Agent access (MCP)

One local MCP connection to your whole library, its key, and Connect a project for one topic.

The app serves your library over MCP on your computer. One connection gives Claude Code, or any agent that speaks MCP, every topic: it can search, read packs and get a hand-off brief while it works.

Set up Claude Code

Run this once, in any terminal. It works in every project from then on.

Terminal
claude mcp add --transport http --scope user startwell-capture http://127.0.0.1:43127/mcp --header "Authorization: Bearer <your key>"

The app shows this with your key filled in: open Settings, Agent access (it is also in the palette, ⌘K or Ctrl+K) and copy it, or click Set up Claude Code. The app asks first, then adds that one server to ~/.claude.json and changes nothing else. Test connection does what an agent does and shows the tools it found.

The server lives in the app. Agents can only reach it while Startwell Capture is running. If the tools fail with a connection error, open the app. It starts when you log in unless you turn that off.

Other agents

Cursor: ~/.cursor/mcp.json
{
  "mcpServers": {
    "startwell-capture": {
      "url": "http://127.0.0.1:43127/mcp",
      "headers": { "Authorization": "Bearer <your key>" }
    }
  }
}
VS Code: .vscode/mcp.json
{
  "servers": {
    "startwell-capture": {
      "type": "http",
      "url": "http://127.0.0.1:43127/mcp",
      "headers": { "Authorization": "Bearer <your key>" }
    }
  }
}

Anything else that speaks MCP over HTTP takes the address http://127.0.0.1:43127/mcp and the header Authorization: Bearer <your key>. Codex goes in ~/.codex/config.toml; the app shows that one too.

What an agent can do

Pass topic (a name or id) to any tool to narrow it to one topic. Leave it out to cover everything.

ToolWhat it does
list_topicsYour topics and what each holds.
topic_contextWhat you have been doing: a topic’s memory and recent captures, or everything.
search_topicSearch notes, transcripts, screenshot text and links.
get_captureOne capture in full.
list_packsContext packs on disk.
get_packOne pack: its README and where its files are.
get_handoffA ready-to-follow brief: what to do, what you raised, and the paths to read.
list_issuesWhat you raised: issues, requests, ideas, questions and decisions, with quotes and screenshots.
add_noteAdd a note to your timeline, for example a decision the agent made.

Connect one project to one topic

Settings, Agent access: the local server, Ask for a key, the Claude Code command and setups for other agents.
Settings, Agent access.

For a repo that should only see one topic: in the topic, open Agent access and click Connect a project, then pick the project’s folder. Two files are written there:

  • .mcp.json: the topic’s server is added. Anything already in the file is kept.
  • .claude/skills/startwell-capture-<topic>/SKILL.md: a short skill that tells Claude Code what the topic is and when to use it, for example at the start of a task or when you paste a pack path.
.mcp.json
{
  "mcpServers": {
    "startwell-capture-checkout-redesign": {
      "type": "http",
      "url": "http://127.0.0.1:43127/mcp/<topic id>"
    }
  }
}

Open Claude Code in that project and approve the server when asked. To use a topic elsewhere, copy its address from the Agent access tab into your agent’s MCP settings.

Who can reach it

The server listens only on this computer (127.0.0.1, port 43127) and turns away requests from web pages. The library address reaches every topic, so it asks for a key made for this install. The key is shown in the setup you copy; New key replaces it, and connected agents then need their setup copied again. An agent with no key, or an old one, is turned away, and Settings says so. You can switch the key off.

A topic’s own address reaches only that topic and needs no key, because it is written into a repo’s .mcp.json, where a secret should not go. It contains the topic’s long random id. The key keeps out other programs that merely find the port; a program running as you can also read the folders on disk. The server only runs while the app does. Turn it off in Settings, Agent access.