Skip to main content

Model Context Protocol (MCP) on ColabHive

Status: Live multi-tenant at https://mcp.colabhive.com/mcp since 2026-05-22. Just point your MCP client at it and authenticate with your existing ColabHive API key.

MCP is how an agentic system reaches ColabHive. It is not just a model catalogue behind a protocol: an authenticated agent discovers the public capabilities and private resources visible to its account — models, services, and model operations — and calls them as ordinary tools. API-key scopes are stored but are not an enforcement boundary yet.

This means 3 lines of config in Claude Desktop, Claude Code, Cursor, or n8n get your agent access to:

  • Expert models — every fine-tune, merge, classifier and regressor you own becomes a callable tool
  • LLMs — 10+ open-source families (Qwen, Mistral, Phi, Llama, Gemma, DeepSeek, …)
  • Embeddings and rerankers — retrieval building blocks
  • Forecasting — PatchTST, TimesFM, Prophet, ARIMA, volatility and regime models
  • Classification — text and tabular classifiers
  • Document services — OCR, speech-to-text, translation, moderation
  • Generative — FLUX, SDXL, MusicGen, Kokoro, Bark
  • Web tools — fetch, search, scrape, geocode
  • Model operations — training.merge and training.retrain_on
  • Account-scoped resources — visibility and invocation follow the calling account; explicit account/workload/task node-eligibility policy independently constrains placement

Model operations and lifecycle​

A merge or retrain requested through MCP preserves lineage and can produce a version carrying the candidate lifecycle label. That label is not evidence of a mandatory human approval gate: mandatory automated evaluation gates and an enforced promote/rollback workflow are outside the published MCP contract. Placement first applies the effective account/workload/task node-eligibility policy, then resolves capacity and runtime fit inside the eligible set.

See Private Agentic Infrastructure for the current separation between account-scoped resources, placement and global Burst controls.

What is MCP?​

The Model Context Protocol is an open standard published by Anthropic for connecting AI applications to external tools and data sources. Think of it as USB-C for AI agents: one protocol, plug anything in.

ColabHive provides two ways to use it:

ModeEndpointUse when
Hosted (recommended)https://mcp.colabhive.com/mcpDefault. Zero install. Multi-tenant: bring your own API key in each request.
Local stdiouvx colabhive-mcpAir-gapped, max privacy, or scripting. Your key stays on your machine.

Either way the underlying contract is identical. Pick the one that fits your client.

Two-minute Quickstart (hosted)​

1. Get your API key​

Log in to console.colabhive.com → Settings → API Keys → Create. Use the same hive_... you'd use for any other ColabHive API call — there is no separate "MCP key".

2. Configure your client​

Pick one:

The short version for any HTTP MCP client:

URL:      https://mcp.colabhive.com/mcp
Method:   POST  (request/response) + GET (SSE)
Auth:     X-API-Key: hive_xxx
          OR  Authorization: Bearer hive_xxx

3. Try it from curl right now​

KEY=hive_xxx
curl -sS -X POST https://mcp.colabhive.com/mcp \
  -H "Content-Type: application/json" \
  -H "X-API-Key: $KEY" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' \
| jq '.result.tools | length'

If you get a number back (the count of tools visible to your account), you're done.

Two-minute Quickstart (local stdio)​

For privacy-sensitive setups or scripting where the API key should never leave your machine:

# Recommended: uv (no Python install needed)
uvx colabhive-mcp@latest version

# Or pipx
pipx install colabhive-mcp

Then configure your client to launch it locally (see per-client pages). Each launch picks up COLABHIVE_API_KEY from your environment.

How auth works​

In hosted mode, every request carries its own API key. There is no shared service key on mcp.colabhive.com — each client passes its own hive_... in the header, and the server proxies through to api.colabhive.com as that account.

  • Your account's tools appear in tools/list
  • Your account's billing applies to invocations
  • Other accounts cannot see your trained models
  • Revoke the key in console.colabhive.com → access stops immediately

This is the same model as the underlying Actions API. MCP just wraps it.

Concepts​

  • Manifests — the JSON contract each tool exposes: name, input/output schema, examples, side-effects, cost & latency hints.
  • Architecture — how the hosted endpoint, the local stdio server, and the ColabHive API fit together.
  • Security — auth, rate limits, audit log, hardening.
  • Tools Reference — every tool ColabHive ships, categorized.

Published capabilities​

PhaseWhatStatus
F0Manifest schema in DB + /mcp/manifest endpoint✅ live
F1colabhive-mcp Python package (stdio + HTTP/SSE)✅ on PyPI: colabhive-mcp==0.3.4
F1.5Hosted multi-tenant endpoint mcp.colabhive.com✅ live since 2026-05-22
F3Client integration walkthroughs✅ pages live

See Also​