Skip to main content
Condor needs a language model for chat and for its agents. You can use a coding-agent CLI you already log in to, a cloud API key, or a model running on your own machine. Condor reaches models in two ways:
  • ACP bridges: Claude, Gemini, GitHub Copilot and Codex run as a separate process that speaks the Agent Client Protocol. You log in to each one once, with its own CLI.
  • OpenAI-compatible APIs: OpenRouter, Ollama, LM Studio and custom endpoints are called directly by Condor. They need an API key, a running local server, or both.

The model options

Every place that offers a model (the setup wizard, the Workspace pickers and Telegram) uses the same list: The two “Pick Model” rows aren’t models. They open a list, and your choice is stored as the full key, for example openrouter:deepseek/deepseek-v4.1-flash.

The default model

The recommended model is Claude (ACP) — Sonnet (claude-acp:sonnet). make pick-model lists it first. When you haven’t picked a model yourself, Condor uses the first of these that is set:
  1. CONDOR_DEFAULT_AGENT in .env
  2. The agent_key in Condor’s own agents/condor/AGENT.md, which ships as claude-acp:sonnet
  3. claude-code
A model you pick in Telegram or on the Workspace chat page overrides this default for your chats.

Choose a model

Run this in the Condor folder:
It lists every option with a status for this machine: ready (✓), installed but login not confirmed (?), or not usable (✗) with the fix. Pick a number, and Condor saves it to .env as CONDOR_DEFAULT_AGENT.
  • If you pick a bridge that isn’t installed, the wizard runs its npm install -g command for you.
  • If the Claude bridge is outdated and you pick a Claude option (or Claude is your default), the wizard runs the upgrade.
  • For Ollama or LM Studio, it lists the models the server has loaded. Press Enter to let the server choose at session start.
  • For OpenRouter, it asks for your API key, checks it with OpenRouter, and lists the 20 cheapest models that support tool calling. It saves the key to .env as well.
  • Custom endpoints aren’t offered here. Add them from the Workspace or Telegram.
Restart Condor (make restart) so the new default takes effect. To print the report without changing anything, run uv run python -m condor.setup_llm --status. make doctor also checks the default model.

Claude

Condor runs Claude through the claude-agent-acp bridge from the npm package @agentclientprotocol/claude-agent-acp.
1

Install the bridge

2

Log in once

Run claude and log in to your Anthropic account. Condor checks for the login in ~/.claude/.credentials.json (or the macOS keychain). See the Claude Code docs to install the claude CLI.
3

Pick a Claude option

Choose Claude (ACP) — Sonnet, Opus or Haiku to pin a model, or Claude Code to use the bridge’s own default model.
The bridge must be version 0.79.0 or newer. The older package @zed-industries/claude-agent-acp still launches, but it is frozen at 0.23. Its bundled Claude CLI bills Sonnet 4.6 at Opus rates and doesn’t know any model newer than the 4.6 family.make pick-model and make doctor mark an old bridge as outdated and show the fix. make pick-model runs the upgrade when you choose a Claude option or when Claude is already your default, and a Condor update (/update) upgrades it automatically. To do it by hand:
The uninstall is needed because both packages install the same claude-agent-acp command.

Gemini, GitHub Copilot and Codex

Condor launches these bridges with npx, but it treats a bridge as installed only when its npm package is already on disk. Install each one globally: Condor checks for a Gemini login in ~/.gemini/oauth_creds.json (or GEMINI_API_KEY) and a Codex login in ~/.codex/auth.json. It can’t check a Copilot login, so Copilot shows as “login could not be verified” until your first chat works. Updating Codex. If a newer Codex model fails because the bridge is outdated, run:
This updates @agentclientprotocol/codex-acp and the Codex CLI bundled with it. Start a new Codex session afterwards. Sessions that are already running keep the old bridge.

OpenRouter

OpenRouter gives you many models through one API key. Condor lists only OpenRouter models that support tool calling, because its agents depend on tools.
  1. Create a key at openrouter.ai/keys.
  2. Add it to Condor’s .env, or let make pick-model save it for you:
  3. Pick a model:
    • Workspace: choose OpenRouter — Pick Model in the model picker.
    • Telegram: /agent → Change LLM → OpenRouter — Pick Model. Page through the list, or tap Enter model manually and send a slug.
    • make pick-model: choose OpenRouter.
The model is stored as openrouter:<slug>, for example openrouter:deepseek/deepseek-v4.1-flash. You can browse the list without a key, but a session can’t start until OPENROUTER_API_KEY is set.

Ollama and LM Studio

Local models run on your own machine with no API cost.
  1. Install Ollama from ollama.com and start it with ollama serve.
  2. Pull a model, for example ollama pull qwen3:32b.
  3. Choose Ollama — Default Model, or set a specific model as ollama:<model>, for example ollama:qwen3:32b.
Condor connects to http://localhost:11434/v1.
Which model the default option uses. With a bare key (ollama: or lmstudio:), Condor picks the model when the session starts:
  1. CONDOR_DEFAULT_LOCAL_MODEL or OLLAMA_MODEL, if either is set
  2. Otherwise the first model the server lists at /v1/models
  3. For Ollama only, the first model from /api/tags
If none of these finds a model, the session fails with “No local model found”. Use an explicit key such as ollama:qwen3:32b, or set CONDOR_DEFAULT_LOCAL_MODEL. Server on another address. For a local server on a different host or port, the simplest option is to add it as a custom endpoint. Chat sessions also read LMSTUDIO_BASE_URL from .env as their base URL for these models. Set it only when you use LM Studio: Condor applies it to every model of this kind in chat, including Ollama and OpenRouter, unless you choose a custom endpoint. Slow first replies. A local model can take a long time to load on its first request. Condor waits up to 600 seconds for a reply. Set LOCAL_MODEL_READ_TIMEOUT (in seconds) to change this.

Custom OpenAI-compatible endpoints

You can connect any API that follows the OpenAI format, such as Venice AI, Together, Fireworks, or your own vLLM or LM Studio server. Endpoints are saved per user, and the Workspace and Telegram share the same list.
  1. Open Settings → LLM Endpoints and click Add endpoint.
  2. Enter the Base URL, for example https://api.venice.ai/api/v1.
  3. Enter an API key, or leave it blank if the endpoint doesn’t need one.
  4. Optionally enter a Name. If you leave it blank, Condor derives one from the URL.
  5. Click Validate & save.
Condor checks the endpoint by fetching <base URL>/models and lists the chat models it finds. Pick one, and Condor stores it as custom@<nickname>:<model>, for example custom@venice:llama-3.3-70b. Optional .env variables:
Leave CUSTOM_LLM_BLOCK_PRIVATE_URLS unset on a personal install, so you can point at a server such as http://localhost:8000/v1. Turn it on for a Condor shared by several users. Cloud metadata addresses are always refused.

Models for agents and loops

Each agent has its own model, set by agent_key in its AGENT.md. Most shipped agents use claude-acp:sonnet, and a new agent starts with the model you’re using. If agent_key is empty, a chat with that agent uses your chat’s model, but its loops and delegations run on claude-code.
You can change it from the agent page (the model next to its name), or by editing AGENT.md. A loop run picks its model in this order:
  1. agent_key in the run’s config
  2. agent_key in the loop’s loop.md (null means “use the agent’s model”)
  3. The agent’s agent_key
A loop won’t start on a model that can’t run. Before it starts a loop, Condor checks the model and refuses with Model '<key>' cannot run: <reason>. Examples are an unknown provider, a key with no model id, a custom endpoint you never saved, or an OpenRouter model without OPENROUTER_API_KEY. A local server that is switched off isn’t caught by this check, because it may be running by the first tick. That fails at run time instead. Fallback when an agent is asked for help. When one agent delegates to another whose model is Ollama, LM Studio, OpenRouter or a custom endpoint, Condor first checks that backend. If it is unreachable, Condor answers with AGENT_FALLBACK_MODEL instead (default claude-code) and notes this in the answer. Set AGENT_FALLBACK_MODEL= (empty) in .env to turn the fallback off, so the request fails with the reason instead.

Troubleshooting

Run make pick-model or make doctor. Each row says what is missing and the command that fixes it, for example not installed — npm install -g @github/copilot or not reachable at http://localhost:11434/v1 — start it with ollama serve.
The bridge needs its own login. Run the login command from the Gemini, GitHub Copilot and Codex table, or claude for Claude, then start a new session.
You are probably running the old @zed-industries/claude-agent-acp bridge. See the warning in Claude.
Restart Condor with make restart. If you already picked a model in Telegram or on the Workspace chat page, that choice wins over CONDOR_DEFAULT_AGENT for your chats.
The local server is running but Condor couldn’t find a model on it. Load a model, use an explicit key such as ollama:qwen3:32b, or set CONDOR_DEFAULT_LOCAL_MODEL.