Other model providers

This page covers what works today in each engine and version-specific differences with OpenCode 2.x.

What works today

"Chat engine" is Maven's built-in engine for OpenAI-compatible services. In a bot's model picker it is called OpenAI-compatible (OpenRouter / Groq) and, for now, sits in the Local group.

The Cline row lists its hosted model service. To run Cline's own coding agent, use the Cline engine.

Other engines (Antigravity, Cursor, Gemini, Qwen, Grok, Droid, Kimi, Hermes, Pi and custom agents) use their own sign-in.

Keys you export in your shell (~/.zshrc, ~/.bashrc) don't reach the desktop app when you open it from the Dock, Finder or an app menu: it copies only PATH from your shell. Use the steps below instead.

OpenCode

OpenCode already supports OpenRouter, Fireworks AI, DeepSeek and ClinePass. It just needs your key.

1. In a terminal, run opencode auth login.
2. Pick OpenRouter, Fireworks AI, DeepSeek or ClinePass and paste your key.
3. In Maven, open a bot's model picker, choose OpenCode and press the refresh button (Refresh models). The provider's models appear in the list. Maven reads the model list from an ACP session on both OpenCode 1.x and 2.x, and each model's effort levels the same way: a model that has them, such as GitHub Copilot's GPT and Claude models, shows them under Reasoning in the picker.

Maven uses your normal OpenCode sign-ins and settings, so there is nothing to set up inside the app. Sign in with the same OpenCode version that your bots use (see OpenCode 2.x).

Your own OpenAI-compatible address (this includes Cline credits, at https://api.cline.bot/api/v1):

1. Run opencode auth login, pick Other, type a short ID such as myrouter, and paste your key.

2. Add the provider to ~/.config/opencode/opencode.json (%USERPROFILE%\.config\opencode\opencode.json on Windows):

   {
     "$schema": "https://opencode.ai/config.json",
     "provider": {
       "myrouter": {
         "npm": "@ai-sdk/openai-compatible",
         "name": "My router",
         "options": { "baseURL": "https://router.example.com/v1" },
         "models": { "model-id": { "name": "Model name" } }
       }
     }
   }

   Keep this file plain JSON, with no comments or trailing commas. Maven edits it when an OpenCode bot uses a model running on this computer.

3. Refresh OpenCode's models in Maven.

OpenCode's provider guide has more examples. OpenCode can also use Anthropic-compatible addresses, but we haven't tested that with Maven.

On a self-hosted server that you start from a terminal, OpenCode also reads provider keys exported in that terminal, such as OPENROUTER_API_KEY, FIREWORKS_API_KEY or TOGETHER_API_KEY. Keys exported this way reach every engine on that server, and any bot that can run commands can read them. On OpenCode 2.x this also has a side effect; see below.

CLINE_API_KEY is scoped to the Cline and OpenCode engines. For ClinePass in OpenCode, opencode auth login also keeps its sign-in in OpenCode's own settings.

Chat engine

The chat engine uses one provider at a time.

1. Open Settings → Connections → Model providers.

2. Paste your key into OpenAI-compatible API key and set OpenAI-compatible base URL:

   Provider	Base URL
   OpenRouter	https://openrouter.ai/api/v1 (the default)
   Fireworks AI	https://api.fireworks.ai/inference/v1
   DeepSeek	https://api.deepseek.com
   Cline	https://api.cline.bot/api/v1
   Anything else	The OpenAI-compatible address your provider gives you, usually ending in /v1

3. Press Test. Then pick OpenAI-compatible (OpenRouter / Groq) in a bot's model picker and choose a model.

On a server you can set OPENAI_COMPAT_API_KEY and OPENAI_COMPAT_URL instead.

To use more than one provider at once, add extra openai-compat engines to config.json (see Bring your own engine). Give each one its own key. An engine without a key uses the key from Connections and sends it to that engine's address.

Free OpenRouter models

A new OpenRouter account with no credit can only use free models. Paid models stop with "This model needs OpenRouter credit".

• Picking one: the model picker marks free models (free). When the key belongs to a free account, free models that can use tools come first and one of them is the default. Pressing Test on the key says so. A model you chose yourself is never changed.
• Privacy setting: if OpenRouter's privacy settings block free models, turns stop with "OpenRouter's privacy settings block this free model". Allow free model endpoints at openrouter.ai/settings/privacy, or pick another model.
• Daily limit: OpenRouter limits how many requests free models take each day, and every step of a bot's turn counts as one. When the limit is used up, wait for it to reset or add credit at openrouter.ai/credits to raise it.

Cline engine

1. Install the standalone CLI with npm install -g cline.
2. Run cline auth in a terminal to choose its provider, model and account.
3. In Maven's bot model picker, select Cline and Cline default (configured).

Maven uses Cline's selected provider and saved credentials. Change the provider or model through cline auth; the configured model is pinned before each prompt. An unconfigured account remains unavailable. Set CLI can select a wrapper with --config or --data-dir for a separate Cline home.

Cline receives the bot's scoped MCP tools and uses Maven's approval broker, including Full Access. Native Cline tool selection is not supported.

Claude Code

The Anthropic API key in Connections is for Anthropic only.

To run Claude bots on DeepSeek, GLM or another host that speaks Claude Code's API, add a custom endpoint in Settings → Engines, in the Claude card:

• Add endpoint takes a name, the host's base URL, its API key and, optionally, a model. Leave Model empty and Maven reads the host's model list at /v1/models; type the model id (such as deepseek-chat) when the host doesn't list its models.
• Each endpoint is its own engine with its own directory, so a bot chooses it like any other AI. It never reads or writes this computer's ~/.claude, and the Claude Code on this computer keeps its own sign-in.
• Each bot can choose a different model from the endpoint's catalog.
• Another gateway must support streaming Messages (POST /v1/messages) with tool use. Enter its root URL without /v1/messages, such as https://gateway.example.com.
• The token is saved in Maven's own config.json and sent only to the base URL you enter. Removing an endpoint removes only that entry.

You can also point Claude Code itself at OpenRouter, Fireworks AI or DeepSeek, and your Claude bots follow it:

• Put ANTHROPIC_BASE_URL and your key in the env block of ~/.claude/settings.json. Fireworks' FireConnect does this for you. OpenRouter's and DeepSeek's guides use shell exports, which the desktop app doesn't see, so copy those lines into the env block.
• If you also set ANTHROPIC_MODEL there, that model appears under Use a local model in the Claude model picker; pick it. So do ANTHROPIC_CUSTOM_MODEL_OPTION and each ANTHROPIC_DEFAULT_*_MODEL that isn't a Claude model. Otherwise bots ask for Claude models, which OpenRouter bills as Claude and DeepSeek replaces with its own model. FireConnect's model list doesn't show up in Maven, so add ANTHROPIC_MODEL yourself. A bot runs only the model you pick: ANTHROPIC_DEFAULT_*_MODEL and CLAUDE_CODE_SUBAGENT_MODEL aren't carried over.
• It applies only while no Anthropic API key is saved in Connections.
• It changes Claude Code everywhere on this computer, not just in Maven.
• On a server, use the env block too, not exports in the terminal that starts the server. Otherwise an Anthropic key saved in Connections can be sent to that provider.

Codex

Codex bots can also run on an API key instead of a ChatGPT sign-in:

1. Open Settings → Engines and expand the Codex/ChatGPT card.
2. Under Codex direct API, choose Add endpoint.
3. Leave the base URL empty for OpenAI, or enter another gateway's URL, such as https://gateway.example.com/v1.
4. Enter a name and API key. Model is optional; then save.
5. Choose the new engine and a model in a bot's settings.

• Usage is billed by the provider behind the URL. Saving an endpoint doesn't start a sign-in or replace an existing account.
• Each endpoint has its own Codex home in Maven's data directory. The key is saved in config.json and is never shown in the app or passed in command arguments. Changing the URL asks for the key again.
• A gateway must support streaming Responses (POST /v1/responses) with tool calls. A gateway that only offers Chat Completions works with the chat engine instead.
• Each bot chooses its model. OpenAI uses the standard catalog; another gateway's list is read from /models under its base URL, unless you typed a model.
• Remove an endpoint after moving its bots and the default model off it.

When an endpoint won't save

Saving a custom gateway without a model reads its model list and says whether the network, the API key, the HTTP response or an empty model list failed. Saving with a typed model, or with the official Anthropic or OpenAI URL, doesn't check the key; a wrong key shows up on the bot's first turn. Refresh models reads a gateway's list again. The key is sent to read the list only over HTTPS or to this computer. Check that:

• The URL is reachable from the computer or container running Maven, not just from the one showing the app. Inside a container, localhost means the container itself.
• Codex endpoints use HTTPS. Plain HTTP is accepted only on localhost or 127.0.0.1.
• The gateway serves the right protocol for the agent and lists models for your key.

Ollama's context window

Below 23 GB of graphics memory, Ollama gives a model a 4,096-token context window unless told otherwise, and it drops whatever doesn't fit without an error. That is less than a bot's instructions and tools, so Maven sets a larger window the first time a bot uses a model.

You'll see the result in ollama list as a second name, such as qwen3.5:4b-maven-64k. It uses the same weights, so it takes no extra disk space. Removing it is safe: Maven makes it again when it needs it. To free the disk space a model uses, remove both names.

The window is 16K, 32K or 64K tokens, depending on your memory and the model's size. To choose your own, set num_ctx in a Modelfile or start Ollama with OLLAMA_CONTEXT_LENGTH. A num_ctx in a Modelfile is always left alone. With OLLAMA_CONTEXT_LENGTH, keep the model loaded before the first turn, because Maven can only see that setting on a running model.

Current limits

There is no shared provider list that configures every engine. Use each engine's setup above, and do not assume a key saved for one engine changes another engine's sign-in or model list.

OpenCode 2.x

Maven installs and tests OpenCode 2.x (the @opencode/cli npm package). The older 1.x line shipped as opencode-ai; both packages install a command called opencode.

• If you installed OpenCode from Maven's Settings, bots use that copy. Otherwise they use the first opencode on your PATH, which is 2.x if you installed @opencode/cli. Settings → Engines shows the version in use.
• Differences from 1.x:
  • Some 2.x releases have reported missing configured providers in new ACP sessions (upstream report #50236). If a provider is missing, check the installed version and upstream updates before changing your credentials.
  • When a provider key is set in the environment (for example OPENROUTER_API_KEY on a server), 2.x hides OpenCode's free models and makes a paid model the default.
  • 2.x keeps its sign-ins apart from 1.x, so a sign-in made with one may not be seen by the other. Run opencode auth login with the version your bots use.