AI coding harnesses, compared
A harness — also called a CLI coding agent — is the tool that sits between you and the model: it reads your files, plans, edits, and runs commands. The model gets the headlines, but the harness decides how many tokens your task actually costs, whether you can choose your own model, and how much of your machine it occupies.
This site tracks the field with measured numbers: token spend on a fixed task suite, pricing model, license, language, and footprint. Updated monthly. Last updated: .
The index
Sorted by GitHub stars (a proxy for mindshare, not quality). BYOK = bring your own key / any provider. Token cost = measured on our fixed benchmark suite where available; see the blog for methodology.
| Harness | Stars | Language | License | BYOK | Notes |
|---|---|---|---|---|---|
| OpenCode | 210K | TypeScript | Open source | Yes | The category leader; 75+ providers, LSP. Development perceived as closed-off |
| Claw Code | 195K | Rust | Open source | Yes | Clean-room Claude Code rewrite from the March 2026 source leak |
| Codex CLI | 126K | Rust | Apache-2.0 | No | OpenAI's agent; OpenAI models only |
| Gemini CLI | 107K | TypeScript | Open source | No | Google's agent; Gemini models only |
| Pi | 109K | TypeScript | Open source | Yes | Minimal harness by badlogic; measured 4.6M tokens on our suite |
| OpenHands | 89K | TypeScript | Open source | Yes | Agentic dev environment, CLI + web |
| 3code | 86 | Nim | MIT | Yes | Token-economy focus, tiny binary, no telemetry. Measured 5.1M tokens, highest task score on our suite. Affiliation disclosure: this site is run by the 3code maintainer — check our raw data. |
| Cline CLI | 69K | TypeScript | Open source | Yes | VS Code hybrid going CLI; large community |
| Goose | 55K | Rust | Open source | Yes | Block's agent; extensible, enterprise-backed |
| Aider | 49K | Python | Apache-2.0 | Yes | The veteran; git-native pair programming. Last push May 2026 |
| Codewhale | 41K | Rust | Open source | Yes | Multi-model role-based agents; hosted or local models |
| Claude Code | 148K | — | Proprietary | No | Anthropic's agent; subscription or API billing, Claude models only |
Head-to-head comparisons
Cost
The number nobody publishes: what a month of real coding agent work costs per harness, on the same models. See the blog for the running cost series.