platform MCP (mcp.vercel.com) does not call the gateway, only buys its credits. Tokens have no markup. BYOK requires the paid tier. Free-tier overage returns 429. On Vercel, OIDC tokens can replace the API key.	—
Voyage AI (embeddings/rerank)	Generate text, code and multimodal embeddings and rerank documents for RAG and semantic search.	API key `Authorization: Bearer <VOYAGE_API_KEY>`	—	free tier: 200M tokens on current models	Base URL: https://api.voyageai.com/v1 Free plan: first 200 million tokens per account on current models (e.g. voyage-4, voyage-4-large), one-time ↗ Rate limits: Per model RPM/TPM by usage tier (Tier 1 = payment method added), e.g. voyage-3.5 2,000 RPM / 8M TPM; Tier 2 2x, Tier 3 3x ↗ Data policy: Voyage may store and use API data for training unless an org admin with a payment method opts out (then zero-day retention) ↗ Async jobs: Batch API: upload .jsonl file, create batch, poll GET retrieve batch ↗ Example call: embed texts: POST https://api.voyageai.com/v1/embeddings ↗ Unit price: voyage-4: $0.06 per 1M tokens (voyage-4-lite $0.02, voyage-4-large $0.12); first 200M tokens free ↗	Owned by MongoDB, but the standalone API at api.voyageai.com is still self-serve. Free tokens do not apply to the Batch API.	—
xAI Grok API	Generate text, images, video and voice with Grok models, with built-in web search and code tools.	API key `Authorization: Bearer <XAI_API_KEY>`	—	paid only (prepaid credits)	Base URL: https://api.x.ai/v1 Free plan: none ↗ Rate limits: Per-model RPS (RPM/60) and TPM by tier; tier based on cumulative spend since 2026-01-01 (Tier 0 $0, Tier 1 $50, Tier 2 $250...) ↗ Async jobs: Batch API: create batch, add requests, poll batch status until num_pending is 0 ↗ MCP: docs-only: the server searches documentation, it cannot call the API ↗	The account must have prepaid credits (console.x.ai). The API is OpenAI-compatible (/v1/responses, /v1/chat/completions). Vendor MCP at https://docs.x.ai/api/mcp only searches the docs.	—

Model APIs and inference APIs and MCP servers for AI agents

Other categories