Coding agents on one API

• Long sessions with repeated context, where prompt caching and cache-read prices matter.
• Tool calls and multi-turn history that hold up on every request.
• Room to switch models per task without a new integration.
• Spend that a team can cap per developer, project or pipeline.

• Claude Code runs over /v1/messages, Codex CLI over /v1/responses, and Cline, OpenCode, Continue and Aider over Chat Completions, all with the same key and catalog.
• Any text model in the catalog can run behind Claude Code by setting ANTHROPIC_MODEL; switching models is one id.
• Cache reads are billed at each model's published cache-read price, and cached input is reported in usage.
• Each API key can carry its own daily or monthly spend limit and per-minute request limit, set in the console.
• If a route fails before the response starts, the request moves to another route of the same model.

1. Create a key per developer or pipeline and set its spend limit in the console.
2. Copy the config for your tool from the Coding page.
3. Run the one-line check on the Coding page before a long session.

• POST /v1/messages/count_tokens returns 404; Claude Code falls back to its local estimate.
• Thinking settings are applied, but thinking text is not returned as thinking blocks.
• Images inside tool results are replaced with a short text placeholder.
• /v1/responses is stateless: previous_response_id and store: true are not supported.
• Each organization can send up to 600 requests per minute.

Coding agent setup →Messages reference →Responses reference →All use cases →