Low-cost model coding agent with harness switching and computer-use QA

What they did

They installed Open Interpreter via the curl/irm one-liner and started a session with `i`. Because the tool is a Codex fork, they kept their existing Codex SDK code and only changed the constructor to `new Codex({ codexPathOverride: "interpreter" })`, so their automation pipeline spoke the same exec protocol without a rewrite. They used `/harness` to switch between native, claude-code, qwen-code, deepseek-tui, and minimal harnesses to find which one got the best behavior out of a low-cost model on their repo. They pointed the agent at the repository's `AGENTS.md` and shared `.agents/skills` directory so instructions and skills stayed portable rather than locked into a product-specific format. For verification, they leaned on the bundled QA skill: the agent drove their web app in a real browser through agent-browser, and exercised a native desktop build through trycua computer use, clicking through flows and reporting what broke. They also wired the agent into an ACP-compatible editor by configuring the client to launch `interpreter acp`, and ran `scripts/test-codex-sdk-compat.sh` as a provider-free check that the protocol layer still matched.

How it went

The setup worked end to end: the Codex SDK override was a genuine one-line change, harness switching let a cheap model produce usable patches, and the QA skill caught a broken form flow in the web app plus a mis-sized dialog in the native build. Polish is decent but the agent still needed human review on multi-file refactors, and browser-driven QA runs were slow enough that they were reserved for pre-merge checks rather than every commit.

Worth knowing

Harness emulation is a compatibility shim, not a guarantee that a low-cost model matches frontier behavior; results varied a lot by harness and provider. Computer-use QA needs a real browser and a working trycua setup, which adds install weight and can be flaky on CI. The ACP and SDK paths assume you already have a client or Codex SDK project. Product-specific state still lives under ~/.openinterpreter, and the install script pipes remote content straight into a shell, so pin and review it for sensitive machines.

Wiring

Install with `curl -fsSL https://www.openinterpreter.com/install | sh` (macOS/Linux) or `irm https://www.openinterpreter.com/install.ps1 | iex` (Windows), then start with `i` or `interpreter`. Order of operations: first place `AGENTS.md` at the repo root and skills under `.agents/skills` or `~/.agents/skills`; second select a provider/model in the TUI with `/model`; third pick a harness with `/harness` (native, claude-code, claude-code-bare, zcode, kimi-code, kimi-cli, qwen-code, deepseek-tui, swe-agent, minimal). For editor integration, configure the ACP client to launch `interpreter acp`. For programmatic use, keep the OpenAI Codex SDK and set `codexPathOverride: "interpreter"` so calls route through Open Interpreter's Codex exec protocol; validate with `scripts/test-codex-sdk-compat.sh`. Computer use connects through the bundled QA skill, which drives web apps via agent-browser and native apps via trycua. MCP servers and shared standards (ACP, MCP, AGENTS.md) are the integration seams; config and runtime state live under `~/.openinterpreter`.

Try it yourself

Set up Open Interpreter in this repo as a low-cost coding agent. Read AGENTS.md and any .agents/skills first, then: (1) confirm `interpreter` is installed and run `scripts/test-codex-sdk-compat.sh`; (2) try `/harness` values native, qwen-code, and minimal on a small real task in this repo and tell me which gave the best patch quality per token; (3) use the QA skill to drive our web app in a browser via agent-browser and report any broken flows with repro steps; (4) if a native build exists, exercise it with trycua and list UI defects. Do not commit anything; summarize findings and the exact commands you ran.