The local tools are the immediate payoff: they read what Claude Code and Cline already record and answer what this session cost, which project has consumed the most, how much caching saved, and — less commonly available — which subagent inside the current session is doing the spending. The gateway half adds shared budgets with admission control, where estimated spend is reserved before the request leaves and an over-budget request is rejected rather than discovered later; that half needs a key the project says it cannot currently issue.
A cost server with two halves. The local half reads the session data Claude Code and Cline already write on your machine and needs no account. The gateway half queries an LLMKit proxy, which reserves estimated spend before dispatching a request and refuses one that will not fit the budget.
- llmkit_local_session totals the current session's cost across every AI coding tool it detects, with no API key involved
- llmkit_local_projects ranks cumulative cost across all projects and sessions by spend
- llmkit_local_agents attributes cost to individual subagents inside the current Claude Code session, so an expensive one is identifiable rather than merely suspected
- llmkit_local_cache reports how much prompt caching actually saved across the detected tools
- llmkit_local_forecast projects a monthly cost from local usage and compares it against a Max subscription
- llmkit_usage_stats returns spend, request count and top models for a period, and llmkit_cost_query breaks the same spend down by provider, model, session or day
- llmkit_budget_status reports budget limits and remaining balance, and llmkit_session_summary lists recent proxy sessions with cost, duration and models used
- llmkit_list_keys shows API keys with status and creation date, and llmkit_health checks proxy reachability and response time
Published as @f3d1/llmkit-mcp-server and launched over stdio with npx. The local tools require nothing at all — they read session files the coding tools already produce. The gateway tools need an existing LLMKit API key, and the repository states that account creation and key management are temporarily unavailable while the authenticated service is restored, so without a key already in hand the local half is what works.
One command plus a key — npx -y @f3d1/llmkit-mcp-server, then supply credentials