Source-backed data on AI coding harnesses.

Compare documented capabilities, operating models, public usage signals, and evidence coverage. Popularity and quality remain separate.

Active catalog entries
51
Latest source check
Sep 8, 2026
Peer-reviewed studies
19
Primary source pages
1081

Public usage signals

Compare source-native activity without turning popularity into a quality score.

Open full usage data
Most used on OpenRouter2026-09-01 to 2026-09-0724 listed; 27 mapped of 51
  1. #1Hermes Agent10.81T tokens
  2. #2Claude Code4.46T tokens
  3. #3Kilo Code2.77T tokens
  4. #4Cline2.58T tokens
  5. #5Pi Agent1.56T tokens

Opt-in traffic attributed to coding apps. Attributed volume does not measure users, task quality, or success. Rankings stay source-specific and never affect capability claims or classifications.

Under the lens: August 2026.

Three contrasting harnesses selected by HarnessMatch from the catalog. Not a ranking, an overall recommendation, or the result of comparative product trials.

Compare these three
  • Integrated product
    Why inspect
    Polished Claude-first agent across terminal, IDE, desktop, and web.
    Check first
    Hosted model access is Claude-only, local open-weight models are unsupported, and Chrome integration requires a direct Anthropic plan
    Catalog checked Aug 22, 2026
  • Extensible harness
    Why inspect
    Extensible LangGraph coding agent with headless runs and optional remote sandboxes.
    Check first
    Local execution is host-first and trusts the working directory; project artifacts can influence the agent before the first approval prompt
    Catalog checked Aug 22, 2026
  • Minimal harness
    Why inspect
    Minimal bash-only coding harness for inspectable CLI use and reproducible research.
    Check first
    The default local environment executes directly with the user's host privileges; isolation is opt-in
    Catalog checked Jul 27, 2026

Browse all harnesses.

Filter by a capability, then open a profile for trade-offs and sources.

51 active profiles

Advanced filtersLayer, role, surface, and runtime
Coding harness65 first-party sources

Polished Claude-first agent across terminal, IDE, desktop, and web.

Role
Coding agent
Interfaces
Terminal, IDE, Web / desktop, Automation
Model access
Enterprise routing
Checked 2026-08-22View profile
Coding harness49 first-party sources

Agentic coding across app, terminal, IDE, cloud, and automation.

Role
Coding agent
Interfaces
Terminal, IDE, Web / desktop, Automation
Model access
Multi-provider
Checked 2026-08-22View profile
Coding harness28 first-party sources

Model-agnostic coding agent with a broad, configurable surface.

Role
Coding agent
Interfaces
Terminal, IDE, Web / desktop, Automation
Model access
Multi-provider
Checked 2026-08-22View profile
Coding harness20 first-party sources

Minimal terminal harness designed to be extended, embedded, or automated.

Role
Extensible harness
Interfaces
Terminal, Automation
Model access
Multi-provider
Checked 2026-08-10View profile
Coding harness24 first-party sources

A batteries-included Pi fork with LSP, debugger, browser, and subagents.

Role
Coding agent
Interfaces
Terminal, IDE, Automation
Model access
Multi-provider
Checked 2026-08-22View profile
Coding harness26 first-party sources

Extensible Rust coding agent with plan review, sandbox profiles, and ACP.

Role
Coding agent
Interfaces
Terminal, IDE, Automation
Model access
Multi-provider
Checked 2026-08-07View profile
Coding harness17 first-party sources

Meta's sandbox-first terminal agent with goals, memory, and multi-agent workflows.

Role
Coding agent
Interfaces
Terminal, Automation
Model access
Single vendor
Checked 2026-08-05View profile
Coding harness18 first-party sources

Focused pair programming built around Git, diffs, and explicit control.

Role
Pair programmer
Interfaces
Terminal, IDE, Web / desktop, Automation
Model access
Multi-provider
Checked 2026-07-27View profile

Capability filters reflect explicit product documentation. They do not compare model intelligence or benchmark performance.

Read the methodology

Latest stable releases

Current version, publication date, and 90-day cadence from reviewed product-scoped GitHub release feeds. Frequency describes activity, not quality.

Open all release feeds
  1. GGCodeFull-source repository
    Stable versionv1.3.236Open release
    ReleasedSep 8, 2026
    Release cadence160 releasespast 90 days
  2. KernFull-source repository
    Stable versionv0.34.0Open release
    ReleasedSep 8, 2026
    Release cadence9 releasespast 90 days
  3. Oh My PiFull-source repository
    Stable versionv18.1.14Open release
    ReleasedSep 7, 2026
    Release cadence163 releasespast 90 days
  4. Letta HarnessClient-source repository
    Stable versionv0.31.13Open release
    ReleasedSep 7, 2026
    Release cadence102 releasespast 90 days
  5. Junie CLISupport repository
    Stable version3110.6Open release
    ReleasedSep 7, 2026
    Release cadence32 releasespast 90 days
  6. Hermes AgentFull-source repository
    Stable versionv2026.9.7Open release
    ReleasedSep 7, 2026
    Release cadence13 releasespast 90 days
  7. ReasonixFull-source repository
    Stable versionv1.38.1Open release
    ReleasedSep 6, 2026
    Release cadence83 releasespast 90 days
  8. Claude CodeSupport repository
    Stable versionv2.1.263Open release
    ReleasedSep 6, 2026
    Release cadence73 releasespast 90 days

Showing 8 of 35 reviewed release feeds, sorted by latest stable release date and then trailing 90-day cadence. Version tags are shown verbatim. Observed Sep 8, 2026.