Integrating your LLM

Condor supports multiple LLM providers through two integration methods:

ACP Protocol: Native integration with Claude Code, Gemini, GitHub Copilot, and Codex
PydanticAI: Direct API access to OpenRouter and to local models via Ollama and LM Studio

See ACP documentation for the full protocol specification.

Before installing Condor: Install and authenticate your LLM provider in the CLI first.
Already installed Condor? Set up your LLM provider, then restart Condor for the changes to take effect.

Using the `/agent` Command

The /agent command in Telegram is how you connect Condor to your LLM and run autonomous Trading Agents. From the /agent menu you can:

Change LLM — pick which installed provider to use for new sessions
Start — launch an agent session with the currently selected LLM
Stop — end the active session
Status — view the currently active provider and model

A typical first-time flow is: install and authenticate your provider in the CLI, then in Telegram run /agent → Change LLM → pick your provider → /agent → Start. Full step-by-step instructions are in Configure in Telegram.

Supported Providers

Cloud Providers (ACP Protocol)

Provider	Agent Key	Installation
Claude Code	`claude-code`	`curl -fsSL https://claude.ai/install.sh \| sh`
Gemini	`gemini`	`npm install -g @google/gemini-cli`
GitHub Copilot	`copilot`	`npm install -g @github/copilot-cli`
Codex	`codex`	`npx @zed-industries/codex-acp`

API Key Providers (PydanticAI)

Provider	Agent Key	Setup
OpenRouter	`openrouter`	Set `OPENROUTER_API_KEY` — see OpenRouter

Local Providers

Provider	Agent Key	Installation
Ollama	`ollama`	`curl -fsSL https://ollama.com/install.sh \| sh`
LM Studio	`lmstudio`	Download from lmstudio.ai

Setup Instructions

Claude Code

Claude Code is the default and most integrated option for Condor Trading Agents. 1. Install Claude Code: Install via CLI:

# macOS/Linux
curl -fsSL https://claude.ai/install.sh | sh

2. Authenticate:

claude auth login

Follow the prompts to sign in with your Anthropic account. 3. Verify installation:

claude --version
claude  # Run Claude and see if it starts up without issues

4. Configure in Telegram: Follow the Configure in Telegram steps and select the Claude Code label.

Gemini

1. Install Gemini CLI:

npm install -g @google/gemini-cli

2. Authenticate: Start Gemini CLI — on first run it will prompt you to authenticate:

gemini

Select Login with Google and follow the OAuth flow in your browser. Your credentials are cached locally for future sessions. Alternatively, set a GEMINI_API_KEY environment variable from Google AI Studio. 3. Verify:

gemini --version

4. Configure in Telegram: Follow the Configure in Telegram steps and select the Gemini CLI label.

GitHub Copilot

1. Install prerequisites:

# Install Copilot CLI
npm install -g @github/copilot-cli

2. Authenticate:

gh auth login
copilot auth login

3. Verify:

copilot --version
copilot  # Run Copilot and see if it starts up without issues

4. Configure in Telegram: Follow the Configure in Telegram steps and select the Github Copilot CLI label.

Codex

Condor connects to Codex through Zed’s ACP bridge (@zed-industries/codex-acp). You’ll need the Codex CLI installed and authenticated, plus the ACP bridge available for Condor to invoke. 1. Install the Codex CLI: Download and install Codex from openai.com/codex. 2. Sign in: Open Codex and sign in with your ChatGPT account or an OpenAI API key. 3. Install the ACP bridge:

npx @zed-industries/codex-acp

Running this once fetches the package so Condor can invoke it via npx. You can also install it globally with npm install -g @zed-industries/codex-acp. 4. Verify installation:

codex --version
codex  # Run Codex and see if it starts up without issues

5. Configure in Telegram: Follow the Configure in Telegram steps and select the ChatGPT Codex label.

OpenRouter

OpenRouter routes requests to many models — including some free ones — through a single API key, often at lower cost than going to a provider directly. Condor connects to it through PydanticAI’s OpenAI-compatible client, so no CLI install is required. 1. Get an API key: Create a key at openrouter.ai/keys. 2. Set the environment variable: Add your key to Condor’s .env:

OPENROUTER_API_KEY=sk-or-...

3. Configure in Telegram: Run /agent → Change LLM → OpenRouter — Pick Model. Condor fetches the live OpenRouter catalog and lists only models that advertise tool-call support (Condor’s agents depend on tool calling, so models without it are filtered out). Pick one from the list, or choose Enter model manually to provide a model slug. Your selection is stored as openrouter:<model> — for example openrouter:anthropic/claude-sonnet-4-5. To run at no cost, choose one of the openrouter/free models.

If a request fails with an insufficient-credits error, add credits to your OpenRouter account or switch to a openrouter/free model.

Ollama (Local)

Perfect for development, testing, and cost-effective experimentation. 1. Install Ollama: Visit ollama.com and download for your platform.

# macOS/Linux
curl -fsSL https://ollama.com/install.sh | sh

2. Start Ollama service:

ollama serve

3. Pull a model:

# Recommended for trading agents
ollama pull llama3.1:70b        # Large, capable model
ollama pull qwen2.5:32b         # Good balance
ollama pull llama3.1:8b         # Fast, lightweight

Verify the downloaded model is listed

ollama list

Model Selection: When you don’t specify an explicit model (e.g., agent_key: ollama), Condor automatically selects one:

Checks environment variables (CONDOR_DEFAULT_LOCAL_MODEL or OLLAMA_MODEL)
Queries the Ollama API (/api/tags) and uses the first available model
If no model is found, prompts you to specify one explicitly

4. Configure in Telegram: Follow the Configure in Telegram steps and select the Ollama - Default Model label.

Local models have a reduced tool surface depending on parameter count — see Tool Filter Modes.

Custom server (optional): To point Condor at a specific model or a remote Ollama server, add the following to your Condor agent configuration (e.g., config/agents.yaml):

agent_key: ollama:llama3.1:70b
model_base_url: http://your-server:11434/v1

LM Studio (Local)

GUI-based local model hosting with OpenAI-compatible API. 1. Install LM Studio: Download from lmstudio.ai 2. Download a model:

Open LM Studio
Browse the model library
Download a model (e.g., Qwen 2.5, Llama 3.1)

3. Start local server:

Go to “Local Server” tab
Click “Start Server”
Default port: 1234
Make sure a model is loaded

4. Configure in Telegram: Follow the Configure in Telegram steps and select the LM Studio - Default Model label.

Local models have a reduced tool surface depending on parameter count — see Tool Filter Modes.

Custom port or model (optional): To target a specific model or a non-default port, add the following to your Condor agent configuration (e.g., config/agents.yaml):

agent_key: lmstudio:your-model-name
model_base_url: http://localhost:1234/v1

Configure in Telegram

Once your provider is installed and authenticated, connect it to Condor from Telegram. 1. Select your provider: Run the /agent command and select Change LLM, then choose your provider using the label from the table below. You should see a confirmation message:

LLM set to <Provider Label>. New sessions will use this model.

Use /agent to continue.

2. Start the agent: Run the /agent command again and click Start. Once the agent is ready you’ll see:

Condor is ready. Send a message to start chatting.

Use /agent to see options or any other command to exit.

You can now send a message to the agent and start chatting.

Provider labels

Provider	Telegram label
Claude Code	`Claude Code`
Gemini	`Gemini CLI`
GitHub Copilot	`Github Copilot CLI`
Codex	`ChatGPT Codex`
OpenRouter	`OpenRouter — Pick Model`
Ollama	`Ollama - Default Model`
LM Studio	`LM Studio - Default Model`

Tool Filter Modes

Local models automatically adjust tool availability based on size:

Essential (≤8B params): Minimal tools, basic operations
Moderate (9-32B params): Common trading operations
Full (>32B params, cloud): All available tools

Larger models handle more complex tool interactions. Cloud providers always get full tool access.

Choosing the Right Provider

For Production Trading

Recommended: Claude Code or ChatGPT Codex

High reliability and uptime
Strong reasoning capabilities
Full tool access
Consistent performance

For Development

Recommended: Ollama or LM Studio

No API costs
Fast iteration
Full control over model
Privacy (runs locally)

Performance Comparison

Provider	Reasoning	Speed	Cost	Reliability
Claude Code	⭐⭐⭐⭐⭐	⭐⭐⭐⭐	💰💰	⭐⭐⭐⭐⭐
ChatGPT Codex	⭐⭐⭐⭐⭐	⭐⭐⭐⭐⭐	💰💰💰	⭐⭐⭐⭐⭐
Gemini CLI	⭐⭐⭐⭐	⭐⭐⭐⭐⭐	💰	⭐⭐⭐⭐
GitHub Copilot CLI	⭐⭐⭐⭐	⭐⭐⭐⭐	💰💰	⭐⭐⭐⭐
Ollama (70B)	⭐⭐⭐⭐	⭐⭐⭐	Free	⭐⭐⭐
Ollama (8B)	⭐⭐⭐	⭐⭐⭐⭐⭐	Free	⭐⭐⭐

Troubleshooting

”Command not found” errors

If Condor can’t find your LLM CLI:

Verify installation:

which claude
which gemini
which copilot
which codex

Check PATH:

echo $PATH

Restart Condor after installing LLM tools:

# If running via make
make stop
make run

# If running via Docker
docker compose restart

Changed LLM but Condor still uses the old one

Change LLM only affects new sessions. Stop any active agent session with /agent → Stop (or send any non-agent command), then run /agent → Start to pick up the new provider. If it still persists, restart Condor:

# If running via make
make stop && make run

# If running via Docker
docker compose restart

Authentication Issues

Claude Code:

claude auth logout
claude auth login

Gemini: Re-run gemini and select a different authentication method, or clear cached credentials and restart the interactive login:

rm -rf ~/.gemini/oauth_creds.json
gemini

Copilot:

gh auth logout
gh auth login
copilot auth login

Local Model Connection Errors

Ollama not running:

# Check if Ollama is running
curl http://localhost:11434/api/tags

# Start Ollama
ollama serve

LM Studio not running:

Open LM Studio app
Go to “Local Server” tab
Click “Start Server”

Model Not Found (Ollama)

If you get “No local model found”:

# List available models
ollama list

# Use exact model name in agent_key
agent_key: ollama:llama3.1:70b-instruct-q4_0

Or set a default:

export CONDOR_DEFAULT_LOCAL_MODEL=llama3.1:70b

Best Practices

Installation Order:

Install LLM provider CLIs first
Authenticate with each provider
Test the CLI tools independently
Then install / start Condor

Security:

Use CLI authentication tools (not environment variables when possible)
Never commit credentials to git
Use separate API keys for development and production

Model Selection:

Start with Claude Code or ChatGPT Codex for best results
Use Ollama 70B+ for cost-effective alternatives
Avoid models under 8B for complex trading strategies

Additional Resources

ACP Documentation

Agent Client Protocol specification

Claude Code Guide

Official Claude Code documentation

Ollama Library

Browse available Ollama models

Trading Agents

Build your first Trading Agent

Documentation Index

​Using the /agent Command

​Supported Providers

​Cloud Providers (ACP Protocol)

​API Key Providers (PydanticAI)

​Local Providers

​Setup Instructions

​Claude Code

​Gemini

​GitHub Copilot

​Codex

​OpenRouter

​Ollama (Local)

​LM Studio (Local)

​Configure in Telegram

​Provider labels

​Tool Filter Modes

​Choosing the Right Provider

​For Production Trading

​For Development

​Performance Comparison

​Troubleshooting

​”Command not found” errors

​Changed LLM but Condor still uses the old one

​Authentication Issues

​Local Model Connection Errors

​Model Not Found (Ollama)

​Best Practices

​Additional Resources

ACP Documentation

Claude Code Guide

Ollama Library

Trading Agents

Using the `/agent` Command

Supported Providers

Cloud Providers (ACP Protocol)

API Key Providers (PydanticAI)

Local Providers

Setup Instructions

Claude Code

Gemini

GitHub Copilot

Codex

OpenRouter

Ollama (Local)

LM Studio (Local)

Configure in Telegram

Provider labels

Tool Filter Modes

Choosing the Right Provider

For Production Trading

For Development

Performance Comparison

Troubleshooting

”Command not found” errors

Changed LLM but Condor still uses the old one

Authentication Issues

Local Model Connection Errors

Model Not Found (Ollama)

Best Practices

Additional Resources