FAQ
Common issues and triage paths for AI app integrations
A single page collecting the common questions from the four integration guides. Jump by section; if something's missing, reach out from the console.
General
401 / 403 unauthorized
Confirm the API key is copied in full and hasn't been revoked in the console; the Authorization header should read Bearer sk-tb-… (or x-api-key: sk-tb-…). Double-check the base URL suffix — OpenAI-compatible needs /v1, Anthropic-compatible does not.
Which base URL?
| Protocol | Base URL |
|---|---|
| OpenAI-compat | https://api.tokenbyte.ai/v1 |
| Anthropic-compat | https://api.tokenbyte.ai |
What do I put in the model field?
Open the Models page or the Console → Models catalog. The title on each model card is the exact value you put in the API model field (e.g. Claude Sonnet 4.5, gpt-5-codex). TokenByte passes the upstream name through as-is — the card title is the model name, no second-level mapping.
Cost and rate control? Configure RPM / TPM and a per-request spending cap on every key individually in the console. Agents that act autonomously (Cline, Claude Code) loop easily; give them dedicated keys with conservative limits.
Usage stats aren't showing up?
Metrics take ~1 minute to settle. If they're still missing, make sure the traffic actually hit api.tokenbyte.ai — some tools fall back to the upstream default endpoint silently.
Claude Code
Model selection? TokenByte forwards Claude Code's model requests to upstream Anthropic as-is. To route to non-Claude models, configure mapping rules in the console.
Rate limited?
Claude Code issues a lot of parallel requests (multiple tools + background summarization). Raise the key's RPM / TPM, and keep CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC set so only essential calls go through.
Which env var?
Claude Code reads ANTHROPIC_AUTH_TOKEN (not ANTHROPIC_API_KEY). TokenByte keys still start with sk-tb-.
Codex CLI
Windows setx didn't take effect?
setx only affects newly spawned terminals. Close and reopen the window — VS Code's integrated terminal too.
Which model should I use?
gpt-5-codex (Codex's dedicated reasoning model). Any TokenByte-supported OpenAI-compatible model ID also works — just set it in the model field of ~/.codex/config.toml.
.codex folder not visible?
Explorer hides dot-directories by default. Enable View → Show → Hidden items, or run mkdir .codex.
What does env_key do?
It tells config.toml which environment variable holds the API key. The value of env_key must match the system variable name exactly — e.g. both set to TOKENBYTE_API_KEY.
Kilo Code
Model list is short? Under OpenAI Compatible, model IDs are typed by hand. Reference the Models page for official IDs.
Homepage doesn't load?
The brand moved from kilocode.ai to kilo.ai. The old domain 308-redirects, but update your bookmarks.
Can Kilo's Anthropic entry take a custom base URL? No — that entry hits Anthropic directly, with no override. Route Claude through TokenByte via OpenAI Compatible instead.
Cline
The Base URL field disappeared? API Provider must be set to OpenAI Compatible — the Base URL field only appears there. Cline's built-in "Cline" quick-login is a managed mode and doesn't allow custom endpoints.
Context overflow? Long tasks hit model context limits. Prefer long-context models (Claude Sonnet 1M, GPT-5.4), and let Cline compact state via Focus Chain or Memory Bank.
How do I cap Auto-approve spend? Set a per-request cap and RPM/TPM on the key itself in the console. During testing, keep Auto-approve off.
Can Plan and Act use different models?
Yes — toggle Use different models for Plan and Act in Settings. Pair a reasoning-heavy model (e.g. claude-opus-4-6) for Plan with a fast coder (e.g. gpt-5-codex) for Act.