tokevenHow Tokeven works
Track every instrumented AI call across Claude, ChatGPT, Gemini and Grok. Tokeven receives usage metadata only. Your provider keys and prompt content stay on your machine.
The data flow
From API call to dashboard - four hops, no prompt content on the wire.
Your machine
The wrapper (or the sidecar) reads response.usage after each call: token counts, model, latency, modality, and a local character count. Prompt content never reaches Tokeven.
Batched push
Events flush to /ingest asynchronously (default: every 5s or 50 events). Authenticated by write-only ingest token.
Tokeven backend
Server recomputes cost_usd (cache-aware, authoritative). Idempotent on (org_id, event_id) - duplicates are silently dropped.
Spend, Save, Proof
Three views query usage_events and usage_summaries. Sub-2-second load. Export to CSV/JSON on the Team plan.
Install the wrapper
Tokeven ships as a lightweight Python package. wrap() takes your existing Anthropic, OpenAI, Google or xAI client and returns an instrumented one - it dispatches on the client you hand it, so the same line works for all four. Grok has no SDK of its own: xAI serves the OpenAI protocol at api.x.ai, so it runs through the OpenAI SDK. No code changes to your prompts, no new dependencies in your hot path.
After each API call, the wrapper reads response.usage and captures: input/output tokens, cache read/write tokens, model name, latency, and a local character count of the prompt. It never sends the prompt content itself to Tokeven.
Cannot wrap the client - a different language, a tool you do not control? Run the self-hosted tokeven-proxy sidecar in your own environment and point the provider client at it instead. Those two paths - wrap() and the sidecar - are the only things that record usage. The MCP server is the pre-send advisor: it prices calls, it does not track them. Traffic that never passes through an instrumented path (the Workbench, a console, an un-instrumented app) will not appear in your dashboard. tokeven status tells you which paths are live on a machine.
What gets captured
- Token counts (input, output, cache_read, cache_creation)
- Provider and model name (e.g. claude-sonnet-4-6)
- Modality, detected from block types - never from content
- Latency in milliseconds
- Local prompt_length (character count, not content)
- session_id and project tags (optional)
- tool_used (detected from tool_use blocks)
What the wrapper never sends to Tokeven
- Prompt and completion content
- Your Anthropic (Claude), OpenAI (ChatGPT), Google (Gemini) or xAI (Grok) API key
- Personal identifiers, unless you turn on the optional actor label - which sends whatever you set it to, such as an email address
Captures usage: token counts, model, latency and cost. Prompt content and your provider key never reach Tokeven.
The same 9,000-token system prompt is billed as fresh input on 755 Claude calls a week. Marking it cacheable bills the repeats at the 90%-off cache-read rate instead.
155 of your 370 claude-opus-4-8 calls this week were single-turn lookups. On the identical prompt, claude-sonnet-4-6 prices 40% lower.
3 gemini-3.1-pro workflows re-send full conversation history every turn. Summarising prior turns would cut fresh input tokens by ~35%.
See the price and the alternatives before you send
Before you send, the advisor prices the call and scores cheaper alternatives across all four providers: what each would cost, and a confidence score for this kind of task. It runs entirely on your machine - as the MCP tool tokeven_estimate, as client.estimate(), or from the CLI. Then you choose. Tokeven never swaps a model on you and never downgrades a call silently; a suggestion you do not accept changes nothing about the request you make.
Prices behind those estimates are verified daily. Confidence scores come from a dated third-party benchmark snapshot, not a live feed: the advisor prints the benchmark week next to its scores and flags the snapshot when it is stale.
After the fact, a weekly digest email gives your account admins the week's spend by model category and provider, with coaching tips drawn from calls your team already made (Team plan).
Efficiency quick wins (shipped with the wrapper)
- Model-tier nudges: 'This task used Opus but Haiku handles it at 80% less'
- Cache hints: 'Your 4,200-token system prompt repeats 8x/day - add cache_control'
- Context trimming: 'This workflow re-sends full history - summarize prior turns to cut 60%'
- Prompt-length flags: 'Reference files instead of pasting - 12K chars is expensive'
- Auto-compaction: Summarize long conversations to cut input token growth
Prove every dollar saved
Three views, in the order the questions actually arrive: where is the money going, what should we change, and did the change work. Each loads in under 2 seconds.
Spend
Where your budget goes, by model, team, and session. Tabs for Overview, Cost Drivers, Models, Tokens, and Activity.
Save
What to change next: recommendations, caching opportunities, and prompt efficiency - each with a dollar figure.
Proof
Every verified dollar saved, traced to the model switch you accepted. Open book - this is the same math we bill from.
Around those three sit the pages you go to on purpose rather than daily: Team for per-person and per-team rollup, Weekly Analysis for the digest, Board Briefs for the monthly summary you forward, plus Export and API & Widgets.
The single biggest lever most teams have is caching: Save shows the cacheable patterns in your usage and what caching them would be worth in dollars - alongside per-call cost, model breakdowns, and team spend across all four providers on Spend.
What changes with Tokeven
See more comparisons across every model pair at /compare.
Everything above ships today
There is no phase list to wait out. Tokeven is one release - capture, advisor, dashboard, widgets, teardown tool, MCP server and integrations all landed together. Here is the whole surface area.
Capture
Runs in your environment, pushes metrics only.- Drop-in Python wrappers for the Anthropic, OpenAI and Google SDKs
- xAI/Grok via the OpenAI SDK against api.x.ai - same one-line wrap
- Self-hosted sidecar (tokeven-proxy) for anything you would rather not wrap
- Write-only ingest tokens, SHA-256 hashed and revocable - never a provider key
- Cache-aware cost, recomputed server-side on ingest so it stays authoritative
- Idempotent on your own event_id - retries and duplicates cost you nothing
Advise before you send
Entirely client-side. You always pick the model.- Pre-send estimate: what this call costs, plus cheaper cross-provider options
- A confidence score per alternative, per task type - not a blind price sort
- Three ways in: the MCP tool tokeven_estimate, client.estimate(), or the CLI
- Efficiency quick wins: auto-compaction, model-tier picker, cache and max_tokens hints
- Prices are verified daily; task rankings are a dated snapshot the advisor prints
- Nothing is routed for you and nothing is downgraded behind your back
Analyze and prove
Three views, then the artifacts you forward upward.- Spend, Save and Proof - the three dashboard views, each under 2 seconds
- Per-person and per-team cost rollup on the Team view
- Weekly Analysis digest built from your own usage (Team plan)
- Monthly Board Briefs for the budget conversation you have to win
- Proof traces every saved dollar to the recommendation you accepted
- Savings are per-call deltas priced at call time - no frozen baseline
Integrate
Get the numbers where your team already looks.- REST API into your existing BI tooling
- Embeddable widgets and a live cost widget for your own surfaces
- n8n, Zapier and outbound webhooks
- Budget thresholds and spend alerts
- CSV and JSON export (Team plan)
- A free, no-login teardown tool for a first look before you sign up
What Tokeven will not do
Not gaps - commitments.- Route your models silently. The advisor shows the options and the cost; the choice stays yours.
- Hold a provider key. The credentials Tokeven stores are password hashes, hashed tokens, and the signing secrets for webhooks you create - Privacy Policy §2.3 lists them.
- Receive your prompts on the ingest path. Two opt-in tools (prompt revision and advisor.improve()) send text you paste, and both say so at the point of use.
- Bill per seat. The fee is 10% of verified savings, so zero savings means zero fee.
Ready to see where your AI budget goes?
Free tier: one developer, the full dashboard, forever. Install in two minutes. Paid plans bill 10% of verified savings. Zero savings means zero fee.
Start free