What is an AI coding harness?

The client that turns a language model into an agent. It runs the loop, exposes tools like file editing and the terminal, manages context and permissions, and decides when the task is done. The model supplies the reasoning; the harness supplies everything else.

Does the harness change how much I pay?

Substantially. The harness controls how much context it re-sends, whether it caches, and how many turns it takes. Two harnesses running the same model on the same task can differ several-fold in output tokens, and output tokens are most of the bill.

Can I use one harness with several model providers?

Yes, if you choose a model-agnostic one. OpenCode and Command Code both let you bring your own keys across providers. Claude Code is locked to Anthropic models; ZCode and Antigravity are built around their vendor's models.

Is Gemini CLI still usable?

Not on consumer plans. It stopped serving requests on 18 June 2026 for Google AI Pro and Ultra subscribers and free Gemini Code Assist users, and was replaced by Antigravity CLI. Access under Code Assist Standard or Enterprise licences was unaffected.

Coding harnesses

The coding agent is not the model

You pick a model for capability and a harness for how the work gets done. These are the harnesses worth knowing in 2026, with every claim sourced and dated — including the one that has already been retired.

Compare themAll rows verified Aug 14, 2026

Read this first

A harness is the client that turns a model into an agent: it holds the loop, the tools, the file access and the permissions. Two people using the same model through different harnesses get very different results — and pay very different amounts, because the harness decides how many tokens the job costs.

One of these is already gone

Google retired Gemini CLI into Antigravity CLI, and consumer access stopped serving requests on 18 June 2026. We keep the row because knowing what replaced it is more useful than a page that quietly drops it.

Labs benchmark through harnesses now

When Z.ai published GLM-5.3's scores, it ran them inside Claude Code — a competitor's harness. That tells you the harness layer has become standard infrastructure, and that a model's published score is partly a statement about the client it was measured in.

Compare the harnesses

Filter by how each one handles models, then select a row for its full detail, limitations and sources. Retired harnesses are hidden by default.

Show retired

Harness	Models	Licence	Runs on	Pricing
can1357 (community)	Any modelAny provider via BYOK — Anthropic, OpenAI, Google, xAI, DeepSeek, Mistral, plus local servers (Ollama, llama.cpp, vLLM)	Open source	Terminal, Zed, Node.js SDK, One-shot CLI	Free, MIT-licensed. You pay only your own provider API bills; no subscription and no vendor account required.
Anthropic	Locked to one vendorAnthropic models only (Claude Fable, Opus, Sonnet, Haiku)	Proprietary	Terminal, VS Code, JetBrains, Desktop, Web, GitHub, Slack	Included with Claude paid plans; also usage-billed via the Anthropic API
SST	Any modelAny provider via BYOK — Anthropic, OpenAI, Google, DeepSeek, GLM, local models	Open source	Terminal, IDE extensions, Web	Free and open-source with your own API keys; OpenCode Go plan $5 first month, then $10/month
Z.ai	Own model, not lockedGLM family (GLM-5.3, GLM-5.2), including a self-hosted GLM endpoint	Proprietary	Desktop (macOS, Windows, Linux), Mobile steering via WeChat / Feishu	Via the GLM Coding Plan's points quota — off-peak calls cost 50% of standard points
Langbase	Any modelAnthropic, OpenAI, Google, DeepSeek, Qwen, Kimi, GLM, MiniMax and more — BYOK with no markup	Proprietary	Terminal	From $1/month with $10 credits, up to $200/month Ultra; Teams $40/month; BYOK carries no markup
DeepSeek	Own model, not lockedBuilt around DeepSeek models (V4.1 Flash on the deepseek-flash API id since 10 Sep 2026); plugin architecture allows others	Open source	Terminal, Web UI	MIT-licensed and free; you pay only for model tokens
Google	Own model, not lockedGemini models	Proprietary	Desktop, CLI, SDK	Via Google AI Pro / Ultra and Gemini Code Assist licences

Oh My Pi (omp)

can1357 (community)

Hashline edits cut output tokens: The model patches by content-hash anchor instead of retyping lines, and a stale anchor is rejected before it corrupts the file. The project reports Grok 4 Fast spending 61% fewer output tokens on the same work (self-reported) Source
Real IDE machinery, not a text box: 31 built-in tools, 14 LSP operations and 28 DAP (debugger) operations, so renames go through the language server and it can drive lldb, dlv or debugpy Source
Adopts your existing config on first run: Inherits rules, skills and MCP servers already on disk from .claude, .cursor, .codex, .gemini, .cline, .windsurf and .github/copilot — no migration script Source

Extension points

SupportedSkills
SupportedHooks
SupportedSubagents
SupportedPlugins
SupportedMCP

The honest limitation

Community-run and moving very fast — three releases shipped on 1 Sep 2026 alone — with no vendor support contract, no SLA and no company behind it. It is also the most complex harness here: ten model roles and 31 tools are a lot of surface to learn, and the headline efficiency numbers are the project's own, not independently verified.

Best for

Technical teams who want IDE-grade tooling and a hard floor on spend, and who can absorb a fast-moving open-source dependency without vendor support.

Source Verified 2026-09-02

Questions people actually ask

What is an AI coding harness?: The client that turns a language model into an agent. It runs the loop, exposes tools like file editing and the terminal, manages context and permissions, and decides when the task is done. The model supplies the reasoning; the harness supplies everything else.
Does the harness change how much I pay?: Substantially. The harness controls how much context it re-sends, whether it caches, and how many turns it takes. Two harnesses running the same model on the same task can differ several-fold in output tokens, and output tokens are most of the bill.
Can I use one harness with several model providers?: Yes, if you choose a model-agnostic one. OpenCode and Command Code both let you bring your own keys across providers. Claude Code is locked to Anthropic models; ZCode and Antigravity are built around their vendor's models.
Is Gemini CLI still usable?: Not on consumer plans. It stopped serving requests on 18 June 2026 for Google AI Pro and Ultra subscribers and free Gemini Code Assist users, and was replaced by Antigravity CLI. Access under Code Assist Standard or Enterprise licences was unaffected.

A harness choice is downstream of a model choice.

Pick the model that clears your quality bar at the lowest cost, then pick the harness that runs it the way your team works. Doing it the other way round is how people end up locked into a vendor they did not mean to choose.

What shipped recently

Dated releases from every major lab, with each vendor's own claims labelled as claims.

Open

Independent scores

Third-party benchmark results — deliberately separate from whatever a lab publishes about itself.

Open

What to actually build

A deployment plan picks the stack for one workflow at your budget, models and tooling included.

Open

We record a source and a check date for every claim on this page, and we show disagreements instead of resolving them silently. How we verify