Connect

MCP

Point your AI tool at TokenCap once. After that the agent asks the local server for context instead of reading the whole repo.

1 / 2

Connect your AI tool

Give Claude Code, Cursor, Codex and other MCP hosts 19 local TokenCap tools.

tokencap serve --mcp --init --client claude-code

Writes .mcp.json in the project root.

1
Write the config
tokencap serve --mcp --init
```
✓ cursor config written to .cursor/mcp.json
  Start the server: tokencap serve --mcp
```
Example output
Detects your host. Add --client <host> to pick one.
Check it works
tokencap serve --mcp --health
List the tools
tokencap serve --mcp --tools

Restart the host after --init. The host starts tokencap serve --mcp itself, builds intelligence in the background and keeps it fresh with a watcher, so you don't need a separate tokencap make.

--init merges into an existing config and keeps other servers. An existing tokencap entry is kept; pass --force to replace it.

The npm package also ships an Agent Skill in skills/tokencap/ for hosts that load skills.

All flags (6)

--init: Write the MCP config for a host
--client <host>: Choose the host instead of detecting it
--force: Replace an existing tokencap entry
--health: One-shot health check
--test: Call every tool and verify output
--tools: List available tools