Early access: your personal sandbox is free, with $5 in model credits included. AQ adds no markup on your model usage. Start free

aq.dev / guides / coding-agent-harness-directory

The Coding Agent Harness Directory (August 2026)

This is a directory of coding agent harnesses and the tools that orchestrate them, checked against each vendor's own site in August 2026. It covers four shapes: CLIs built by model labs, open-source model-neutral harnesses, desktop apps that run several harnesses in parallel, and cloud platforms that run agents on hosted machines. It also lists what shut down or was replaced this year, the part every static awesome-list gets wrong: two well-known entries died in February 2026 alone. For the definition of the term, read what is a coding harness; for a decision framework, read how to choose one. This page is the map.

How to read this directory

Four facets separate these tools better than feature lists:

CLIs from the model labs

Each lab now ships its own harness, tuned for its own models. All of these run locally in your terminal by default, so persistence is your problem (see keeping sessions alive past a closed laptop).

HarnessBuilt byModel accessNotes (as of August 2026)
Claude CodeAnthropicClaude Pro/Max plan, Team/Enterprise premium seat, or API keyTerminal, IDE extensions, desktop app, and the web at claude.ai/code; macOS, Linux, Windows
Codex CLIOpenAIChatGPT plan sign-in or API keyOpen-source terminal agent; one account also covers Codex in ChatGPT, the editor, and the cloud
Antigravity CLIGoogleGoogle account plansThe agy command; replaced Gemini CLI, which Google shut down for free, Pro, and Ultra users on June 18, 2026; multi-agent by default, with skills, hooks, and MCP support
Grok BuildxAISuperGrok or X Premium Plus subscriptionAnnounced May 25, 2026 as an early beta; open-source TUI in Rust, with plan mode, headless CI use, and editor embedding over ACP
Kimi Code CLIMoonshot AIKimi models by default; other compatible providers configurableMIT-licensed TypeScript agent with built-in coder, explore, and plan subagents; succeeds the earlier kimi-cli
AmpAmp (ampcode.com)Amp accountRuns on the web, in the terminal, and on a phone; threads run in orbs, per-thread machines billed by the minute, and workspace members can join a thread while it runs; see AQ vs Amp

Open-source, model-neutral harnesses

These decouple the harness from the model vendor: bring an API key, a subscription, or a locally served open-weight model.

HarnessShapeNotes (as of August 2026)
OpenCodeTerminal, IDE, or desktop appConnects to any provider with your own keys, including local models; optional OpenCode Go subscription ($5 first month, then $10 per month) for hosted open-source models
PiMinimal terminal harnessShips four tools and a system prompt under 1,000 tokens; everything else arrives as extensions, skills, and packages shared over npm or git
AiderTerminal pair programmerAmong the oldest harnesses still in wide use; connects to most LLMs including local ones and git-commits every change
GooseLocal agent frameworkStarted at Block, now under the Agentic AI Foundation; model-configurable, extends through MCP
OpenHandsCLI, browser client, and agent serverMIT-licensed and model-agnostic; runs locally, self-hosted, or on its managed cloud

Desktop apps that run harnesses in parallel

These are orchestrators rather than harnesses: they run the stock CLIs side by side on your machine, usually one git worktree per task. This layer, the harness of harnesses, is where churn is fastest.

ToolPlatformsNotes (as of August 2026)
ConductormacOSRuns Claude Code, Codex, Cursor, and OpenCode in parallel, each workspace in its own worktree
EmdashmacOS, Windows, LinuxFree and open source (YC W26); one worktree per task, works with 25+ CLI agents
SupersetDesktopPitches running 100+ parallel coding agents on your machine
ProliferateDesktop, self-hostableOpen-source IDE for Claude Code, Codex, and OpenCode; runs agents in parallel, locally or in the cloud
XirpWeb dashboard over local sessionsSpotify's vendor-neutral agentic development environment, opened as a free public beta on August 10, 2026; manages Claude Code, Codex, and Gemini CLI sessions, one working tree per session
Vibe KanbanDesktop, open sourceKanban board over parallel agents with automatic worktrees; sunsetting as a company, continuing as community-maintained open source

Cloud platforms that run agents on hosted machines

Here execution moves off your machine: the vendor provisions a sandbox, runs the agent, and hands back a pull request.

PlatformVendorNotes (as of August 2026)
Claude Code on the webAnthropicClaude Code sessions in Anthropic-managed sandboxes, at claude.ai/code
Codex cloudOpenAIRuns many tasks in parallel, each in its own cloud sandbox preloaded with your repository
Cursor cloud agentsCursorAgents in isolated VMs; hand a CLI conversation to the cloud by prefixing a message with an ampersand, resume on web or mobile
Copilot coding agentGitHubAssign an issue to Copilot; it boots a GitHub Actions VM, writes the code, and opens a PR
DevinCognitionAutonomous engineer with a sandboxed VM per session; runs many tasks in parallel
Warp OzWarpLaunched February 2026; cloud agents in Docker sandboxes, driven by Claude Code, Codex, or Warp's own harness
SuperconductorSuperconductorCloud workspace running Claude Code, Codex, Pi, Amp, and others in parallel, with previews and one-click PRs from phone or desktop

Gone this year: the churn a directory has to track

This market retires tools faster than lists update. Status checked on each vendor's own site in August 2026:

One adjacent entry worth knowing: qm, the MIT-licensed multiplayer agent harness Y Combinator open-sourced in July 2026, gives every employee a work agent over Slack and the web, driven by Pi, OpenCode, Codex, or Claude Code at its core. It is a company-wide work harness rather than a coding harness, and a sign of where the vocabulary is heading.

The facet the directory keeps surfacing: who sees the work

Scan the tables and a pattern shows: nearly everything local is single-player, and nearly everything multiplayer runs on the vendor's cloud. The default matters: a July 2026 LeadDev analysis of 25,264 agent-generated pull requests across 2,361 popular GitHub repositories found that in 79 percent of agentic PRs the same developer both reviewed and modified the agent's contribution, and only about one in eight workflows involved multiple humans. When a session is invisible, the only reviewable artifact is the diff, and the person closest to the diff is the person who prompted it. The missing quadrant, multiplayer sessions on infrastructure you own, is what multiplayer coding agents names.

Where AQ fits

AQ is the multiplayer coding harness where engineering teams run AI coding agents like Claude Code and Codex together: shared live terminals, a code editor, and app previews, in your own cloud. In this directory's facets, AQ sits in that missing quadrant. It is not another harness: agents run as the real CLIs (Claude Code, Codex, Cursor Agent, Kimi, Grok, or plain shells) in persistent tmux sessions on your team's VM, streamed live to the browser. Sessions survive a closed laptop and resume from any device, each workspace gets one isolated git worktree, and teammates open the same workspace and watch the same live session (typing into someone else's terminal happens only after its owner approves a control request in one click; visibility is owner-managed, team-visible or private and shared with specific people).

Model access stays as the tables describe: each engineer signs into the CLIs with their own Claude or OpenAI account, and AQ never marks up model usage. Agents commit, push, and open PRs with per-user GitHub auth, every workspace can serve a live dev-server preview with links that are viewable without an account, and Linear issues become workspaces by labeling them ai-task, with a human starting the agent, never an auto-launch.

Pricing is two plans: Free is a personal sandbox for one person (AQ creates a private machine in an isolated network, nothing to install, no time limit), and Team is $50 per user per month in early access (standard $200, billed monthly), covering VMs you connect from your own cloud or a dedicated always-on AQ-managed VM in its own isolated network, with your rate locked for your first 12 months.

Plainly: pick any harness from the tables above, or several. AQ is the layer that makes whichever ones your team picks persistent, isolated, and visible to more than one person, without moving execution into someone else's cloud.

Frequently asked questions

What is a coding agent harness?

The software wrapped around a language model that turns it into a working coding agent: the system prompt, tool definitions, context management, execution loop, and failure handling. Claude Code, Codex CLI, OpenCode, and Pi are all harnesses; the model underneath supplies the reasoning. The glossary page covers the term in depth, including why the same model scores differently in different harnesses.

Which coding agent harnesses are open source?

As of August 2026: Codex CLI, OpenCode, Pi, Aider, Goose, OpenHands, Kimi Code CLI, and Grok Build among the harnesses themselves, plus Emdash, Proliferate, and Vibe Kanban among the tools that orchestrate them. Claude Code, Amp, and the cloud platforms (Devin, Warp Oz, Superconductor) are proprietary. Open source here usually means the harness code; the frontier models it drives remain hosted APIs unless you serve open-weight models yourself.

Which coding agent tools shut down or were replaced in 2026?

Terragon shut down on February 9, 2026, and Crystal was deprecated on February 26, 2026 (its maker now ships Nimbalyst). Google shut down Gemini CLI for free, Pro, and Ultra users on June 18, 2026, replacing it with the Antigravity CLI, and Moonshot AI succeeded kimi-cli with Kimi Code CLI. Vibe Kanban is sunsetting as a company and continuing as community-maintained open source. Check the vendor's own site before adopting anything in this category.

Can I run multiple harnesses against the same repository?

Yes, and it is increasingly the norm rather than the exception. The safe pattern is one git worktree per session, so parallel agents never edit the same checkout: desktop orchestrators like Conductor and Emdash automate exactly that on one machine, and AQ does it on a team VM, where each workspace is an isolated worktree and any CLI can run inside it. Keep shared instruction files (CLAUDE.md, AGENTS.md) consistent across the harnesses you use.

Which coding harness should I choose?

Start from your model access: an Anthropic subscription points at Claude Code, a ChatGPT plan at Codex CLI, own keys or local models at OpenCode or Pi. Then run a small bake-off on your own repository rather than trusting benchmarks, because scores are earned by a model and harness pairing, not a model. The chooser guide walks the full decision, and treats the harness as a per-engineer pick you revisit as models change.