runcycles/cycles-mcp-server
MCP server that gives any MCP-compatible AI agent (Claude Code, Cursor, Windsurf, custom agents) runtime budget, action, and audit authority.
MCP budget and risk authority for Cycles — runtime governance for Claude, Cursor, and Windsurf
- TypeScript96.5%
- Shell2.9%
- JavaScript0.6%
1 Review
cycles-mcp-server is a strong fit for the Cycles ecosystem because it puts budget governance where a lot of agent work is actually happening now: inside MCP-compatible hosts like Claude Code, Claude Desktop, Cursor, Windsurf, and custom agents. The value proposition is clear. Instead of asking every agent implementation to import a language SDK, the server exposes Cycles as MCP tools so an agent can reserve, commit, release, extend, check balances, list reservations, and create usage events from its normal tool loop.
The README is very practical. It explains concrete use cases like per-task dollar caps, multi-tenant SaaS budgets, shared multi-agent workflow budgets, long-running jobs with heartbeats, and fire-and-forget metering. The setup coverage is also good: Claude Desktop config, Claude Code command, generic stdio setup, HTTP transport, environment variables, and mock mode for local evaluation. The tool/resource/prompt surface is well chosen: nine protocol tools, four cycles:// resources, and prompts for integration code, overrun diagnosis, and budget strategy design. That makes the server useful both as runtime infrastructure and as a guided onboarding surface for AI assistants.
The implementation signals are solid. The repo has 150 commits, Apache-2.0 licensing, CI, npm publishing with provenance, MCP Registry publishing, Scorecard, TypeScript build tooling, ESLint, Vitest, Zod schemas, and a clear package manifest for @runcycles/mcp-server plus MCP registry metadata under io.github.runcycles/cycles-mcp-server. The AUDIT.md is especially helpful: it maps all nine MCP tools to Cycles protocol endpoints, checks Zod input constraints, validates resources/prompts, documents previously fixed schema issues, and reports 157 tests with about 97% line and 95% branch coverage.
The biggest improvement I’d suggest is tightening the agent-facing safety guidance. The README says every reservation must be finalized, but MCP agents can be unreliable at following process rules. A short “recommended agent instruction” block, plus examples of how an agent should respond to DENY, ALLOW_WITH_CAPS, expired reservations, and failed commits, would make adoption safer. I’d also add a small compatibility matrix for MCP hosts and transport modes. Overall, this is a polished and well-documented MCP server with a clear purpose and good operational discipline for a young project.
