Updated
Vibe coding

AI Coding Agents

Hand a coding task to an agent that reads your repo, edits files and runs commands — in your terminal, or in a cloud machine that comes back with a pull request.

Ask your AI about this, with this page as the source:ChatGPT ↗Claude ↗Perplexity ↗

Trends

In Sep 2026, OpenAI Codex commented on the most public pull requests (898.6k). GitHub Copilot fell to 0.3×. Of pull requests opened, GitHub Copilot opened the most (60.2k); Claude Code grew fastest, 7.1× since Apr.

Public counts each month: pull requests on GitHub, package downloads, stars, Product Hunt followers. They show direction, not market share.

Public pull requests each tool commented on · log scale, by month the PR was created, Apr 2026 to Sep 2026
10k20k50k100k200k500k1MAprMayJunJulAugSepGitHub Copilot: 19,927 in Sep 2026Devin: 70,363 in Sep 2026Claude Code: 109,065 in Sep 2026Jules: 115,258 in Sep 2026OpenAI Codex: 898,558 in Sep 2026OpenAI Codex 898.6kJules 115.3kClaude Code 109.1kDevin 70.4kGitHub Copilot 19.9k
Public repositories only, from GitHub search (counted Oct 21). Recent months keep growing as older pull requests get reviews.
Public pull requests each tool opened · log scale, by month the PR was created, Apr 2026 to Sep 2026
1k10k100kAprMayJunJulAugSepJules: 10,045 in Sep 2026Claude Code: 19,775 in Sep 2026Devin: 39,781 in Sep 2026GitHub Copilot: 60,196 in Sep 2026GitHub Copilot 60.2kDevin 39.8kClaude Code 19.8kJules 10k
Public repositories only, from GitHub search (counted Oct 21). Recent months keep growing as older pull requests get reviews.
ToolGitHub PRs, Sep 2026DownloadsStarsProduct Hunt
OpenAI Codex898.6k reviewed 3.4× since Apr—★ 129.4k1.7k followers
Jules115.3k reviewed 0.6× since Apr10k opened 1.7× since Apr8.5k/wk——
Claude Code109.1k reviewed 2.0× since Apr19.8k opened 7.1× since Apr12.9M/wk−13% vs npm—11.9k followers
Devin70.4k reviewed 1.1× since Apr39.8k opened 1.4× since Apr———
GitHub Copilot19.9k reviewed 0.3× since Apr60.2k opened 0.2× since Apr——1.2k followers
Factory867 reviewed10 opened9.8k/wk——
OpenHands178 reviewed3.9k/wk★ 91.2k0 reviews
Pi—2.8M/wk★ 116.1k254 followers
Mistral Vibe CLI—1.1M/wk★ 5.1k—
Gemini CLI—374.5k/wk−64% vs npm★ 107.4k—
Oh My Pi—151.4k/wk★ 35.7k—
Qwen Code—65k/wk★ 28.5k—
Kimi CodeRenamed—30.9k/wk★ 7.9k—
Amp—19.2k/wk——
Crush—6.8k/wk★ 28.6k—
OpenCode——★ 216.5k1.5k followers
goose——★ 55.4k3 followers
Grok Build——★ 27.4k—
Lately in this groupEverything that changed →

The real choice

Where the agent runs, and whose models it is tied to. A terminal agent works in your checkout on your machine, so you watch it, approve commands and stop it mid-task; a cloud agent takes the task to its own VM and returns a branch or pull request, so you can run several at once but review only the result. Most agents from a model lab use that lab's models and are paid through its subscription; the open-source, model-agnostic agents work with any provider or a local model, and the bill is whatever your API keys cost. AI-first editors and editor plugins are on their own page, AI Code Editors.

What makers chose886 live products, each with a source. Not market share.

Pick by situation

Tap the ones that are you — the tools that fit light up below.
  • You're comfortable in a terminal and want to hand off whole tasksClaude CodeOpenAI Codex
  • You already pay for ChatGPTOpenAI Codex
  • You want to switch between model providers freely or keep your tools open sourceOpenCodePigoose
  • You want to try a terminal agent without paying for a subscriptionGemini CLI
  • Tasks start as GitHub issues and you want pull requests back without running anything locallyGitHub CopilotJules
  • Tasks arrive in Slack, Linear or a ticket tracker and you want an agent to pick them up thereDevinFactory
  • You want agents running on a machine you control, with any model including local onesOpenHands

Compare the tools

Grouped by the side of the choice they answer, not ranked. Open a row for when to use it, the trade-off and what makers say.

ToolBest for
Terminal agent6
Claude CodeappHanding off multi-step tasks — features, refactors, bug hunts — to an agent that reads the repo, edits files and runs commands, with web sessions for running tasks in the cloud.772+740 open source$20Pro12.9M/wk−13% vs npm

Use it whenYou're comfortable in a terminal and want the agent to do the work while you review.

Trade-offUses Anthropic models only, and needs a Claude subscription or API credits.

llms.txt

Used by Accordio AI, agent-manager, Agentcard, AgentCenter for OpenClaw, AgentConnect and 767 more · in 740 open-source projects.

We build HarnessRouter with Claude Code, and HarnessRouter runs Claude Code as a product backend for other teams. It is both our daily driver and one of the harnesses we route to.
HarnessRouter, the makerSep 2026 ↗
OpenAI Codexapp · open source · free tierDelegating tasks to an agent in the terminal, IDE or desktop app, or to cloud tasks that run in the background and open pull requests.145$8.00ChatGPT Go★ 129.4k

Use it whenYou already pay for ChatGPT, or want an open-source CLI backed by OpenAI models.

Trade-offBuilt around OpenAI models, and usage limits depend on your ChatGPT plan.

llms.txt

Used by agent-manager, Almanac, Chert, Chronock, Click and 140 more.

Codex is one of the three harnesses HarnessRouter routes to, and we use it in our own build loop. Its structured output made it one of the cleanest harnesses to normalize against.
HarnessRouter, the makerSep 2026 ↗
OpenCodeapp · open source · free tierAn open-source coding agent that works with almost any model provider, local models, or an existing Copilot or ChatGPT subscription.44+22 open source$10Go★ 216.5k

Use it whenYou want to switch models freely or keep your agent tooling open source.

Trade-offResult quality depends on the model you connect, and you manage provider keys and costs yourself.

CLI

Used by agent-manager, Brev.io, CoChat, Cosyra, Cushion and 39 more · in 22 open-source projects.

A strong open local coding-agent runtime. We like it because it keeps agent work close to the developer environment while still fitting into a shared workspace through Vokal.
Vokal, the makerSep 2026 ↗
Piapp · open source · free tierA deliberately minimal open-source (MIT) terminal agent that works with 15+ model providers through API keys or existing subscriptions, extended with packages instead of built-in features.4+48 open source—2.8M/wk

Use it whenYou want a small agent you understand completely and extend yourself, on whichever model you prefer.

Trade-offNo sub-agents or plan mode out of the box; you add what you need, and results depend on the model you connect.

Used by bb, MausBot, Side, SonOf · in 48 open-source projects.

Pi is one of the agent engines integrated into MausBot. Its integration gives users another agent option within the same local app and conversation workflow.
MausBot, the makerOct 2026 ↗
Gemini CLIapp · open source · free tierAn open-source terminal agent on Gemini models with a free tier, for reading the repo, editing files and running commands.42 open source—374.5k/wk−64% vs npm

Use it whenYou want to try a terminal agent without a paid subscription, or already use Google's AI plans.

Trade-offBuilt around Gemini models, and the free tier has daily request limits.

llms.txtCLI

No maker's product we track shows it for this yet · in 42 open-source projects.

gooseapp · open source · free tierAn open-source (Apache-2.0) agent as a desktop app, CLI and API, for coding and other workflows, with 15+ providers or an existing Claude, ChatGPT or Gemini subscription, and 70+ MCP extensions.2 open source—★ 55.4k

Use it whenYou want a model-agnostic agent with a desktop app, governed by a foundation rather than one company.

Trade-offA general-purpose agent rather than a coding specialist; it moved from Block to the Agentic AI Foundation, so older docs and links point to block.github.io.

No maker's product we track shows it for this yet · in 2 open-source projects.

Cloud agent1
GitHub Copilotapp · free tierAssigning a GitHub issue to Copilot's cloud agent and getting a pull request back — also from Slack, Jira or Linear — or running the Copilot CLI in a terminal, on the Copilot plan you may already have for your editor.27+173 open source$10Pro

Use it whenYour code and issues live on GitHub and you already pay for Copilot.

Trade-offThe cloud agent is built around GitHub repositories and issues, and agent sessions draw on your plan's AI credits.

llms.txt

Used by Beatoven.ai, Bolt, Edgee, Open Paper, Packmind Open Source and 22 more · in 173 open-source projects.

Our token compression technology has to be compatible and efficient with Github Copilot, so we have to use it extensively to understand how we can optimize its context.
Edgee, the makerSep 2026 ↗

Cost: the cheapest plan that fits developers 1 seat, from list prices. Try your own numbers →

Cost as you grow

Each contender's cheapest usable plan as usage rises.

Who switches to what

Public pull requests on GitHub since Oct 2024 whose title says "X to Y" — real code changes, by developers in general rather than makers only. Pick a flow to see its pull requests.

Claude Code → OpenCode: 44 pull requestsOpenAI Codex → Claude Code: 43+ pull requestsGitHub Copilot → Claude Code: 40 pull requestsClaude Code → OpenAI Codex: 34+ pull requestsClaude Code → GitHub Copilot: 28 pull requestsOpenCode → Claude Code: 28 pull requestsOpenAI Codex → GitHub Copilot: 22 pull requestsGitHub Copilot → OpenAI Codex: 14 pull requestsClaude Code 106OpenAI Codex 65GitHub Copilot 54OpenCode 28Claude Code 111GitHub Copilot 50OpenAI Codex 48OpenCode 44Moving fromMoving to
Claude Code → OpenCode44 PRs · Compare →

What Hacker News says

Hacker News commenters choose Claude Code most often for AI coding agents, and like OpenCode best (net praise +32%). Among products launched since 2024 that show their choice, 85% use Claude Code.

ToolHN comments choosing itNet praise on HNMakers' launches since 2024
Claude Code2,642−4%85% 686
OpenAI Codex905+9%16% 132
GitHub Copilot845−0%3% 23
OpenCode725+32%5% 38
Pi246+19%0% 2
Aider185+44%0% 1
Gemini CLI77−14%0% 0
Oh My Pi42+39%0% 0

Comments January 2023 to August 2026 that name the tool in a sentence like "I use" or "we moved to", estimated from a sample Claude read; net praise is comments recommending it minus comments criticizing it, as a share of all comments about it. Makers: of 811 launches that show what they use here. HN commenters and makers are different crowds — where they agree and where they split.

Before you choose

How to approach it

Keep a short instructions file in the repo (AGENTS.md or the agent's own file: stack, conventions, commands) so every agent starts with the same context. Start with one terminal agent on small, well-scoped tasks, have it run the tests or the app, and read every diff before you commit. Move a task type to a cloud agent once you trust the result enough to review only the pull request. Try one agent properly for a week before switching; most of the gain comes from learning how to prompt and review.

Common mistakes
  • Accepting large diffs without reading them, so bugs and security holes you never saw end up in your codebase.
  • Letting an agent call a task done without running the tests or the app, so plausible-looking code that doesn't work lands in main.
  • Giving an agent production credentials or unrestricted shell access on your main machine, where one wrong command can do real damage.

Other options

Real choices most makers here won't need to weigh.

AiderTerminal agentMaintenance only · Feb 2026The original open-source terminal pair programmer, which edits files in your git repo with your own model keys or local models and commits every change.1 makers' productAmpTerminal agentA coding agent in the terminal and editors, with runs in the cloud for longer tasks.DevinCloud agentHanding a whole task to an agent that works in its own cloud environment and comes back with a pull request.JulesCloud agentAsynchronous tasks on a GitHub repo, run in a Google cloud VM, with a REST API and CLI to start them.FactoryCloud agentCoding agents you can start from the terminal, the web, Slack or Linear, through to a pull request.OpenHandsCloud agentAn open-source (MIT) control center for running coding agents — its own, or Claude Code, Codex and Gemini CLI — on your laptop, in Docker, on a server or in OpenHands Cloud, including automations triggered by schedules or GitHub events.Kimi CodeTerminal agentRenamedMoonshot's coding agent — CLI, desktop app and VS Code extension — on its Kimi K3 model, included in Kimi membership.Mistral Vibe CLITerminal agentMistral's minimal open-source (Apache-2.0) terminal agent on its Devstral coding models.Grok BuildTerminal agentxAI's open-source terminal agent on Grok models, which also runs headless and over ACP inside editors.CrushTerminal agentCharm's terminal agent with many model providers, language-server context and MCP support.Qwen CodeTerminal agentThe Qwen team's open-source (Apache-2.0) terminal agent, with VS Code, Zed and JetBrains integrations, built for Qwen models and usable with other providers.Oh My PiTerminal agentA fork of Pi with more built in — language-server and debugger tools, 60+ providers — for people who like Pi's approach but want fewer packages to assemble.CodegenCloud agentAcquiredA platform for orchestrating coding agents in the cloud for teams — now part of ClickUp.

Decided alongside

What the 886 makers' products here chose for their other decisions.

Further reading