Harness

Harness mode is an Open Interpreter addition. It changes the model-facing prompt, tool schema, message conversion, and response handling while keeping the native Open Interpreter runtime.

The public runtime does not shell out to the real external agent CLI.

Configure

harness = "kimi-code"
harness_guidance = true

For one run:

interpreter -c harness='"kimi-code"' "solve this task"

Harness IDs

Reference Projects

These links are the external products or repositories whose user-facing surfaces informed the corresponding harness modes:

• Claude Code from Anthropic.
• ZCode, a GLM coding-agent surface used with Anthropic Messages-compatible endpoints such as Z.AI's Coding Plan endpoint.
• DeepSeek TUI, with product context at CodeWhale.
• Current Kimi Code and its MoonshotAI/kimi-code repository.
• The legacy Python Kimi CLI and its MoonshotAI/kimi-cli repository.
• Qwen Code CLI and its QwenLM/qwen-code repository.
• SWE-agent and its SWE-agent/SWE-agent repository.

Route Compatibility

Harness routing is strict:

This means an Anthropic-style provider normally needs:

model_provider = "anthropic"
harness = "claude-code"

Most OpenAI-compatible hosted providers use wire_api = "chat" and can either run through generic chat compatibility or through a matching chat harness.

Automatic Harness Defaults

When the config does not set harness, Open Interpreter may infer one from the selected provider/model:

Explicit harness = "..." always wins.

What Each Harness Changes

claude-code

Builds Claude Code-shaped requests over the selected provider's Responses, Chat, or Anthropic Messages transport. It adds a Claude Code system prompt, thinking configuration, context-management settings, title-generation requests, and a Claude-shaped tool surface. Transport-specific headers and message conversion are applied when the provider requires them.

It maps supported tools into Anthropic tool definitions and includes handlers for Bash/PowerShell, Read, Write, Edit, TodoWrite, Glob, Grep, web search/fetch, LSP, scheduled wakeups, and Claude-style subagents.

claude-code-bare

Uses the same provider transport as claude-code, but with the bare Claude Code profile. The bare profile uses a smaller prompt/profile shape and different output config defaults. DeepSeek selects this profile automatically.

zcode

Uses Anthropic Messages-compatible requests with a ZCode-shaped system prompt, headers, tools, todo and plan behavior, skills, session context, and subagent surface. Tool calls still run inside Open Interpreter's native Rust runtime.

The harness requires wire_api = "messages". The built-in Z.AI and Z.AI Coding Plan providers use wire_api = "chat", so use the Messages endpoint configuration in Z.AI, GLM, and ZCode when you want the actual ZCode route.

kimi-code

Uses Chat Completions-compatible requests with the current Kimi Code prompt, tool definitions, prompt-cache key, thinking configuration, and message format. Supported tool calls are executed by Open Interpreter's native Rust runtime; it does not invoke the external Kimi executable.

This is the default for Kimi and Moonshot providers, including Kimi K3. Kimi Code subscriptions use the kimi-for-coding provider and can authenticate through the built-in Kimi sign-in flow. Moonshot Platform API keys use the moonshotai provider.

kimi-cli

Uses Chat Completions-compatible requests with the legacy Python Kimi CLI-shaped system prompt. It includes working-directory listing, AGENTS.md loading, Kimi skill discovery, prompt-cache keys, reasoning-effort mapping, and Kimi tool schemas.

Kimi tool handlers include Shell, ReadFile, WriteFile, StrReplaceFile, Glob, Grep, ReadMediaFile, SearchWeb, FetchUrl, SetTodoList, plan-mode controls, background task list/output/stop, AskUserQuestion, and Agent.

Keep this mode only when you specifically need compatibility with the retired Python CLI profile. New Kimi sessions should use kimi-code.

deepseek-tui

Uses Chat Completions-compatible requests with a DeepSeek TUI/CodeWhale-shaped system prompt. It adds turn metadata, repository context, generated project instructions when none are found, and DeepSeek TUI tool schemas.

DeepSeek TUI tool handlers include shell, apply patch, edit/write/read file, list directory, grep/file search, git status/diff, diagnostics, checklist, plan, and tool search.

qwen-code

Uses Chat Completions-compatible requests with Qwen Code startup context. It adds a synthetic setup exchange containing date, OS, current directory, and a small folder listing before the user conversation.

Qwen handlers include read file, write file, edit, shell command, glob, grep, todo write, ask user question, plan exit, and agent.

swe-agent

Uses a SWE-agent-style discussion/command loop rather than tool schemas. The assistant response is parsed for a shell command, Open Interpreter injects the corresponding action, and command output returns as an observation. The default command timeout is 30 seconds.

minimal

Uses a compact software-agent system prompt and maps function tools into a plain Chat Completions tool list. It is useful when a provider supports chat tools but does not need a provider-specific harness surface.

Guidance

harness_guidance = true is enabled by default. Today it only adds extra guidance for kimi-cli; other harnesses ignore it.

Disable it for a stricter harness-shaped run:

harness_guidance = false

Related Config

model_provider = "moonshotai"
model = "kimi-k3"
harness = "kimi-code"

[model_providers.moonshotai]
name = "Moonshot AI"
base_url = "https://api.moonshot.ai/v1"
env_key = "MOONSHOT_API_KEY"
wire_api = "chat"

For provider and wire-api details, see Providers.