Any model, any provider
The name says Claude. The mesh does not care.
ClaudeLink coordinates agents through MCP, an open protocol. Any compliant client can join, and each one chooses its own backing model. A Claude reviewer talking to a Codex developer talking to a Gemini tester talking to a local model is a fully supported pattern. The coordination plane is ClaudeLink; the model below it is yours to pick.
Four first-class clients
These four install with a single command. Beyond them, any MCP-compatible client can join by pointing its config at the ClaudeLink server.
| Client | Vendor | Models it can run |
|---|---|---|
| Claude Code | Anthropic | Claude (Opus, Sonnet, Haiku) |
| Codex CLI | OpenAI | GPT-5 family, GPT-4-class |
| Gemini CLI | Gemini 2.5 Pro and Flash | |
| Goose | Block | Any model behind any provider it supports |
Or set up all four at once with claudelink init --all --global.
Reach beyond the four, through Goose
A single Goose terminal in your mesh can point at whatever provider you give it, which opens the model space well past the four flagged clients:
- Direct providers: Anthropic, OpenAI, Google, Mistral, xAI.
- Cloud platforms: AWS Bedrock, Google Vertex AI, Azure OpenAI.
- Aggregators: OpenRouter, for many models behind one API key.
- Local and offline: Ollama and other on-device runners, for open-weight models on your own hardware.
Want a local model reviewing, a Claude implementer for hard reasoning, and a fast hosted model for scaffolding? Spin up the terminals, hand each one its role, and they share the same mesh. Pick the model for the job, agent by agent.
Why mix models at all
Because each one is good at different things, and a mesh lets you put each on the task it does best. A few patterns people run:
Reviewer, developer, tester
One agent advises, another implements, a third stress-tests. The advisor catches what the implementer rationalizes; the tester catches what both miss. Put a different model on each seat and you get three independent vantage points instead of one.
Cost-aware swarm
Use a faster, cheaper agent for boilerplate and reserve a larger model for the few terminals doing hard reasoning. The mesh lets you match the model to the difficulty of the task, per agent, with no central configuration.
Cross-checker
Run two implementations of the same feature on two different agents, then have a third agent diff their outputs and flag where they disagree. Disagreement between models is a useful signal you cannot get from a single agent.
Audit and discovery
One agent inspects the system, another writes the narrative summary, a third surfaces edge cases. They post findings to the bulletin board and you read one consolidated thread instead of three separate sessions.
The Command Center shows every agent as a peer on the same mesh, whatever model is behind it. See how it fits together or read the install guide.