One server, only the tools you need
MCP servers bundle dozens of tools. Capa builds one server that exposes only what your skills actually use.
Open source CLI
Capa is an open-source capabilities manager for AI coding agents. Declare skills, tools, and MCP servers in one file. Capa installs them into Cursor, Claude Code, Codex, and 35+ other agents, or wraps a shadow workspace so nothing lands in your repo.
Step 1
Skills, MCP servers, tools, rules, and hooks live in capabilities.yaml. If it is not in the file, capa will not manage it.
Step 2
capa install fans the file out to your agents and starts the local server. capa wrap does the same in a shadow workspace.
Step 3
Skills, MCP configs, and rules are written into .cursor/, .claude/, and the rest — or into ~/.capa/workspaces if you wrap.
Step 4
localhost:5912 is the same product as the CLI: edit skills and servers, fill secrets, browse registries.
Step 5
The activity monitor traces MCP and shell calls live — duration, errors, and which skill pulled the tool in.
providers:
- cursor
- claude-code
options:
toolExposure: on-demand
skills:
- id: find-skills
type: github
def:
repo: vercel-labs/skills@find-skills
requires:
- '@brave.search' $ capa add vercel-labs/skills@find-skills
Added skill find-skills to capabilities.yaml
$ capa install
Installing for cursor, claude-code
✓ skills/find-skills
Server listening on http://localhost:5912
$ capa wrap cursor
Shadow workspace ~/.capa/workspaces/my-app
Launching Cursor Project Configuration
Edit capabilities, credentials, and activity for this project
Find agent skills on GitHub and add them to the capabilities file.
.cursor/skills/find-skills
Activity
Opening events for each agent run — open a run to follow spans live
Capa is an open-source capabilities manager for AI coding agents. You declare skills, MCP
servers, tools, rules, hooks, plugins, and agent instructions in one capabilities.yaml file,
then run capa install or capa wrap so Cursor, Claude Code, Codex, and
35+ other agents get the same setup. Capa does not auto-discover leftover files under
.cursor/ or .claude/: if a skill or server is not in the file, capa
will not manage it. Managed install writes provider configs into the project and starts a local
server around localhost:5912. Wrap runs the agent from a shadow workspace under
~/.capa/workspaces so provider files never land in git. Passthrough writes native
provider files and skips the capa proxy. Skills can come from GitHub, GitLab, registries, or a
local path, and secrets stay out of the file as ${VarName} placeholders. Capa
is built by Infragate.
MCP servers bundle dozens of tools. Capa builds one server that exposes only what your skills actually use.
Stop duplicating skills across projects. Reference them from GitHub, GitLab, registries, or a local path.
Run Cursor, Claude Code, and other agents from ~/.capa/workspaces so provider configs never land in git.
Every configured tool is a command: capa sh <server> <tool>. Same credentials, no wrappers.
Agents start with a minimal set and pull the rest when a task needs them. Flip options.toolExposure anytime.
Block dangerous phrases and sanitize skill content before install. Allowed character sets and blocked lists live next to the capabilities.
One binary on your PATH. Then capa init, add a skill, and
capa install — or skip the in-repo configs and
capa wrap instead.
Need the full path? The installation guide covers upgrades, PATH, and Windows details.
curl -LsSf https://capa.sh/install.sh | sh powershell -ExecutionPolicy ByPass -c "irm https://capa.sh/install.ps1 | iex"