Guide — Knovaryn as an MCP training-data server¶
Knovaryn is an MCP-native training-data server: a Model Context Protocol
server (knovaryn_mcp) that any MCP-capable agent — Claude, Cursor, a custom
agent host — can drive to build and export training datasets. "MCP training-data
server" and "Model Context Protocol dataset tools" are its packaging story.
What the server is¶
knovaryn_mcp exposes the same application-services core as the CLI, REST
API, and SDK. The agent calls tools; Knovaryn owns the durable state, jobs,
provenance, quality gates, and release artifacts.
| Property | Value |
|---|---|
| Server ID | knovaryn_mcp |
| Resource URI scheme | knovaryn:// |
| Env prefix | KNOVARYN_ |
| Canonical command (stdio) | knovaryn-mcp (or knovaryn mcp) |
| Remote (Streamable-HTTP) | knovaryn-mcp --transport streamable-http --host 127.0.0.1 --port 8000 |
| Dev start from a checkout | uv run knovaryn mcp |
Install & start¶
pip install "knovaryn[mcp]" # includes the MCP extra / framework
knovaryn-mcp # stdio, the default MCP host transport
Or from a source checkout:
uv sync --extra mcp
uv run knovaryn mcp
Connect a client¶
Point your agent at the server. For a source checkout:
{
"mcpServers": {
"knovaryn_mcp": {
"command": "uv",
"args": ["run", "knovaryn", "mcp"],
"cwd": "/absolute/path/to/knovaryn",
"env": {}
}
}
}
Keep env empty (or minimal) in client config — provider keys belong in the
server process environment, never in a shared/committed JSON file.
The toolset (what the agent can drive)¶
The server exposes the full pipeline as tools — project/source lifecycle, dry-run estimation, durable jobs (start/poll/cancel/resume), example preview, lineage, immutable-revision review, validation, versioning, export, and a confirm-gated publish. The authoritative catalogue is generated from the server's real registration: see MCP tools.
With it connected you can ask an agent things like:
"Create project
launch-dataset, add./handbook.pdf(CC0), estimate a run of 500 SFT+preference examples, then start it."
or
"Show me one rejected preference example and its reason; then export version
1.0.0to TRL and Parquet."
Deployment modes¶
- stdio — default, ideal for a local agent session.
- Streamable-HTTP — remote/network hosting behind bearer-token auth: point
the server at a project with real provider config on the server side, and
connect over
--transport streamable-http --host ... --port ....
Both transports are exercised by the automated MCP acceptance matrix across the
supported SDK range (mcp>=1.28,<3) — see the
tool reference.
Related¶
- Guides — MCP client setup, first real project.
- Concepts — overview.