Models
Choosing and switching models across supported providers.
synapse modelOpens the model picker: choose a provider and model without code changes or lock-in. Mid-conversation, use /model [provider:model] — the switch takes effect on your next message and the conversation continues seamlessly.
Supported providers
The repository bundles 35+ provider profiles as plugins, including: OpenRouter, OpenAI, Anthropic, Gemini, DeepSeek, Ollama Cloud, AWS Bedrock, Azure Foundry, Vertex AI, xAI, GitHub Copilot, HuggingFace, Fireworks, Together-class endpoints via custom profiles, and a generic custom endpoint.
| Provider | Good for |
|---|---|
| OpenRouter | One key, hundreds of models — easiest start |
| OpenAI | GPT models with first-class tool calling |
| Anthropic | Claude models, strong long-context work |
| Gemini | Huge context windows, multimodal input |
| DeepSeek | Cost-effective reasoning |
| Ollama Cloud | Open-weight models without local hardware |
| custom | Any OpenAI-compatible base URL you control |
Choosing a model
- Daily driver — a mid-tier model with solid tool calling keeps costs sane for routine work
- Heavy reasoning — swap to a stronger model with
/modelfor architecture or debugging sessions, then switch back - Cheap background jobs — point cron and subagent workloads at an economical model; they do not need your best one
Model health checks
synapse doctor includes health checks for API-key provider profiles, so you can verify a model is reachable before relying on it. If a provider starts failing mid-day, run doctor first — it is usually a key or quota issue.
Reasoning & thinking (always-on)
Synapse keeps reasoning/thinking enabled by default and never lets a config value turn it off — the always-on reasoning policy. Models that support reasoning tokens always use them; models that don't are unaffected (no illegal provider parameters are sent).
| Level | Trade-off |
|---|---|
| Medium | Balanced speed and cost (the default) |
| High | Deeper reasoning — slower and costlier per turn |
| Max | Strongest reasoning — slowest and costliest per turn |
Set the level via /reasoning medium | high | max in the CLI, the Reasoning picker in the dashboard sidebar, or agent.reasoning_effort: medium in config.yaml. Per-model overrides use agent.reasoning_overrides: { "model-id": "high" }. Legacy none/false/off/disabled values all resolve to medium so reasoning is never silently turned off.