Use ColabHive with SuperClaw / OpenCode

SuperClaw is Intel's hybrid, local-first agentic AI: it runs agents on the AI PC and keeps sensitive, high-frequency work on-device, while offloading heavy reasoning to a cloud model. ColabHive is a natural fit for that cloud tier — an OpenAI-compatible inference cloud that runs natively on Intel Arc — so the stack can stay on Intel silicon, edge to cloud, when the cloud tier runs on Arc nodes.

Because SuperClaw is built on OpenCode, the same setup works for OpenCode and any OpenAI-compatible agentic client. For OpenCode on its own, start with Coding Agents.

Why ColabHive as the cloud tier

• OpenAI-compatible with tool calling. Agents pass their tools; the model returns tool_calls; the agent executes them locally. See Tool Calling.
• Arc-native. The heavy tier runs on Intel Arc GPUs instead of requiring every user to operate a dedicated multi-GPU workstation — elastic and pay-per-use.
• Private lanes for regulated data. Run the cloud tier on Arc nodes you control, with an Enterprise Private Cluster (or, for labs and research, self-hosted Community, free under BSL 1.1), so you choose where the heavy reasoning is processed while SuperClaw keeps its PII handling on the device.

Configure ColabHive as the cloud provider

In SuperClaw → Advanced → Model Routing, add a cloud model provider with:

Field	Value
Base URL	https://api.colabhive.com/v1
API key	your ColabHive key (hive_...) — see Authentication
Model	a model id from GET /v1/models that returns tool calls — colabhive agents models lists them, warm models first

For OpenCode, colabhive agents init writes the provider for you and checks that the model returns tool calls — see the Coding Agents quickstart. The provider block by hand is in OpenCode setup.

Verify the connection:

curl https://api.colabhive.com/v1/chat/completions \
  -H "Authorization: Bearer $COLABHIVE_API_KEY" -H "Content-Type: application/json" \
  -d '{"model":"<model id>","messages":[{"role":"user","content":"hello"}]}'

Drive the ColabHive platform from your agent (MCP)

Beyond the cloud model, your agent can call ColabHive's platform — fine-tune on proprietary data, run specialists (embeddings, rerank, forecasting), deploy a model as an endpoint — via the Model Context Protocol. This complements the chat model: the model reasons, MCP tools act on the platform.

The MCP server exposes two kinds of tools, and they follow different naming conventions:

• Built-in management tools — snake_case (e.g. list_endpoints, run_inference, list_trainable_models, create_training_run). These are fixed and documented in the MCP tools reference.
• Action tools — one per public endpoint, named by its kebab-case slug (which is the endpoint's endpoint_name, e.g. qwen-2.5-7b-instruct-public, embeddings-public). These are discovered live from the catalog, so the exact set is whatever is public right now.

Notes

• Check that the model returns tool calls before relying on it — see Known limits.
• A cold model makes the first request wait while it loads — see Known limits.
• For data-residency deployments, ask about private lanes so the heavy tier runs on your own Arc nodes.