start_within. An agent key holds the deadline itself, so every request the tool makes routes by it.
Create the key
- Open the dashboard and go to API, then FlexInference keys
- Press Create agent key
- Pick the deadline
- Turn on Store traces if you want this key’s requests kept in your logs
- Press create, then copy the secret. The dashboard shows it once
- Flex race tries the cheaper tier for a duration you set, from 5 seconds to 10 minutes, then falls back to standard. It costs less and starts later, so a tool you sit and watch usually wants a tier instead.
- Standard serves the normal tier. It works with every model and every route.
- Priority starts sooner and costs more. Worth it when a person waits on the answer.
- Auto lets the provider pick. Not available on Gemini models.
400 invalid_key_defaults.
A key can also carry an upstream retries policy. Leave it unset unless you have a reason, because most tools retry on their own and the two stack.
Store traces puts trace_storage: true on every request through the key, so the request and response show up in your dashboard logs. It only works on managed serving. The toggle is off until you turn it on, and a request that sets trace_storage itself still wins.
The request always wins
We fill a stored field into any request that arrives without it. We never replace a value the request sent, even a wrong one, so an explicit badstart_within still fails.
Every response we filled in carries x-flexinference-defaults-applied. Most tools don’t show response headers, so read the request in the dashboard under Logs to confirm the key applied.
A key duration falls back instead of failing
A duration in a request body fails on a model that can’t race. The same duration on a key falls back to its tier instead, so the key always has an answer. That fallback is also why a key duration never races Claude’s cheaper tier. Send the duration in the request body when you want that race.Edit or revoke
Edit the deadline on the key row. Live traffic picks it up within seconds, and you keep the same key. Revoking a key clears its defaults with it.Per-tool guides
Troubleshooting
These come up in every tool, so they live here once. Each tool’s page keeps only what’s specific to it, and errors carries every refusal we return.400 missing_start_within on every request. The key came from Create key rather than Create agent key, so it holds no deadline. Create an agent key and swap it in. A key made without defaults can’t gain them later.
400 unsupported_parameter. The tool sent a Chat Completions parameter only OpenAI serves, on a model OpenAI doesn’t run. The full list is seed, stop, presence_penalty, frequency_penalty, logit_bias, logprobs, top_logprobs, prediction, audio, modalities, and web_search_options. Switch to a gpt-* model or remove the parameter.
400 key_default_not_applicable. Your key holds a duration and the request sent something a race can’t go with: one of those parameters, or a pinned cloud route. Edit the key to a tier, or remove the conflict from the request.
400 flex_unsupported_for_anthropic. A duration in the tool’s config reached a claude-* model on your own Anthropic key, and that race takes managed keys. Remove the duration from the config and let the key hold the deadline, which falls back instead.
Every request spends a full retry budget. The tool retries on its own, and a key retry policy stacks with it. Leave retry off a key an agent uses.
The cursor sits waiting. Flex race costs less and starts later, which you feel in an interactive loop. Edit the key to Priority or Standard. The next request picks it up with no restart.
Does the tool need to know it holds an agent key? No. It authenticates the same way as any key, and every key starts with flex_live_.
Can one key serve a tool and my own code? Yes. Your code sends start_within and wins. The tool sends nothing and takes the key’s deadline.
What does this cost? Nothing from us on your own provider key, and your provider bills the tokens. On Managed Keys each request draws your balance.