🪩 DiscoMCP

Your agent, meet your tools and context.

Teach any AI agent how you use the tools it connects to — safely, in one command.

Agents are great at code. They get lost inside your tools — they don't know which of a hundred actions matter, how your data connects, or what's safe to touch. So they guess, or they freeze.

DiscoMCP fixes that. Point it at any MCP server and it hands your agent a skill for how you use it — not a generic tool list, but your workflows on your data: the views, tables and records you actually work with, the sequences that answer your real questions, and how one result leads into the next. Learned by looking at your own workspace, and without ever changing a thing.

A tool catalogue tells an agent the server has 90 actions. DiscoMCP tells it the five you actually use, in the order you use them, against the data that's really there.

Why teams use it

Related MCP server: Agent Delegation MCP

Read-only, guaranteed

This is the part that lets you actually turn an agent loose on a real system.

DiscoMCP explores behind a default-deny gate: it runs an action only when it can prove that action is a read. A safe lookup runs. Anything that could write, change, or delete — even if a tool claims it's harmless — is refused. Secrets are stripped from everything it saves.

So your agent can learn your production tools without you holding your breath.

Does it help?

Same question, same server, same model — with and without the generated skill. Read-only, against a genuinely wide, unfamiliar server. n=2 per row.

The harder and less familiar the task, the more it helps: the skill front-loads the map a cold agent has to rediscover by trial. Both reach correct answers — the skill reaches them in far fewer round-trips.

Small sample, directional — not a guarantee. On a trivial task or a narrow server the skill's own prompt cost can wash out what it saves; the durable win is fewer round-trips and steadier behavior on complex servers. Full method in benchmarks/METRICS.md.

Get started

1. Run it — no install needed:

npx @ieranama/discomcp --help

2. Point it at a server — the whole config is a few lines:

[targets.example]
transport = "stdio"
command = "npx"
args = ["-y", "some-mcp-server"]

3. Hand it to your agent and let it explore:

discomcp serve --config ./discomcp.toml

Your agent does the exploring; DiscoMCP keeps it safe and writes the skill. The result lands in .discomcp/profiles/<server>/SKILL.md — ready to drop into your agent.

Under the hood

Built in Rust: the model does the thinking, a small deterministic core enforces every safety check. Every claim in a generated skill is tagged with how it was known — declared, documented, observed, or inferred — so an agent never mistakes a guess for a fact.

• Architecture

• Threat model

• Extension guide

• Configuration example

• Contributing · Security policy

License

Licensed under either of Apache License, Version 2.0 or MIT License at your option.

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in DiscoMCP by you, as defined in the Apache-2.0 license, shall be dual-licensed as above, without any additional terms or conditions.