Give every engineer a coding agent. Cap what each one can spend.
TokenRouter is the gateway between Claude Code, Codex, your apps and the model providers. Each engineer gets a key with a hard budget, each request is logged with its cost, and a judge model reviews the answers.
One key per engineer. One bill for the company.
Claude Code, Codex and opencode connect with a base URL and a key. Nothing about the tool changes. Every session shows up in the console under the engineer who ran it, their team and their budget.
Issue a key per engineer
Each key carries its own rate limit and budget. When someone leaves, revoke their key. Your provider keys never leave the gateway.
export ANTHROPIC_BASE_URL=https://api.tokenrouter.ioexport ANTHROPIC_AUTH_TOKEN=tr_your_key_here claudeConnect with two environment variables
Point the agent at the gateway and keep working. Requests to the same provider pass through untouched, so extended thinking, tool use and prompt caching behave exactly as before.
Budgets that block, not just warn
Agents retry, re-read and re-plan. When a budget is spent, the next request is refused before it reaches the provider. No overage to argue about at the end of the month.
Every request logged, no prompts stored
Each request is recorded with its model, tokens, cost and latency. Prompts and completions are stored only where you switch on capture, encrypted and with a retention limit.
Choose which models each team can use
Allow a whole provider or a single model per team. Anything outside the list gets a 403 and never reaches a provider.
Catch the bad answer
before your customer does.
A judge model reads a sample of your live traffic and asks one question of each answer: does this break the rules you wrote? A made-up refund policy, a leaked phone number, a force-push from a coding agent. Findings arrive within minutes, with the judge’s reasoning attached.
01
Write the rules in plain English
Choose a scope, a sample rate and a judge model, then describe what a bad answer looks like. No regex, no labelled data. Judging runs seconds after each response and never in its path.
Monitors in the docs
02
Findings arrive with reasoning
Every finding carries a severity, a category and the judge's explanation. Email the owners above a severity you choose, or let a critical finding disable the key on the spot.
How monitors work
03
Prove the fix with an eval
Add the failure to a dataset. Run it against a new prompt version or a cheaper model, graded by the same rubric on the same real conversations. Switch when the numbers hold.
Datasets and evals
Thirteen providers behind the SDK you already use.
Keep your OpenAI or Anthropic client. Change the base URL and the key, and every model in the catalog is one string away. A routing rule can move a team to a cheaper model without a deploy.
from openai import OpenAI client = OpenAI( base_url="https://api.tokenrouter.io/v1", api_key="tr_your_key_here",) response = client.chat.completions.create( model="openai/gpt-5-mini", messages=[{"role": "user", "content": "Hello, TokenRouter!"}],)print(response.choices[0].message.content)Your code stays the same
Requests to the provider you named are forwarded byte for byte. Requests across providers are translated at the gateway, so Codex can run on Claude or DeepSeek without a plugin.
Route by rule, not by redeploy
Match on team, model, token count or content, then block, reroute, redact or alert. Rules run before the provider is called, so a blocked request costs nothing.
Built for production traffic
Your provider keys, your rates, your data. TokenRouter adds control without adding markup or latency.
Read the docs- Your rates, not ours
- Providers bill your keys at their rates. TokenRouter is a flat monthly subscription, whatever your volume.
- Nothing stored by default
- Request logs hold tokens, cost and latency. Prompts and completions are kept only where you switch on capture, encrypted, retention-capped and deletable.
- Verbatim passthrough
- Anthropic clients hit /v1/messages, OpenAI clients hit /v1/responses, and the body is forwarded byte for byte. Thinking blocks, tools and cache_control arrive intact.
- Never in the request path
- Budget checks and rules add one Redis round trip. Judging runs after the response, off the hot path, and can never slow one down.
Flat pricing. Zero token markup.
Start free with no card. Paid plans carry a 14-day trial. Tokens are always billed by your providers, at their rates.
Free
$0/ month
Try the gateway, no credit card required
- 1 seat · 1 team · 2 providers
- 1,000 requests / month
- Budgets, hard caps, alerts
- Usage analytics and logs
Starter
$29/ month
For solo builders getting AI spend under control
- 5 seats · all 13 providers
- 1M requests / month
- model="auto" routing
- 1 judge monitor · 7-day capture
Team
Popular$99/ month
For teams managing AI spend across projects
- 25 seats · unlimited teams
- 10M requests / month
- Per-team and per-member budgets
- 5 judge monitors · datasets and evals
Enterprise$499+ / month
Unlimited seats, requests and monitors · audit logs · custom retention · self-hosted option · SSO coming soon
Start with one engineer.
Add the team when you’re ready.
The free plan needs no card. Point Claude Code at the gateway, set a budget, and watch the first session land in the console.