An MCP server for ML experiment tracking
Before your coding agent starts the next run, it should know what your team already tried. Probe's hosted MCP server gives any MCP client read access to your team's experiment record: projects, experiments, runs, metrics, notes and papers.
https://mcp.research.prbe.ai/mcp
Connect
When your agent records its own work, the Probe plugin wires the server in for you; in the default setup, where the Probe daemon records, your agent asks with probe ask instead. Tell your coding agent:
Install Probe on this machine and its coding agents. Read https://research.prbe.ai/install.md and follow it.
Or add it to any MCP client yourself. In Claude Code:
claude mcp add --transport http probe-research https://mcp.research.prbe.ai/mcp
The server signs you in with OAuth the first time a client connects. It only reads; writes go through the probe CLI and SDK.
Six read-only tools
| Tool | Answers |
|---|---|
browse | What exists here: the project, experiment and run tree |
search | Find things about X, with where each result came from |
entity | Show me this thing, through a purpose-shaped view such as card, metrics, trajectory, lineage or branches |
metrics | What are the numbers, for one run or several runs side by side |
query_sql | What exactly do the tables say, read-only SQL over your team's research tables |
find_papers | What does the literature report |
Responses stay inside a token budget the agent sets, 2,000 tokens by default. See MCP server.
What changes for you
- Your agent won't repeat what your team already tried. It can search past experiments before it proposes one.
- Answers come with sources. Each result says which run, experiment or note it came from.
- Nothing gets written by accident. The server is read-only.