Every token that entered the model, in load order: system prompt, tool specs, injected reminders, tool results. Not a summary, the real breakdown, against real cost, over a real MCP connection. No rewritten agent loop, no wrapper.
One click, no password, and no account to create here. Sign in with Google and your token, dashboard, and setup command are ready on the next screen.
What you get
Per-session tokens, cache-hit rate, and dollar cost, updated within seconds of each turn. Eight MCP tools, seven read-only.
Exactly what entered the model's context window, block by block, in load order, with honest token estimates for every segment.
Your agent connects with a genuine Model Context Protocol handshake and streams its own telemetry. Nothing mocked, no fake data.
Trust & privacy
Sign-in mints a bearer token tied to your Google account; CtxWindow never sees your Google password. Session metrics (tokens, cost, tool names) are stored by default. Full prompt/response content is only captured if you turn on the optional raw-body flag, and even then it passes through basic redaction before storage. Your data is never visible to another signed-in user, only to you and the server owner.
MCP (Model Context Protocol) is an open standard that lets an LLM or agent call tools over a
normal HTTP connection. This server exposes read/write tools for agent execution
data, and is built to support two integration paths: Bedrock-based agents calling the
record_session tool directly, and Claude Code.
No account to create or password to remember here. Google verifies who you are, this server just checks the signed proof and hands you a token scoped to your account. That token, not your Google identity itself, is what your agent actually uses afterward.
Every session recorded through your token is tagged with your account. Reads are filtered the same way, you only ever see, list, or query sessions you recorded. Even guessing another person's session ID reads back as "not found," identical to one that never existed.