What the agent gets
15 meta-tools. The agent never holds a credential.
The server exposes a small set of tools that route to the right action. Every call runs under the session's mandate; requests matching your approval rules wait for review; requests outside hard limits are blocked.
Sandbox run · demo data
Hosted or local
Same tools, same mandate. Pick where it runs.
Hosted
connect.codespar.dev/mcp
Nothing to install. Keys stay in your environment; the server runs on ours. Rate limits per key.
Local
npx -y @codespar/mcp serve
Runs on your machine over stdio. Same 15 tools, same rules. MIT on npm.
Sandbox run · demo data
Works with
Any client that speaks MCP.
Claude Code
One command.
claude mcp add --transport http codespar https://connect.codespar.dev/mcp --header "Authorization: Bearer csk_test_…"
Cursor
mcp.json at the project root.
Add the hosted endpoint under mcpServers with the Authorization header.
Codex CLI and others
Same endpoint, same header.
Any MCP client with HTTP transport. Local stdio server also available: npx -y @codespar/mcp serve.
Install in one click
Cursor and VS Code open with the hosted endpoint and the Authorization header filled in. Replace csk_test_your_key_here with your key.
- Run the command in a terminal. It registers the hosted endpoint; nothing runs on your machine.
- Prefer a local server? The second command runs @codespar/mcp over stdio, with the same tools.
- Prefer code? Claude also has a first-party adapter, @codespar/claude, with getTools(session).
Every snippet carries csk_test_your_key_here: replace it with a key from Dashboard · API keys. A csk_test_ key runs in the sandbox; a csk_live_ key is the switch to production.
For AI agents reading this page
This block is written for an AI agent. Paste it, or the link to this page, into an agent: it adds the codespar MCP server, finds the right tools and reports back.