What is an AI agent harness?

An agent harness is the program that wraps a language model in a working loop: it reads context such as a repository or files, decides which tool to call, executes shell commands or edits, reads back results and iterates until the task is done. The model is the weights; the harness is everything around them — permissions, tool schemas, memory, subagents, plugins and the user interface. In 2026 the two layers separated: models swap in a config line, harnesses do not, which is why open, model-agnostic harnesses matter.

What is the best open-source AI agent harness in 2026?

By adoption, OpenCode: about 209,000 GitHub stars on September 22, 2026, MIT-licensed, a mature terminal coding agent that works with any model. By momentum, DeepSeek Harness: MIT-licensed, plugin-first, 140,000+ stars within four days of its August 13, 2026 release, and in worldwide public preview with macOS and Windows desktop apps since October 1, 2026. By scope beyond coding, OpenClaw: the open-source personal-agent gateway that connects models to chat channels, browsers and devices, at v2026.9.x after its 2.0 release on August 31, 2026. By measured efficiency, Unreal Agent: a Go harness that matched Codex's 57.9% on Terminal-Bench 4.0 for 39% less money.

Are Codex CLI and Gemini CLI really open source?

Yes. OpenAI's Codex CLI (about 126,000 GitHub stars on September 22, 2026) and Google's Gemini CLI (about 107,000) are both Apache-2.0 licensed harnesses, so you can read, fork and modify the loop. What is not open is the default model behind them, and each is tuned for its maker's models and billing. Claude Code, by contrast, is a proprietary harness. If you want a harness you can point at any model including local open weights, OpenCode, DeepSeek Harness, OpenClaw, Pi, Cline, Goose and Aider are the model-agnostic choices.

Which harness is cheapest to run?

Harness cost is token overhead, not licence fees — all of these are free software. Unreal Labs' September 22, 2026 benchmarks are the clearest public data: Unreal Agent completed Terminal-Bench 4.0 at 57.9% for $1,428 versus $2,350 for Codex at the same score and $1,827 for Pi at 55.0%, using 28 model turns and 1.73M input tokens per trial versus Pi's 44 turns and 2.83M. Fewer turns and a smaller system prompt mean lower cost at equal pass rate. For any harness, the biggest savings come from prompt-cache hit rate and effort level, not the harness brand.

Quick Answer

Best Open-Source AI Agent Harnesses 2026: Ranked by Job

Published: October 2, 2026

The short answer

For most developers in 2026 the answer is OpenCode: the most-adopted open harness (~209K stars), MIT, model-agnostic, stable. Pick DeepSeek Harness if you want the plugin-first architecture and a desktop app for non-coding work, OpenClaw if the agent needs to live in your chat apps and devices rather than a terminal, Unreal Agent if you benchmark cost per task, and Codex CLI or Gemini CLI if you have already committed to OpenAI or Google models. Rankings reflect adoption, licence, model freedom and measured efficiency as of October 2, 2026.

The ranking

#	Harness	Licence	GitHub stars (Sep 22, 2026)	Model-agnostic	Best for
1	OpenCode	MIT	~209,400	Yes	Default terminal coding agent
2	DeepSeek Harness	MIT	140,000+ within 4 days of Aug 13 launch	Yes	Plugin-first; desktop app for docs, data, research
3	OpenClaw	Open source	—	Yes	Personal agent across chat channels, browser, devices
4	OpenAI Codex CLI	Apache-2.0	~126,000	Tuned for OpenAI	Teams on GPT-6 Sol/Astra
5	Gemini CLI	Apache-2.0	~107,100	Tuned for Google	Free tier, Gemini 3.8 Flash users
6	Unreal Agent	MIT	New (Sep 22, 2026)	Yes	Lowest measured cost per benchmark task
7	Pi	Open source	—	Yes	Minimal, hackable reference harness
8	Cline	Apache-2.0	~69,100	Yes	VS Code-native agent
9	Goose	Apache-2.0	~54,600	Yes	Block’s extensible desktop/CLI agent
10	Aider	Apache-2.0	~49,100	Yes	Git-centric pair programming

Star counts are a snapshot from a September 22, 2026 survey and move daily; treat them as order-of-magnitude adoption signals, not rankings.

1. OpenCode — the default

OpenCode (anomalyco/opencode) is the most-used open harness: ~209,000 stars, 950 contributors and a claimed 16 million monthly developers, MIT-licensed, with a terminal UI and a steady release cadence (v1.18.31 on September 14, 2026). It supports any model provider, including local servers, and has the plugin, MCP and permission surface a team needs without a vendor’s billing attached. Its weakness is that it is a coding harness; it does not try to be a general assistant. If you want one open-source answer to “what replaces Claude Code without lock-in,” this is it. See Unreal Agent vs Codex vs Pi for how harness choice changes cost at equal scores.

2. DeepSeek Harness — plugin-first, now a desktop app

DeepSeek Harness shipped August 13, 2026 as an MIT-licensed developer preview and set a GitHub record — 20,000 stars in about an hour, 140,000+ in four days. Its architecture, built on Cordis’s “everything is a plugin” model, makes the built-in tools (read, write, shell, search) plugins loaded through the same interface third parties use; model backends are plugins too, so it runs Claude, GPT, GLM, Kimi or local weights as readily as DeepSeek.

On October 1, 2026 it moved to worldwide public preview with macOS and Windows desktop apps plus a Node-based web UI, positioned beyond coding: organise files, analyse data, draft documents and slides, research with citations, run background scripts. The default model is DeepSeek-V4.1-Flash (MIT open weights, $0.15/$0.60 per MTok off-peak), which makes the whole stack open. Execution tracing and runtime inspection are built in, and plugins can be authored inside the interface. Caveat: still a preview with breaking changes expected. Background in what is DeepSeek Harness.

3. OpenClaw — the personal-agent gateway

OpenClaw is a different shape of harness: a gateway that connects models to the places you already are — Discord, Telegram, WhatsApp, Slack, iMessage, a browser, paired phones and desktops — with scheduling, memory, skills and multi-agent routing. It began as Peter Steinberger’s project in November 2025 (Steinberger joined OpenAI in February 2026; the project stayed independent and open source), hit OpenClaw 2.0 on August 31, 2026 (v2026.8.1, ~16,000 PRs, rebuilt UI) and offers desktop installers for macOS, Windows and Linux at v2026.9.x, with a gateway-only extended-stable channel released September 24, 2026. Choose it when the agent’s job is to run your life and your ops, not just your repo. Compared against commercial personal agents in best personal AI agents 2026.

4–5. Codex CLI and Gemini CLI — open harness, closed defaults

Both are Apache-2.0 and both are genuinely readable, forkable code. OpenAI’s Codex CLI (~126,000 stars) is the reference harness for GPT-6 Sol and Astra, with the Codex cloud-agent integration; Google’s Gemini CLI (~107,000) ships a generous free tier on Gemini 3.8 Flash. The trade-off is tuning: prompts, tool schemas and billing paths assume the maker’s models. Fork them if you want to study a production harness; pick them as-is if you have already standardised on that vendor. Note that Claude Code is not in this list — it is a proprietary harness, which is why open alternatives tuned for Claude models (OpenCode, Pi) exist.

6. Unreal Agent — efficiency by design

Unreal Labs open-sourced Unreal Agent on September 22, 2026: an MIT-licensed Go library, a runner comparable to codex exec, and a Harbor-compatible benchmark runner. Its design choice is fully asynchronous tool execution — every tool call is logged as in-progress and the model keeps working, so no tokens are spent polling. Published results with GPT-6 Astra at xhigh: Terminal-Bench 4.0 57.9% for $1,428 against Codex’s 57.9% for $2,350 (39% cheaper) and Pi’s 55.0% for $1,827; DeepSWE 1.1 72.4% for $1,367 vs Codex 69.0% for $1,633. No subagents, minimal prompt footprint. It is young and library-first; adopt it for cost-measured pipelines rather than as a daily driver yet.

7. Pi — the minimal reference

Pi is the small, hackable harness that other projects benchmark against — 44 turns and 2.83M input tokens per Terminal-Bench trial in Unreal’s data, more than Unreal Agent but with a mature, widely used codebase. Pick it when you want to read a complete harness in an afternoon and modify it.

8–10. Cline, Goose, Aider

Cline (~69,100 stars, Apache-2.0) is the VS Code-native agent: plan/act modes, MCP, bring-your-own model. Best when the editor is where you live.
Goose (~54,600, Apache-2.0) is Block’s extensible agent with desktop and CLI, extension-based tools and any-model support; strong for non-coding automation inside a company.
Aider (~49,100, Apache-2.0) is the git-centric pair programmer — every change a commit, repo-map context, any model. Older design, still the cleanest for disciplined diffs.

How to choose

Replace a proprietary coding agent with no lock-in: OpenCode.
General desktop assistant for documents, data and research, fully open stack: DeepSeek Harness on DeepSeek-V4.1-Flash.
Agent that lives in chat apps and controls devices: OpenClaw.
Already on OpenAI or Google: Codex CLI or Gemini CLI.
You measure cost per task: Unreal Agent, then compare on your own harness with prompt caching tuned.
Editor-first: Cline. Company automation: Goose. Git discipline: Aider.

Whatever you pick, read how to secure AI coding agent plugins and skills first — the Plugin4Shell zero-click RCE in September 2026 hit open and closed harnesses alike, and the model-agnostic ones put the security burden on you.

Last verified: October 2, 2026. Star counts from a September 22, 2026 survey of GitHub; release dates from project repositories and announcements.