Skip to content

Skills and Tools (MCP)

Custom agents can carry more than a model and a system prompt. Two libraries in Settings extend what an agent knows and what it can call: Skills (instruction documents loaded on demand) and Tools (MCP servers). Both are attached per agent in the custom-agent editor, and both are snapshotted onto a session when it spawns — editing a skill or server later never changes an agent that’s already running.

A skill is a named Markdown document of instructions — a code-review checklist, your team’s migration conventions, how to write a changelog entry. Manage them under Settings › Skills: each has a name, a one-line description, and a Markdown body.

Skills load on demand rather than being stuffed into every prompt. At spawn, each attached skill is written as a file into the agent’s sandbox, and a compact index — one line per skill: name, description, file path — is appended to the agent’s instructions. The description is all the agent sees until it decides the task matches and reads the file. That keeps the per-turn cost of carrying ten skills at ten lines, not ten documents.

Skills are provider-neutral: the same skill works on any base agent, because it’s just a file and an index line.

Settings › Tools holds your library of Model Context Protocol servers. Each server is a name plus a transport:

  • stdio — a command to run, with optional environment variables.
  • HTTP — a URL, with optional headers (for auth tokens and the like).

Attach servers to a custom agent in the agent editor and the agent gets those tools in every session. What can actually be delivered depends on the base agent, and the editor filters accordingly: Claude Code accepts both stdio and HTTP servers, Codex accepts stdio only, and the other agents currently have no MCP surface Fletch can drive.

The exact server configuration — command, env, URL, headers — is snapshotted onto the session at spawn, so a running agent’s toolset is stable even while you edit the library.