Harnesses - Anpord

A harness is the coding agent itself. Every harness runs the same cases, so switching one is a variant change rather than a rewrite.

Harness	`harness`	Credential	Reports
Codex	`codex`	API key, or ChatGPT	Commands, files, tokens
Claude Code	`claude`	API key	Commands, files, tokens
OpenCode	`opencode`	Auth file	Commands, files, tokens
Pi	`pi`	Auth file	Commands, files, tokens
Gemini CLI	`gemini`	API key	Commands, files, tokens
Qwen Code	`qwen`	API key	Commands, files, tokens
Cursor Agent	`cursor`	API key	Commands, files
FX	`fx`	AI Gateway key, or ChatGPT auth file	Verdict only

Each needs a connection before it can run. A harness that reports less still passes or fails on its verifier; it just leaves a thinner trajectory, and no cost.

Models

Model ids are per harness — opus for Claude Code, provider/model for OpenCode, a bare id for Codex. codex and opencode list theirs live, the rest come from a fixed set.

const { models } = await anpord.evals.models({ harness: "claude" });

Compare two

Same case, same sandbox, one field apart:

tasks: [
  { harness: "codex", model: "gpt-5.6-sol", provider: "daytona" },
  { harness: "claude", model: "opus", provider: "daytona" },
]

Each is its own cell with its own baseline, so their histories never mix.

⌘I