Supported AI providers
Max Harness ships with 25 presets. A preset fills in the API endpoint, request format and model discovery, so connecting a provider takes a name and a key. Anything with an OpenAI-compatible, Anthropic or Gemini API can also be added by hand.
| Provider | API format | Endpoint | Models | Notes |
|---|---|---|---|---|
|
OpenAI
API documentation |
OpenAI-compatible | https://api.openai.com/v1 |
Fetched live | Uses the Chat Completions endpoint. Models that only run on the Responses API are not available through this preset. |
|
Anthropic Claude
API documentation |
Anthropic Messages | https://api.anthropic.com/v1 |
Fetched live | Native Messages API with image and PDF input. Recent Claude models reject temperature and top_p, so leave those unset. |
|
Google Gemini
API documentation |
Gemini native | https://generativelanguage.googleapis.com/v1beta |
Fetched live | Native Gemini API with image and PDF input. Uses an AI Studio API key, not a Google Cloud service account. |
|
OpenRouter
API documentation |
OpenAI-compatible | https://openrouter.ai/api/v1 |
Fetched live | One key for models from many labs. The model list includes per-token prices, which Max Harness uses for cost estimates. |
|
OpenCode Zen
API documentation |
OpenAI-compatible | https://opencode.ai/inference/openai/v1 |
Fetched live | Uses the Chat Completions endpoint of the OpenCode gateway. OpenCode documents separate native endpoints for GPT, Claude and Gemini families; models in those families may not respond through this endpoint. Free-tier models are restricted to the OpenCode client. |
|
Command Code
API documentation |
OpenAI-compatible | https://api.commandcode.ai/provider/v1 |
Fetched live | Provider API on paid plans other than Go. Claude models are only served on the Anthropic-style /messages endpoint, so they are hidden here. To use them, add a custom provider with the Anthropic format and the same base URL. |
|
Cline (ClinePass)
API documentation |
OpenAI-compatible | https://api.cline.bot/api/v1 |
Fetched live | The Cline API. A ClinePass subscription covers model IDs that start with cline-pass/. Other models in the list are billed against Cline credits. |
|
DeepSeek
API documentation |
OpenAI-compatible | https://api.deepseek.com |
Fetched live | Standard setup with an API key from the provider console. |
|
Mistral AI
API documentation |
OpenAI-compatible | https://api.mistral.ai/v1 |
Fetched live | Standard setup with an API key from the provider console. |
|
xAI
API documentation |
OpenAI-compatible | https://api.x.ai/v1 |
Fetched live | Uses Chat Completions. xAI now leads its documentation with the Responses API; models that need it are not available here. |
|
Groq
API documentation |
OpenAI-compatible | https://api.groq.com/openai/v1 |
Fetched live | Standard setup with an API key from the provider console. |
|
Together AI
API documentation |
OpenAI-compatible | https://api.together.ai/v1 |
Fetched live | Standard setup with an API key from the provider console. |
|
Fireworks AI
API documentation |
OpenAI-compatible | https://api.fireworks.ai/inference/v1 |
Fetched live | Standard setup with an API key from the provider console. |
|
Perplexity
API documentation |
OpenAI-compatible | https://api.perplexity.ai |
Built-in list | Sonar models search the web themselves and return citations. Perplexity is moving Sonar requests to its newer Agent API, so check its changelog if requests start to fail. Function tools are not sent to this provider. |
|
Cohere
API documentation |
OpenAI-compatible | https://api.cohere.ai/compatibility/v1 |
Fetched live | Uses the Cohere compatibility API. Cohere-specific features such as connectors and citations are not available through it. |
|
AI21 Labs
API documentation |
OpenAI-compatible | https://api.ai21.com/studio/v1 |
Built-in list | AI21 does not allow streaming together with tools, so tools are turned off for this provider. Text only. |
|
Cerebras
API documentation |
OpenAI-compatible | https://api.cerebras.ai/v1 |
Fetched live | Standard setup with an API key from the provider console. |
|
NVIDIA NIM
API documentation |
OpenAI-compatible | https://integrate.api.nvidia.com/v1 |
Fetched live | The catalog mixes chat, embedding and vision models. Disable the ones you do not use. |
|
Hugging Face
API documentation |
OpenAI-compatible | https://router.huggingface.co/v1 |
Fetched live | Inference Providers router. Needs a fine-grained token with the "Make calls to Inference Providers" permission. Append :cheapest or :fastest to a model ID to pick a routing policy. |
|
Moonshot AI (Kimi)
API documentation |
OpenAI-compatible | https://api.moonshot.ai/v1 |
Fetched live | Global endpoint. Accounts created on the China platform use https://api.moonshot.cn/v1 instead; edit the base URL after adding. |
|
Z.AI (GLM)
API documentation |
OpenAI-compatible | https://api.z.ai/api/paas/v4 |
Fetched live | General pay-as-you-go endpoint. GLM Coding Plan keys use https://api.z.ai/api/coding/paas/v4, which Z.AI limits to coding tools. |
|
DeepInfra
API documentation |
OpenAI-compatible | https://api.deepinfra.com/v1/openai |
Fetched live | Standard setup with an API key from the provider console. |
|
SambaNova
API documentation |
OpenAI-compatible | https://api.sambanova.ai/v1 |
Fetched live | Standard setup with an API key from the provider console. |
|
Ollama Cloud
API documentation |
OpenAI-compatible | https://ollama.com/v1 |
Fetched live | Standard setup with an API key from the provider console. |
|
Ollama (local)
API documentation |
OpenAI-compatible | http://localhost:11434/v1 |
Fetched live
No key needed |
Connects to an Ollama server reachable from the web server, not from your browser. On shared hosting that usually means a server you run yourself. Change the host if Ollama runs on another machine. |
Endpoints and authentication were checked against each provider's documentation on November 10, 2026. Providers change their APIs, so a preset can go out of date. You can edit the base URL of any connection, and administrators can hide presets that no longer apply.
How presets were tested
Every preset's request and response handling is covered by automated tests that replay recorded-style provider responses. Each host and model list path was also confirmed to exist with a live request.
Full conversations with real credentials have not been run for these providers as part of the release, because that requires a paid account with each one. The connection test in the workspace sends a real request with your key, so you find out immediately if something does not work.
Requested but not included
OpenDesign. OpenDesign offers a cloud API on paid plans, but it publishes the base URL only inside each account's "Connection setup" panel, so no endpoint could be verified. Add it as a custom provider with the base URL and format shown in your OpenDesign dashboard.
Subscriptions are not API keys
A consumer plan such as ChatGPT Plus, Claude Pro or Gemini Advanced does not include API access. Each provider issues API keys from a separate developer console, usually billed by usage.
Adding a provider that is not listed
Choose "Custom provider" in the workspace and enter a name, a base URL and a key.
OpenAI-compatible (Chat Completions)
Works with any service that exposes /chat/completions, including self-hosted servers such as vLLM, LM Studio and LiteLLM.
Anthropic Messages API
For gateways that expose an Anthropic-style /messages endpoint with an x-api-key header.
Google Gemini API
For endpoints that implement Google's generateContent API with an x-goog-api-key header.