VERIFIED, NOT IMPLIED
Harness compatibility.
AblitAI publishes integration evidence separately from release support. Release A is text-only; standard tool-bearing requests remain unsupported until parser and exact-GPU validation passes.
Last reviewed 4 September 2026. No coding harness is advertised as production-supported yet. Reference-only configurations are available for local/mock protocol review; do not use paid credits with a pending integration.
Current matrix
| Client | Protocol | Status | Blocking requirement |
|---|---|---|---|
| OpenAI Python 3.8.0 and Node.js 7.10.0 SDKs | Chat Completions | Local SDK pass · GPU validating | Repeat model listing, text, streaming, errors, and accounting on the exact launch GPU. |
| OpenCode | OpenAI-compatible Chat Completions | Reference-only · local/mock evidence | Exact-GPU parser, text, error, accounting, and load tests must pass before standard tool-bearing requests can be supported. |
| Codex | OpenAI Responses API | Reference-only · local/mock evidence | Exact-GPU Responses parser, text, error, accounting, and load tests must pass before standard tool-bearing requests can be supported. |
| Claude Code | Anthropic Messages API | Reference-only · 8K default blocked · unofficial | Default mode exceeds the current context before sending; --bare still sends tools that Release A rejects. Anthropic does not support non-Claude model routing. |
What “compatible” will mean
- The documented reference configuration works from a clean install.
- The client discovers or selects the exact AblitAI model ID.
- Text streams incrementally and terminates normally.
- Sequential and parallel tool-call behavior is a future exact-GPU validation gate, not a Release A capability.
- Authentication, balance, context, concurrency, and service errors are readable.
- Every completed turn settles credits once; an interrupted stream follows the disclosed reservation rule.
Known launch constraints
The current launch configuration is text-only, has an 8,192-token total context and a 4,096-token maximum output, and runs two concurrent sequences on one RTX 3090 Ti. That window is already too small for Claude Code 2.1.246 default startup and may also be too small for large coding repositories, so no coding harness is advertised as supported.
Use today
For launch validation, use cURL or the OpenAI Python/Node.js SDK with POST /v1/chat/completions. Follow the API quickstart. The OpenCode, Codex, and unofficial Claude Code guides are reference-only and copy-ready for local review; standard tool-bearing requests are not release-supported.