2026 年给本地大模型挑外壳：11 个开源选择，按文档扎实程度排Choosing a Harness for Local LLMs in 2026: 11 Open-Source Options, Ranked by How Solid the Docs Are

An agent is made of two parts: a model, plus a harness. The harness calls tools, stores state, manages permissions, and feeds context back to the model. Once you switch to a model running locally, the harness carries more weight — short context windows and weak tool calling mean any corner cut in the design gets magnified.

An agent = a model + a harness. Local models have short context windows and weak tool calling, so any corner cut in harness design gets magnified into a visible defect. This list ranks 11 open-source harnesses by how well local inference is documented, with repository details read from GitHub on September 18, 2026.

Why the Harness Matters More Than the Model

An agent is made of two parts: a model, plus a harness. The harness calls tools, stores state, manages permissions, and feeds context back to the model. Once you switch to a model running locally, the harness carries more weight — short context windows and weak tool calling mean any corner cut in the design gets magnified.

This list ranks 11 open-source harnesses by how well local inference is documented. All repository details were read from GitHub on September 18, 2026. The ranking looks at four things: whether the license is OSI-approved, whether local runtimes are documented, whether maintenance is still active, and how many security controls are available.

Three Rules That Hold for Every Harness

① Turn the context window up first. According to Ollama's context length docs, the default follows VRAM: under 24 GiB of VRAM gets 4k, 24 to 48 GiB gets 32k, and above 48 GiB gets 256k. The same page notes that agent and coding tools should get at least 64,000 tokens. The fix is one line: OLLAMA_CONTEXT_LENGTH=64000 ollama serve.

② Pick a model that supports tool calling. Goose's provider docs say a model without tool-calling ability can only do chat completions. With llama.cpp, Pi's docs mention that the --jinja flag enables a compatible chat template and tool calling.

③ Do the memory math honestly. Cline's local guide lays out a comparison: 16 to 32GB of RAM pairs with small quantized models, 32 to 64GB with mid-size coding models, and above 64GB you can consider something larger. Ollama's Hermes page puts gemma4 at roughly 16GB of VRAM and qwen3.6 at roughly 24GB.

1 | OpenCode

OpenCode's own provider docs lay out three local routes: Ollama, LM Studio, and llama.cpp's llama-server. All three rely on the @ai-sdk/openai-compatible package with a local baseURL. The docs say it supports more than 75 providers in total.

Installing it may take just one command. Ollama's OpenCode page lists ollama launch opencode and suggests a context window of at least 64k tokens. OpenCode's own docs add a practical tip: if tool calls fail, try raising num_ctx to around 16k to 32k.

It ships with two agents built in: build has full permissions; plan is read-only and asks before running bash.

Best for: developers who want the most complete local setup options inside a single terminal tool.

2 | Pi

Pi takes the minimalist route. The README gives the model only four tools: read, write, edit, and run commands. It deliberately omits MCP, subagents, plan mode, and permission prompts — those are left to TypeScript extensions and packages.

Pi natively supports llama.cpp's router server. The router scans multiple GGUF files itself and loads them on demand. Inside Pi you manage models with /llama. Ollama is supported too: ollama launch pi installs Pi, configures the provider and opens a session in one command.

There's one catch for local use: Pi has no built-in permission system and runs with your current user privileges. The README recommends confining it with Docker, micro-VM extensions, or a policy sandbox.

The old badlogic/pi-mono address now redirects to earendil-works/pi. Earendil acquired Pi in April 2026, and author Mario Zechner joined the company. The Pragmatic Engineer reported that OpenClaw grew out of Pi.

Best for: small local models — a short tool list leaves more context for the code.

3 | Goose

Goose documents more local runtimes than anything else on this list. Its provider docs list Ollama, LM Studio, Docker Model Runner, Ramalama and Atomic Chat. vLLM and KServe come in through OpenAI-compatible providers. Custom providers facing a local server can skip the API key.

The real differentiator is governance. The Linux Foundation established the Agentic AI Foundation on December 9, 2025, and Block donated goose to it; the repository now lives under aaif-goose/goose. Goose is written in Rust and ships as a desktop app, a CLI and an API, and the README mentions more than 70 MCP extensions.

Configuring Ollama is short: run goose configure, pick Ollama, and enter a model name.

Best for: general automation where you want neutral foundation governance, not just code writing.

4 | Cline

Cline is the strongest of the editor-native options. Its local guide puts one setting ahead of all its other advice: always turn on Use Compact Prompt for local inference. It also suggests keeping tasks focused and starting a new session once context grows.

By default, every file edit and every command needs your approval; auto-approve is optional. The Plan and Act modes separate thinking from doing.

One licensing detail worth noting: Cline's own README states that the JetBrains plugin is not open source — only the VS Code extension, the CLI and the SDK live in the Apache-2.0 repository.

Best for: VS Code users who want to run local models while keeping a human in the loop.

5 | OpenHands

OpenHands' local guidance is the most specific of the bunch. Its local LLM guide recommends Qwen3.6-35B-A3B as the first local model to try (that advice was updated on May 21, 2026). The hardware requirements are equally blunt: at least 24GB of VRAM for a quantized version, or an Apple Silicon Mac with 64GB of unified memory.

Context sizing is stated just as directly: at least 22,000 tokens of context length, with 32,768 recommended. The guide also warns that Ollama's 4,096 default can't even fit the system prompt.

Linux users hit one trap: LM Studio binds to 127.0.0.1 by default, so OpenHands inside a container can't reach it. Turning on Serve on Local Network fixes it.

Best for: containerized, long-running tasks on a workstation or server with a GPU.

6 | Aider

Aider solves weak tool calling differently: its edit formats have the model emit changes as text. The whole format returns the entire file, and the diff format returns search-and-replace blocks. Each request also carries a repo map of key symbols.

Its Ollama docs call out a real trap: Ollama silently drops context that exceeds the window. Aider's answer is to size the window per request and reserve another 8k tokens for the reply. Note that the page still cites the older 2k default.

What's genuinely worrying here is the maintenance cadence. Version 0.86.2 on PyPI was released on February 12, 2026, and the previous release goes back to August 13, 2025.

Best for: Git-native pair programming where the model's function calling isn't great.

7 | Codex CLI

Codex CLI is Apache-2.0 and ships with two local providers built in. The source hardcodes ollama (port 11434) and lmstudio (port 1234). According to Ollama's Codex page, codex --oss runs gpt-oss:20b by default, and you switch models with -m.

There's one hard requirement: Codex only speaks the Responses API now. The source rejects wire_api = "chat" outright and points to the relevant discussion. So your local server has to expose that interface.

The repository ships separate sandbox crates for Linux and Windows.

Best for: teams standardized on gpt-oss that also want a built-in sandbox.

8 | Qwen Code

Qwen Code's README lists four protocol families — OpenAI, Anthropic, Gemini and Qwen — and names Ollama and vLLM for local models. The project was forked from Google Gemini CLI v0.8.2 and stopped syncing with upstream after v0.1. The npm install requires Node.js 22 or newer.

Best for: open-weight Qwen models paired with a harness tuned by the same lab.

9 | Kilo Code

Kilo's README says Kilo CLI is a fork of OpenCode, and also notes that it forked from Roo in 2025. On April 2, 2026 it shipped a rewritten VS Code extension. Its local model docs cover Ollama, LM Studio and Atomic Chat, and the same page warns that local models often lack prompt caching and computer control.

Best for: people coming from Roo Code who want a path that's still maintained and supports local models.

10 | Hermes Agent

Nous Research's Hermes Agent is a general-purpose agent, not a coding one. The README describes a learning loop that grows skills from experience. Ollama says it ships with more than 70 skills and cross-session memory. Setup is just pointing Hermes at Ollama on localhost port 11434, and context length is detected automatically. Messaging channels include Telegram, Discord, Slack, WhatsApp, Signal and email.

Best for: growing a long-running personal agent on a local model.

11 | OpenClaw

OpenClaw has the most stars on this list. Ollama describes it as a personal assistant that connects messaging services to an AI agent through a central gateway. For local models, Ollama recommends a context window of at least 64k. The first launch shows a safety notice explaining the risks of open tool access. Read that notice carefully — it's wired into your own messaging accounts.

Best for: a messaging-first assistant, provided you're willing to watch the security side yourself.

Key Takeaways

• Among coding harnesses, OpenCode documents the most local routes: Ollama, LM Studio, llama.cpp
• Before blaming the harness or the model, set context to 64,000 tokens
• Pi's four tools suit small models, but you have to bring your own sandbox
• To run Codex CLI locally, the server must expose the Responses API
• Check licenses component by component: Crush is FSL, and Cline's JetBrains plugin is closed source

This is a complete translation of the original article, rewritten sentence by sentence into natural English, with images taken from the original and supplemented by public screenshots from media coverage and project repositories. Original author: Asif Razzaq, published September 18, 2026. Rankings, licenses and version information are translated as stated in the original, with nothing added or removed.