TokenRouter
A managed API gateway tool that supports creating individual keys for each engineer and configuring hard daily spending limits. It performs budget validation before forwarding requests, focusing on fine-grained access control and cost management for LLM API traffic.
Timeline (last 90 days)
TokenRouter's Dataset entries are snapshots, and retention cleanup and wipe operations do not break evaluations; evaluation runs support exact_match, contains, regex, or llm_rubric.
In TokenRouter, Monitors are asynchronous, sampling-run LLM judges that run a few seconds after the response completes, do not occupy the request path, and add no latency.
TokenRouter's Monitors cap the input of each judgment at about 4k tokens.
In TokenRouter, the judge must be a model that supports tool calling, because the verdict is returned via a tool call, and the model selector only lists such models.
In TokenRouter, each monitor has a daily judge budget, and when it is used up the monitor automatically pauses and automatically resumes at UTC midnight; leaving it blank means no limit.
Judge calls go through the organization's own provider key, with zero markup, are rate-limited through the gateway, are logged with source=monitor, and are counted toward the organization's default team.
The official documentation gives a rough figure: with 100,000 requests per month and 10% sampling, a mini-level judge costs roughly 2 to 10 USD per month of the organization's own provider spending.
Monitors take tiered follow-up actions on findings: Flag is always on and enters the review queue; Notify must be enabled and sends at most one email per cooldown window; Disable key must be enabled and has a severity threshold, and automatic disabling should be regarded as a circuit breaker in severe scenarios rather than a content filter.
TokenRouter does not store prompt and response bodies by default, and request logs record only token counts, cost, and latency; captured bodies are compressed and encrypted before leaving the API process, and capture never fails, slows down, or changes a request.
TokenRouter's captured body retention period is limited by plan: Starter 7 days, Team 30 days, Enterprise 90 days, and a shorter window can also be chosen, and administrators can wipe all data with one click.
10 more timeline events
All Facts (20)
TokenRouter's Dataset entries are snapshots, and retention cleanup and wipe operations do not break evaluations; evaluation runs support exact_match, contains, regex, or llm_rubric.
50%UnverifiedIn TokenRouter, Monitors are asynchronous, sampling-run LLM judges that run a few seconds after the response completes, do not occupy the request path, and add no latency.
50%UnverifiedTokenRouter's Monitors cap the input of each judgment at about 4k tokens.
50%UnverifiedIn TokenRouter, the judge must be a model that supports tool calling, because the verdict is returned via a tool call, and the model selector only lists such models.
50%UnverifiedIn TokenRouter, each monitor has a daily judge budget, and when it is used up the monitor automatically pauses and automatically resumes at UTC midnight; leaving it blank means no limit.
50%UnverifiedJudge calls go through the organization's own provider key, with zero markup, are rate-limited through the gateway, are logged with source=monitor, and are counted toward the organization's default team.
50%UnverifiedThe official documentation gives a rough figure: with 100,000 requests per month and 10% sampling, a mini-level judge costs roughly 2 to 10 USD per month of the organization's own provider spending.
50%UnverifiedMonitors take tiered follow-up actions on findings: Flag is always on and enters the review queue; Notify must be enabled and sends at most one email per cooldown window; Disable key must be enabled and has a severity threshold, and automatic disabling should be regarded as a circuit breaker in severe scenarios rather than a content filter.
50%UnverifiedTokenRouter does not store prompt and response bodies by default, and request logs record only token counts, cost, and latency; captured bodies are compressed and encrypted before leaving the API process, and capture never fails, slows down, or changes a request.
50%UnverifiedTokenRouter's captured body retention period is limited by plan: Starter 7 days, Team 30 days, Enterprise 90 days, and a shorter window can also be chosen, and administrators can wipe all data with one click.
50%UnverifiedIn TokenRouter, Rules evaluate expressions in-process on every request or response, supporting interception, masking, transformation, routing, or alerting, and provide an observe mode for safe canary rollout.
50%UnverifiedTokenRouter improves decoding throughput by 2× to 64× through token-level routing.
50%UnverifiedThe number of seats in TokenRouter depends on the plan: the free tier offers 1 seat and 1 team, Starter offers 5 seats, and Team offers 25 seats.
50%UnverifiedWhen Hard cap is enabled in TokenRouter, requests that exceed the limit are rejected with a 429 response until the period resets.
50%UnverifiedTokenRouter's budget cycle uses UTC calendar day boundaries, with daily budgets resetting at 00:00 UTC.
50%UnverifiedWhen configuring Claude Code to point to TokenRouter, you must not append /v1 after the base URL, otherwise it becomes /v1/v1/messages, causing a 404.
50%UnverifiedKeys created by members are automatically bound to that member so the member budget takes effect, while keys created by administrators are subject only to the team budget.
50%UnverifiedTokenRouter's quota check is completed before the request is forwarded, reserving the worst-case cost; if the limit is exceeded, the request is rejected with a 429 budget_exceeded and never reaches Anthropic.
50%UnverifiedTokenRouter's vendor keys are AES-encrypted at rest, decrypted in memory only at request time, and are never written to logs.
50%UnverifiedTokenRouter is a managed gateway positioned to support per-person budgets, charging a fixed subscription fee rather than taking a cut per token.
50%