MCP server
The Node package ships commoncompute-mcp, a Model Context Protocol server that gives AI coding agents — Claude Code, Cursor, Continue, anything that speaks MCP — typed tools for chat, embeddings, transcription, image generation, and the full job lifecycle.
Claude Code
Cursor
Add to ~/.cursor/mcp.json:
Continue
Add to the mcpServers block of your Continue config:
Authentication
The server reuses your CLI login — run commoncompute login once and it picks up the stored credentials. Alternatively, set CC_API_KEY in the server's env block.
Tool catalog
- cc_chat — OpenAI-compatible chat completion
- cc_embed — text → vector
- cc_transcribe_audio — Whisper transcription (path or base64)
- cc_generate_image — prompt → image
- cc_quote_job — cost / latency estimate before submitting
- cc_submit_job — submit a generic workload
- cc_get_job / cc_wait_for_job / cc_get_job_events — job snapshot, poll to terminal, activity feed
- cc_get_job_result_url / cc_preview_job_result — presigned result URL, bounded inline preview
- cc_list_workloads / cc_get_workload — catalog discovery
- cc_list_models / cc_get_model — model discovery
- cc_whoami, cc_get_balance, cc_get_billing_history, cc_get_usage, cc_list_keys — read account state
Agents can also browse resources before tool-calling: cc://workloads, cc://models, cc://jobs/{id}, and cc://jobs/{id}/events.
Scopes
Destructive tools are default-deny: cc_cancel_job requires the +jobs:cancel scope, and cc_create_key / cc_revoke_key require +account:write. Opt in via CC_MCP_SCOPE in the server env: