Changelog
What shipped
Newest first. Only what merged — no roadmap dressed up as release notes.
Findings that say what to do — and when to stop
Every finding now carries a verdict (STOP, FIX, WATCH) and a concrete next step. New detectors catch what actually drives a long session's cost: context re-read on every request, and command cycles whose arguments drift (sed -n 858,866p … 331 times). A live runaway gets one in-session prompt to stop, with the reason. Sessions sort by recency or spend and filter by lane. Subscription seats are now detected from the plan your login reports — Max machines were being labelled API billing.
- spend
- guard
- replay
- cli
Ask the docs from any public page
A research assistant now rests at the bottom of the homepage and docs, answering product questions from the docs themselves. It is mounted on public pages only — never inside the app, where a URL can carry a customer's session id.
- site
Waitlist signups that actually land
Signups were silently failing to reach the database; they now do, deduplicated on a normalized address and rate-limited per client so the one unauthenticated endpoint can't be hammered for free.
- site
- platform
CLI prepared for npm
Packaging for the tokenoscopy CLI: bundled build, bin entry, README, and install guidance that prefers a global binary so the guard skips package resolution on every tool call.
- cli
First Flight: setup that verifies itself
The Setup screen hands you one command with your ingest key already in it, then waits for your first session and shows it to you as it lands — so 'is it working?' is answered by watching it work.
- cli
- replay
Two lanes: seat capacity and billed spend
Each machine now reports how it pays for tokens. API-key and Console-billed spend is real money; subscription usage is shown as what it would have cost, and is never added to money. OAuth no longer implies subscription. Unknown machines say so instead of guessing.
- spend
Loop alerts that fire before the bill
A session that repeats an identical call and edits code that runs unattended — a workflow, a cron, a scheduler — raises one alert linking to the replay, before the schedule it just wired starts billing.
- spend
- replay
Guard latency: 1190ms → 77ms
An allow verdict is now cached locally, so a typical guarded tool call costs a process start rather than a network round trip. Deny and ask are never cached, and every failure path still allows.
- guard
- cli
One design, marketing to product
The site, sign-in and dashboard now share one warm editorial look. The replay stays a dark instrument panel — the recording reads as captured evidence, not another dashboard.
- site
- replay
Attribution and the Flight Report
Spend attributed per session, repo, branch, developer and model, and a weekly report that answers what happened, what it cost, and what to do next.
- spend
OTLP intake
Point Claude Code's native OpenTelemetry logs exporter at Tokenoscopy for per-request spend, including Bedrock and Vertex deployments. The same request from any surface is counted exactly once.
- platform
Budgets and the guard
Budgets per session, repo or org with 50/80/100% alerts, and a PreToolUse guard that stops the next tool call at the cap — fail-open, with every block recorded on the timeline.
- guard
- spend
Session replay
Every Claude Code session becomes a scrubbable timeline of prompts, model calls, tool I/O and diffs, redacted on the machine before upload.
- replay
- cli