Closed Beta · Limited Spots

Whichdrives your LLM cost, latency, and errors?

Toolken is the LLM gateway that tags every request with the dimension you care about, so you see cost, performance, and reliability in one place. Control comes next.

Join the Beta See how it works
No SDK 5-minute setup Free up to 10k req/mo

Who it's for

Built for your team

One gateway. Every buyer finally seeing the same numbers.

For SaaS teams

Per-feature and per-customer visibility

"Which feature is eating my LLM bill? Which customer is unprofitable? And why did latency spike for enterprise tier last Tuesday?"

  • ✓ Tag any request with arbitrary metadata, one header change
  • ✓ See cost, latency, and error rates grouped by feature or customer
  • ✓ Spot the enterprise customer whose AI assistant is burning margin before it shows up on the invoice
For agent builders

Per-agent cost, latency, and reliability

"My agent loops. I have no idea which agent in the swarm is burning tokens, running slow, or stuck in a retry storm."

  • ✓ Tag each agent run, see cost and p95 latency per agent
  • ✓ Catch high error-rate agents before they cascade
  • ✓ Built for OpenClaw, Hermes, LangGraph, and CrewAI users
FinOps Coming Soon

Per-seat team usage with planned controls

"My 80-person eng team has a shared OpenAI key. I have no idea who is using what, how fast their calls are, or why the bill jumped 40%."

  • ✓ One team key, full breakdown by individual seat
  • ✓ Planned: per-user and per-key budget caps enforced at the edge
  • ✓ Catch the engineer running batch evals on the company account

Three lines. That's the whole integration.

One URL change.

Point your existing LLM client at Toolken. No new SDK. No code refactor. Full attribution and budget control from day one.

import OpenAI from "openai";


const client = new OpenAI({
  apiKey: process.env.OPENAI_API_KEY,
  baseURL: "https://gateway.toolken.ai/v1",
  defaultHeaders: {
    "X-Toolken-Key": process.env.TOOLKEN_KEY,
  },
});


const response = await client.chat.completions.create({
  model: "",
  messages: [{ role: "user", content: prompt }],
});

How it works

From zero to attributing in 5 minutes

One URL change. No SDK. Full visibility across every feature, team, and customer.

01

Point your requests at Toolken

Change your LLM client's base URL to gateway.toolken.ai/v1. Pass X-Toolken-Key and optionally X-Toolken-Metadata-Feature in headers. Zero other changes required.

02

We attribute and forward

Toolken reads your metadata headers, records token counts and cost, then forwards to the provider. Response arrives untouched.

03

See the breakdown in real time

Your dashboard shows spend by feature, tenant, model, and time, updated in seconds. Export via CSV or API. Set alerts on any budget threshold.

Pricing

Simple pricing, zero token markup.

From free prototyping to enterprise scale. Pay only for what Toolken does. LLM costs pass through at cost.

See full pricing Start for free

Questions? [email protected]

Closed Beta · Apply Now

Know every LLM call costs and usagebefore your next invoice.

Join the teams already using Toolken to attribute costs, enforce budgets, and ship AI features with financial confidence.

Join the Beta Talk to us

Already have an account? Sign in

FAQ

Common questions

What is Toolken?
An edge gateway that sits between your code and any LLM provider, giving you visibility across cost, latency, and reliability in one place. Tag a request once and every metric shows up grouped by whatever dimension matters to you: feature, customer, agent, or environment. You bring your own provider keys; we never pool or replace them.
How does it work?
You change one line: point your SDK at gateway.toolken.ai/v1 instead of the provider's URL, and add an X-Toolken-Key header. We forward the request to the provider with your key untouched, read the usage back (tokens, latency, status), and write a row to your dashboard. Same SDK, same response, same model strings.
How long does integration take?
Usually about five minutes. There is no SDK to install and no code paths to fork: a base URL swap and one header. Drop-in samples for TypeScript, Python, cURL, and Ruby live on the homepage.
Will it slow down my requests?
Barely. We run on Cloudflare's edge, so the gateway hop happens close to your user instead of in a single far-off region. The added overhead is negligible next to the provider's own response time.
Which providers do you support?
Thirteen today, available on every plan: OpenAI, Anthropic, Google Gemini, Groq, Mistral, DeepSeek, Together, MiniMax, xAI (Grok), OpenRouter, Cerebras, Fireworks, and Perplexity. Already calling OpenAI or Anthropic? Those work natively: keep your SDK, swap the URL. Everything else you reach through the OpenAI-compatible format and we handle the rest.
How quickly do you support new models?
New versions of providers we already support (a new Claude, a new GPT, a new Gemini) usually work the same day they ship: model strings pass through and our pricing comes from a database we keep current. A brand-new provider integration typically takes one to two weeks.
Can I cap my spend?
Not yet. Spend-cap enforcement is on the roadmap. When it ships, you will be able to set a limit per key or per metadata dimension and Toolken will block the request before it reaches the provider, so a runaway loop cannot quietly drain your budget overnight. Drop your email on the homepage and we will let you know the day it is live.
Is Toolken only a cost tracker?
No. Cost is one signal among three. Toolken records spend (tokens x price), performance (latency per call), and reliability (error rates and status codes) for every request. Today you can slice all three by any metadata dimension you tag. Budget enforcement, smart routing, caching, and Vault-enforced key control are coming next.
What happens if Toolken has a hiccup?
Your traffic is not hostage to our analytics. If our logging layer cannot be reached, the request still goes straight through to the provider; you just lose the log line for that one call, never the call itself. The one thing we always check first is your Toolken key: if we cannot verify it we return an error instead of letting an unauthenticated request through.
Do you store my prompts?
No. We keep token counts, model names, cost, latency, status codes, and the metadata you send us. We never store the prompt or the response text. Optional full request logging (for teams who want audit or replay) is on the roadmap and will always be opt-in.
What is on the roadmap?
A few things we are building next: budget caps enforced at the edge (Vault), smart routing with fallback and load balancing across providers, a response and semantic cache for repeated prompts, a PII anonymizer, MCP gateway support, and optional full transcript logging. None of these are live yet; drop your email and we will tell you as each one ships.
How do I get help?
Email us at [email protected]. That reaches a real person today. A community Slack is on the way; we will share the invite once it is open.

Still have questions? Email us: [email protected]