Ollama

The ollama provider calls a locally running Ollama server over HTTP. No API keys, no network calls to third-party services — all inference runs on your machine. This makes ollama the best choice for privacy-sensitive codebases or air-gapped environments.

Prerequisites

• Ollama installed and running locally
• The model you intend to use pulled to your local Ollama instance

Install Ollama and pull a model:

# Install Ollama (macOS)
brew install ollama

# Start the Ollama server
ollama serve

# Pull the default model
ollama pull qwen2.5-coder:7b

Verify the server is reachable:

curl http://127.0.0.1:11434/api/tags

Authentication

No authentication is required. Corydora connects to the Ollama server at the address specified by OLLAMA_HOST, which defaults to http://127.0.0.1:11434.

To use a remote or non-default Ollama instance, set OLLAMA_HOST in your environment or in .corydora/.env.local:

OLLAMA_HOST=http://192.168.1.100:11434

corydora doctor reports this check as ollama-auth. Because Ollama has no authentication layer, the check always passes as long as the host is configured:

ollama-auth   ✓  Ollama host configured at http://127.0.0.1:11434.

Models

Model	Notes
qwen2.5-coder:7b	Default. Strong code understanding at a manageable size.
deepseek-coder-v2	Alternative with strong multi-language support.

Any model available in your local Ollama instance can be used. Pass the model name exactly as it appears in ollama list.

Pull additional models before running:

ollama pull deepseek-coder-v2

Execution mode

ollama uses single-file-json execution. For each file being processed:

1. Corydora reads the file content and builds a structured prompt.
2. A POST request is sent to <OLLAMA_HOST>/api/generate with stream: false.
3. The response field of the JSON payload is extracted and parsed as scan findings or fix instructions with fileEdits.
4. If fileEdits are present, Corydora writes the replacement content to disk.

The stream: false parameter ensures Corydora receives the full response as a single object.

Example configuration

{
  "runtime": {
    "provider": "ollama",
    "model": "qwen2.5-coder:7b"
  }
}

With a custom host:

{
  "runtime": {
    "provider": "ollama",
    "model": "qwen2.5-coder:7b"
  }
}

# .corydora/.env.local
OLLAMA_HOST=http://192.168.1.100:11434

Troubleshooting

Ollama request failed with 404

The specified model is not available in your local Ollama instance. Run ollama list to see what is installed and ollama pull <model> to add it.

Ollama request failed with connection error

The Ollama server is not running. Start it with ollama serve, or verify the address in OLLAMA_HOST is correct and reachable.

Response did not include valid scan JSON

The model responded but Corydora could not parse a JSON object from the output. Smaller models may not reliably follow the structured output format required by Corydora. Consider switching to qwen2.5-coder:7b or a larger variant if you see this consistently.

Slow performance

Corydora respects the maxConcurrentScans setting in .corydora.json. Reduce this value if your machine cannot handle parallel inference requests without degraded performance:

{
  "scan": {
    "maxConcurrentScans": 1
  }
}