Skip to main content
Glama

Capsa: MCP server and capsa

Your saved prompts, skills, MCP servers and design boards, readable by your AI tools and your terminal. Read-only, offline, no dependencies.

Capsa is a local-first library for what you learn about working with AI. Every item keeps its evidence: the post, video moment or README line it came from, quoted exactly. This repository is Capsa's public source. It grows in releases. Release 1 (0.1.0) has two parts:

Part

What it does

packages/capsa-mcp

A local MCP server. Claude, Codex, Cursor, VS Code and other MCP clients can search your library, read items, fill a prompt's blanks and read a board's design brief.

packages/capsa

The same library in your terminal: search, list, show, fill and copy.

Both read one file, capsa-library.json. Neither writes files, opens a network connection or needs npm install.

Status: early. Version 0.1.0. The file format is capsa.mcp.v1 and may grow, but new fields are added, not renamed. The Capsa app that writes the file comes to this repository in later releases; see the changelog.

Requirements

Node.js 20 or newer. The tests run on Node 20, 22 and 24 on Linux in CI, and were run on Node 24 on macOS for this release. Windows hasn't been checked.

Related MCP server: Skills Registry

Try it in a minute

git clone https://github.com/obileonardo/capsa-oss.git
cd capsa-oss
node packages/capsa/capsa.mjs search release --file packages/capsa-mcp/examples/sample-library.json

examples/sample-library.json is a small made-up library: two prompts, an MCP server and a board.

Your library file

The Capsa app writes it: Your AI › Download the library file. You can also write or generate one yourself; the format is plain JSON. Keep it anywhere, for example ~/capsa-library.json. Download it again when your library changes: both tools notice a newer file and read it again.

Use it from your AI tools

Use absolute paths. ~/ works inside the server's arguments, but not every client expands it elsewhere.

Claude Code

claude mcp add --transport stdio capsa -- node /path/to/capsa-oss/packages/capsa-mcp/server.mjs /path/to/capsa-library.json

Codex

codex mcp add capsa -- node /path/to/capsa-oss/packages/capsa-mcp/server.mjs /path/to/capsa-library.json

Cursor (.cursor/mcp.json) and Claude Desktop (claude_desktop_config.json)

{
  "mcpServers": {
    "capsa": {
      "command": "node",
      "args": ["/path/to/capsa-oss/packages/capsa-mcp/server.mjs", "/path/to/capsa-library.json"]
    }
  }
}

VS Code (.vscode/mcp.json)

{
  "servers": {
    "capsa": {
      "type": "stdio",
      "command": "node",
      "args": ["/path/to/capsa-oss/packages/capsa-mcp/server.mjs", "/path/to/capsa-library.json"]
    }
  }
}

Instead of the second argument you can set CAPSA_FILE.

What the server offers

Kind

What

Tools

search_library, list_items, get_item, list_collections, render_prompt, list_boards, get_board_brief. All are marked read-only.

Prompts

Each saved prompt or template, with its blanks as arguments (client support varies).

Resources

capsa://item/<id> and capsa://board/<id>/brief, as Markdown.

Transport: stdio, newline-delimited JSON-RPC 2.0. Protocol versions 2025-06-18, 2025-03-26 and 2024-11-05.

Saved text is quoted content, not instructions. Items come from posts, READMEs and transcripts you saved. The server fences that text and tells the client it is material to use, not instructions to follow. A client can still be fooled by text in your library, so save from sources you trust.

Use it in your terminal

alias capsa="node /path/to/capsa-oss/packages/capsa/capsa.mjs"
export CAPSA_FILE=~/capsa-library.json

capsa search deploy checklist        # every word must match
capsa list --type prompt --starred
capsa show <id>                      # the text, and where it came from
capsa render <id> subject="a lighthouse"
capsa copy <id> subject="a lighthouse"   # to the clipboard
capsa collections
capsa boards
capsa brief "calm onboarding"

Without --file or CAPSA_FILE, capsa looks for ~/Downloads/capsa-library.json. Add --json for machine-readable output. copy uses your system's clipboard command (pbcopy, wl-copy, xclip, xsel or clip); nothing else runs.

From npm: npm install -g capsa-cli (the command is capsa), and npx -y capsa-mcp@latest in your AI tool's config. From a clone, keep the two folders side by side, as in this repository.

The file format

capsa.mcp.v1 is one JSON object (files from before 29 Oct 2026, when Capsa was called LibraryKit, say librarykit.mcp.v1 and still load, as do the old librarykit-library.json name and LIBRARYKIT_FILE):

Field

What

format

Always "capsa.mcp.v1". Other files are refused.

items

Saved items: id, type (prompt, template, skill, mcp_server, workflow, technique, command, tool, reference...), typeLabel, title, text, exact, yours, summary, evidence, quote, source (title, url, kind), tags, starred, collections, projects, savedAt

collections, projects

id and name (projects also note)

boards

Optional. id, title, project, references (credit, link, rights, palette, what you like and avoid) and brief (Markdown)

The sample file shows every field. Blanks in a prompt are written [like this] or {{ like this }}.

Checks

npm test     # both packages; no install needed

Contributing

Issues and pull requests are welcome for what is in this repository. Read CONTRIBUTING.md first: every commit is signed off under the DCO. For security problems, see SECURITY.md.

Licence

The code is under the MIT License. The Capsa name and the mark are not; see TRADEMARKS.md.

Related MCP Connectors

Related MCP Servers

  • A
    license
    Not graded
    quality
    D
    maintenance
    Enables AI assistants to search and read offline documentation from synced libraries via MCP tools, providing hybrid keyword and semantic search without network calls.
    18 npm
    1
    MIT
  • A
    license
    Not graded
    quality
    B
    maintenance
    Enables AI clients to discover, search, retrieve, and read supporting files from a centrally published, versioned skill registry through four read-only MCP tools.
    29 npm
    MIT
  • A
    license
    A
    quality
    C
    maintenance
    Enables MCP-capable clients to search team knowledge, browse an integration catalog, and get sourced product answers through deterministic, offline tools.
    3
    5 npm
    MIT