What did your agents spend last night?What did your agents spend last night… and on what?
Your agents are flying blind
An agent in CI can burn $3K of API spend overnight, and the only record is the invoice.
Tokenoscopy is the flight recorder: every prompt, tool call and dollar, replayable and capped.
npm i -g tokenoscopy && tokenoscopy initEarly-access teams get a key for this on day one.
- Works with
- Claude Code
- Codex CLI· next
- Cursor· next
- Gemini CLI· next
Move through the fog — that's what your agents are doing without a recorder.Drag through the fog — that's what your agents are doing without a recorder.
An overnight backfill stopped by the budget guard at $24.81 — before the next ~$186.
[01] How it works
Record. Replay. Govern.
It rides Claude Code's own hooks and transcripts. No SDK rewrite, no gateway, no proxy.
One command. No proxy.
The CLI wires Claude Code's own hooks. Events stream live; on each stop the transcript is parsed, priced, and redacted on the machine before upload. Your other hooks are never touched, and re-running is idempotent.
CI: init --project + TOKENOSCOPY_KEY · Fleet: init --print → MDM
$ npm i -g tokenoscopy && tokenoscopy init✓ config written✓ hooks installed → ~/.claude/settings.json (backup saved)✓ redaction on (secrets scrubbed before upload)✓ recording — run any Claude Code session and watch it land$ tokenoscopy doctor✓ live hooks installed✓ transcript flush installed✓ budget guard installed✓ ingest key accepted
[02] Runtime contract
Engineered to stay out of your way
It runs inside your agent's hooks, so it is never the reason a session stalled.
1,500ms
guard ceiling
The one hook in your critical path answers inside this or the tool call proceeds. Down, slow, or wrong — it allows.
How →
15s
allow cache
An allow is reused locally, so most tool calls never touch the network. Deny and ask are never cached.
How →
0
proxies in your path
No gateway, no SDK wrapper. Live hooks are async — Claude Code never waits on them. Your traffic never routes through us.
How →
On-device
redaction
Keys, tokens, JWTs, private keys and emails are scrubbed before upload, then again server-side.
How →
[03] Why a recorder
Dashboards watch.
Tokenoscopy can stop.
Observability reports spend after it's gone. Gateways cap tokens but can't see actions.
The agent's own machine is the one place both signals exist.
The dashboard way
Trusted for charts, but the money is already gone.
Spend alert: $1,214
yesterday
Spend alert: $2,891
yesterday
Spend alert: $3,406
2 days ago
Invoice: $9,511
end of month
The Tokenoscopy way
Watch it live, replay it later, stop it in the act.
invoice-backfill · running in CI
$8.11 last request · 2,500 / 8,412 invoices
⛔ Guard: session cap $25.00 reached — Bash blocked
projected ~$186 not spent
Recommendation on the timeline
Batch verifications 50:1 → resume at ~$4 instead of ~$186.
Stopped. Before the next $186.
Elsewhere, budgets are monthly, per-seat or Enterprise-gated.
Ours land mid-session, fail open by design, and show up on the replay.
Caps so good, you'll let agents run overnight
Guard checks answer from a local cache or one fast round trip, and fail open on any wobble.
The worst case is our outage, never your stalled session.
budget · every session
$24.81 spent · Bash blocked · ~$186 not spent
[04] Pricing
Cheaper than what it catches
You're paying for agent capacity either way — by the token or by the seat.
Solo pays for itself the first time it stops a runaway loop.
Solo
$29 / month
14-day free trial · no card
- ✓One developer, every session recorded
- ✓Full replay — prompts, tool calls, diffs, dollars
- ✓Every budget enforced — per session, repo, or org
- ✓Email & Slack alerts before a cap is hit
- ✓Weekly Flight Report · 90-day replay
Team
Custom pricing
for teams of any size
- ✓Everything in Solo, for every developer
- ✓Per-repo, per-dev, per-model attribution
- ✓MDM rollout, SSO, audit export, org policies
- ✓Billed-vs-estimated reconciliation
- ✓Priority support
Every workspace starts with a 14-day Solo trial.
After that it stays read-only until you subscribe — still recording, nothing lost.
[05] FAQ
The questions devs ask first
Every answer links to the docs that back it.
Do you proxy my model traffic?
No. Tokenoscopy rides Claude Code's own hooks and transcripts. Requests go straight from the agent to the model provider; we read what happened, from the machine it happened on. Read more →
What leaves my machine?
Session events and transcripts, redacted on your machine before upload — credential shapes, JWTs, private-key blocks, Authorization headers, connection-string passwords, secret assignments and emails in content. The server runs the same scrub again. Replays include the paths of files the agent read or edited, plus branch drift counted from local git refs — no GitHub access. Read more →
What happens if Tokenoscopy is down?
Nothing you'd notice. The budget guard has a hard 1,500ms timeout and every failure path allows the tool call — outage, network flake, rate limit, bad response. Only an explicit over-budget verdict blocks. We are never the reason an agent stopped. Read more →
Is that dollar figure money?
Depends on the lane, and we always say which. On an API key it's real money, estimated at list price. On a Pro/Max subscription it's what the work would have cost at API rates — nobody is invoiced; what runs out is your usage limit. The two are never added together. Read more →
I sign in with OAuth — doesn't that mean subscription?
Not necessarily. Claude Code also signs into Console accounts that bill per token. We read the plan your Claude login reports instead of assuming, and a machine we can't classify is labelled unclassified rather than guessed. Read more →
Which agents are supported?
Claude Code today, on laptops and in CI. Codex CLI, Cursor and Gemini CLI are next. Read more →
Can I roll it out without touching every laptop?
Yes. Render the hook block once with init --print and ship it through managed settings (MDM). Managed policy can't be disabled by users. CI checkouts use init --project with the key in the environment. Read more →
How is it priced?
Solo is $29 a month for one developer. Team is custom pricing that covers everyone. Every workspace starts with a 14-day Solo trial, no card; after that it's read-only until you subscribe — still recording, with 7 days of replays, but nothing blocked or alerted. Never priced on your spend. Read more →
Stop reading invoices.
Start reading flight logs.
We're onboarding design-partner teams running Claude Code in production now.
Asking about Team? Leave your email and we'll set up a call.