If you have only heard of Claude Code, Codex CLI, or Gemini CLI, you are missing a big part of the picture. A handful of smaller, lesser-known coding agent harnesses are quietly matching or beating the big names on benchmarks. The standout is pi, and there is a solid supporting cast behind it: mini-swe-agent, gptme, Crush, Plandex, and Amp, plus some deeper cuts worth knowing about.
Why the harness matters more than the model
A useful definition comes from an independently maintained catalog of over 100 harnesses that gets rescored weekly: a model answers, an agent acts, and the harness is the runtime that turns one into the other. The same catalog found that swapping the harness around a model can move its pass rate more than upgrading the model itself does. One example: the identical model scored 23% versus 52% on SWE-bench Pro depending only on which harness wrapped it.
Independent leaderboards back this up. Claude Opus scores 77% on Terminal-Bench inside Claude Code but 93% inside Cursor, according to a comparison called "The Harness Effect." The official Terminal-Bench leaderboard shows the same underlying models spread from 51.82% for Claude Code with Opus 5 down to 20.30% for xAI's Grok Build, with Codex CLI landing in between at 37.27%. And the actual top of the Terminal-Bench 2.0 board was not a household name at all: a research harness called "vix" running Claude Opus 4.7 hit 90.2%, ahead of Codex CLI's 82.2% with GPT-5.5. Obscure clearly does not mean worse here.
The lesser known but genuinely excellent ones
pi is the minimal, hackable option, and it is the one people keep bringing up. It was built by Mario Zechner, known for creating libGDX and RoboVM, who joined Armin Ronacher's company Earendil in 2026. Ronacher is the creator of Flask and Jinja. The project moved from Zechner's personal badlogic/pi-mono repository to earendil-works/pi in May 2026, with the core staying MIT licensed even as some future enterprise features may shift to a Fair Source or proprietary model.
The whole idea behind pi is "adapt pi to your workflows, not the other way around." It ships with a small core, just the agent loop, file tools, bash, and model providers, and everything else comes through TypeScript extensions, skills, prompt templates, and themes that get shared as packages via npm or git. There is no sub-agent system and no plan mode built in. If you want a feature, you ask pi to build it for you.
Zechner has been vocal about why this matters to him: exactly controlling what goes into the model's context yields better outputs, while most existing harnesses inject things behind your back. He has described Claude Code as a spaceship packed with 80 percent functionality he never uses. And this restraint is not just philosophy for its own sake. With nothing but read, write, edit, and bash tools, no web search, no compaction, and a system prompt under 1,000 tokens, pi took 8th place on Terminal-Bench running Opus 4.5.
A few practical details worth knowing if you are considering it: it routes to Anthropic, OpenAI, Google, and more than 20 other providers using your own API keys. People run local GGUF models like Qwen through llama.cpp inside it. It works on Windows through Git Bash in Windows Terminal. Hooks are stateful TypeScript modules with UI helpers built in. It is also the engine powering OpenClaw. Star counts quoted across different sources range widely, from around 61,000 up to roughly 101,000, depending on when the snapshot was taken. The trade-off is real: you get more control and better token efficiency, but you end up assembling more of your own workflow.
mini-swe-agent is the radically simple, research-grade choice. It comes from the Princeton and Stanford team behind SWE-bench and SWE-agent, and the whole thing is genuinely about 100 lines of code. Despite that, it scores over 74 percent on SWE-bench Verified, is MIT licensed, and has around 6,900 stars. It is best for people who want an agent they can read and audit in an afternoon, and it doubles nicely as a research baseline for anyone studying how these systems work.
gptme is the local-first veteran of the group. It was one of the very first agent CLIs, dating back to spring 2023, and it still runs anywhere a terminal does, whether that is your laptop, an SSH session, tmux, a headless server, or a CI pipeline. It is provider-agnostic, comes with shell, Python, and vision tools plus a plugin system, and supports fully local operation through llama.cpp. Its community is smaller, around 4,400 stars, but it has stayed active and relevant for years, which says something.
Crush is the best-looking terminal agent you will find. Built by Charmbracelet under the tagline "glamorous agentic coding," it has around 27,900 stars. It works with any LLM through OpenAI- or Anthropic-compatible APIs, lets you switch models mid-session without losing context, integrates with LSP, extends through MCP, and runs cleanly across macOS, Linux, Windows, and the BSDs. It is technically Charm's fork of the original OpenCode, released under the FSL-1.1-MIT license.
Plandex is built for big, multi-step tasks. It is a terminal tool designed to plan and execute large coding jobs that span many steps and dozens of files, and it keeps a cumulative diff review sandbox so AI-made changes stay separate from your actual files until you approve them. It has around 15,600 stars.
Amp is the polished, opinionated commercial option. Sourcegraph launched it in May 2025, and it spun out as its own independent company, Amp Inc, in December 2025 under co-founder Quinn Slack. It gives you one agent across web, Mac, iPhone, iPad, and terminal, along with a feature called Orbs, which are per-thread cloud machines that let you send a prompt, close your laptop, and have the agent keep working. The team's stated editorial approach is blunt: if they do not use and love a feature, they kill it. It is well liked among people who do heavy agentic coding work, though it is not open source.
The deeper cuts
A few more names worth knowing if you like exploring the edges of this space. Tau, built by Hugging Face, has around 2,600 stars and is a small, readable Python terminal agent inspired by pi that doubles as a teaching codebase for understanding how harnesses work under the hood, released under MIT. hax has around 700 stars and is a minimalist, terminal-native agent written in C that runs in just a few megabytes of RAM, with strong llama.cpp support. Paseo has around 15,600 stars and is a no-telemetry parallel runner that can drive Claude Code, Codex, Copilot, OpenCode, and pi agents side by side. Orca, from Stably, has around 58,000 stars and runs a fleet of parallel agents, each working inside its own git worktree.
The familiar names, for reference
To put all of this in context, it helps to know where the popular options stand. opencode is the most-starred open-source agent out there, with around 203,000 stars, MIT licensed, and provider-agnostic. Goose started at Block and now lives under the Linux Foundation's Agentic AI Foundation, Apache-2.0 licensed, with around 53,900 stars and support for more than 70 MCP extensions. Aider remains the classic diff-based pair programmer, with around 48,600 stars. OpenHands sits around 85,600 stars, and Cline, which is VS Code native with Plan and Act modes plus MCP support, has around 67,000.
Then there is the first-party trio everyone already knows: Claude Code, which tops the Terminal-Bench 4.0 board at 51.82 percent when paired with Opus 5, Codex CLI, and Gemini CLI. They anchor the benchmark tables and are worth knowing, but at this point they are not really the interesting edge of the design space anymore.
How to actually choose one
- If you want control, transparency, and token efficiency, go with pi, or mini-swe-agent if you would rather read the entire agent in one sitting.
- If local models and privacy matter most, look at gptme or hax for their llama.cpp-first approach, or Goose paired with local providers.
- If you want a batteries-included open source option that works with any provider, opencode is the safe pick, with Crush as the prettier alternative.
- If you are tackling big multi-file tasks and want reviewable diffs, Plandex is built exactly for that.
- If you want something polished and commercial with team threads and background cloud agents, Amp fits the bill.
- If you need to run parallel fleets across many tasks at once, layer Orca or Paseo on top of whichever harness you already use.
A few caveats worth keeping in mind before you commit to any of these. Star counts and benchmark scores are point-in-time snapshots, and this space shifts on a monthly basis. Some Terminal-Bench links now point to the current 4.0 board rather than the historical 2.0 results referenced above. Amp's documentation does not list pricing. And pi's 8th place finish on Terminal-Bench was a single full run reported by its own author, so treat it as a strong signal rather than an independently verified result.
Top comments (0)