Agent Usage

New

• See what your AI tools actually burn — A new Usage view reads the records your local AI tools already write — Claude Code, Codex, Grok, Cursor, Qwen, OpenCode, Hermes, and OpenClaw — and turns them into tokens, sessions, and list-price cost estimates, all on your machine

• Browse every session — A paginated Sessions table lists turns, tokens, and cost per session, with one click to reveal the source record

• Watch quota pressure before it bites — The Quota tab lines up each provider's month — tokens, plan, cap-hit days — alongside the rate-limit windows vendors report

• Get a monthly plan recommendation — The Advisor reviews your covered history and suggests the plan tier that fits each provider, showing the arithmetic behind every suggestion — it stays quiet until at least 30 days of real coverage exist

• Capture Claude's own quota numbers — An opt-in statusline shim records the rate-limit percentages Claude Code reports, while your existing statusline keeps working underneath it

• Private by construction — Collection is off until you turn it on, prompts are stored as hashes and byte counts only — never prompt text — and Clear removes the derived data without touching the original records

• Define MCP servers once, use them everywhere — A new Registry view holds your MCP server definitions in the vault; flip a binding to choose which AI clients receive each server

• Apply with a reviewed preview — Apply opens a per-file diff of adds, updates, and removals before anything is written; every write is atomic and keeps a backup, and a file that changed since the preview refuses to apply

• Adopt what your tools already have — Pull a client's existing MCP entries into the registry as managed definitions; env values are discarded and only variable names are kept as references

• Keep secrets as references — Env values are stored encrypted in the OS keychain and shown as names only; they are written into client config only when you explicitly enable materialization for that apply

• See drift at a glance — Each client reports in sync, drifted, not applied, or no file, with a reveal button that opens the exact config being tracked

• Safety rails built in — Entries that point at vendor API endpoints or look like pasted credentials are refused, and configs the vault cannot parse cleanly are never rewritten

• Publish a portable export — An opt-in export writes registry definitions and vault skill content — no env values, no secrets — to a JSON file other tools can import