Source-side AI coding telemetry

Know what your AI
coding tools cost.

Tokenbase reads the session logs your terminal coding agents already write to disk, normalizes them, and shows engineering leaders team-level spend and utilization. Never in the model request path. Metadata-only by default.

Spend · last 14 days
$9.0k
AFactual
Claude Code
60%
Codex CLI
31%
Copilot CLI
9%
Cursor
0%
Adapters
Coverage stated per tool
Claude CodeCore
Codex CLICore
Copilot CLICore · partial
CursorBeta
Healthy

Reads the logs your agents already write

Claude CodeCodex CLICopilot CLICursor

Every developer runs several AI coding tools, and nobody can say what they cost or whether they pay off. The data already sits on every developer's disk — siloed per vendor, per machine. Tokenbase collects it safely and makes it legible by team.

How it works

Collected on the device. Analyzed in the cloud.

Not local-first — source-side collection with configurable residency. The collector buffers locally; analysis is cloud-side, or your VPC, or a local model for zero egress.

01
Collect

A signed, device-scoped collector tails local session logs — read-only, checkpointed, never the request path.

02
Normalize

One adapter per tool, each versioned with golden fixtures, maps every harness into one schema with coverage stats.

03
Attribute

Facts roll up into team-level spend and utilization — aggregate only, behind a cohort threshold.

Built for security review

You choose what leaves the device.

The deployment matrix is the first thing security asks for, so it leads. By default raw transcripts never leave the machine — only structured metadata does.

ModeLeaves the device
SaaS · metadata-onlyStructured metadata only
SaaS · redacted textRedacted excerpts, opt-in
VPC / on-premConfigurable
Local-only pilotNothing
Honest by construction

Token counts answer “how much,” not “was it worth it.”

Tokenbase labels every number by confidence tier and never prints an inferred figure without saying so. Spend is a fact from the logs — not a guess dressed up as one.

AFactual

Factual. Tokens, models, session counts, tool calls, estimated cost — straight from the logs.

BInferred

Inferred. Abandonment, retry loops, likely workflow type — heuristics on local data, labeled as such.

Spend, attributed by team
AFactual
Estimated from token counts × versioned model prices — a fact from the logs, never the model request path.
Spend · 14 days
$9.0k
Period tokens63.0M

Never in the request path

Tokenbase reads logs the tools already write. It never proxies or intercepts live model traffic.

Metadata-only by default

Message bodies are an explicit org unlock — admin opt-in, developer-visible, with a redaction report.

No individual surveillance

Aggregate-only manager views, cohort thresholds, no leaderboards, no drill-down to a person.

Adapter health, surfaced

Parse-success and unknown-record rates per tool version — Tokenbase knows when a vendor update breaks coverage.

Signed, device-scoped collection

Each collector enrolls as a device with its own scoped token and signs every batch — no shared keys, no live browser session.

Coverage stated, not hidden

Per-tool, per-metric coverage labels. Cursor is sold as visibility, not precise spend — honestly tiered.

See where the spend goes.

Two inspectable agents, collected safely, with team-level spend and no source exposure. Start with a metadata-only pilot.