Hayz Labs · https://hayzlabs.com/reviews/top/coding-agents/
reviews · top 20 list · checked 19 Oct
Top 20 AI coding agents
The tools that put a model to work on your code: they read the project, make changes and run commands.
Best: pi (90, free). 9 of 20 scored.
Security notes for tools on this list
Scored ones first, highest score first; on a tie, the better independent result. Unscored ones follow, by their public signal. Open the Table to see the number that placed each one and how it was scored.
- 1piEarendil
Small terminal coding agent you can reshape with extensions; best harness score in one independent study.
Best for Developers who want a small, free agent they can shape around their own modelYou give up No default model; you bring and pay for your own key or subscription
90ChecklistFree - 2OpenCodeAnomaly
Free open-source coding agent for 75+ model providers; scored 54 with GLM-5.3. Skip it for one flat monthly plan.
Best for Developers who want a free agent and the choice of 75+ model providersYou give up Scores below Codex and Claude Code; several free Zen models may train on prompts
80ChecklistFree - 3OpenHandsOpenHands (All Hands AI)
Open coding agent you can run locally or in their cloud; scored near the top in an independent harness study
Best for Teams who want to self-host an open agent, or try it free in the cloudYou give up Free cloud tier allows 10 conversations a day; the app has had no release since July
80ChecklistFree - 4Claude CodeAnthropic
Highest published score on a coding-agent test, from $20/month; OpenAI's Codex scores close for far less per task.
Best for Developers who already pay for Claude and want the strongest current coding modelYou give up No free plan; costs far more per task than Codex, which scores close
70Checklist$20/mo - 5Devin (Devin CLI with Fusion)Cognition
Cloud and terminal agent that pairs a top model with a cheaper helper model to cut cost
Best for Teams that hand tickets from Slack or Linear to an agent that returns pull requestsYou give up Usage quotas are not published, so it is hard to know how far $20 goes
70ChecklistFree - 6Muse CodeMeta
Cheapest way into a near-top coding agent; the best model needs the $50 plan
Best for Coders on a tight budget who still want a strong agentYou give up Only 10 to 50 prompts per five hours; a contributor tier trains on your data
70ChecklistFrom $5/mo - 7Google Antigravity CLIGoogle
Google's free terminal coding agent, replacing Gemini CLI; turn off data use in settings if code privacy matters.
Best for Anyone who wants to try a terminal coding agent for free, or who used Gemini CLIYou give up Pricing page doesn't name models or limits; some Gemini CLI features are missing
Not scored yetFree - 8OpenAI Codex CLIOpenAI
Close to the top coding-agent score at a fraction of the cost per task; needs ChatGPT Plus or an API key.
Best for ChatGPT Plus subscribers who want a terminal coding agent at no extra costYou give up Default model unnamed; Plus limits vary widely and personal plans train by default
70Checklist$20/mo - 9Grok BuildxAI
Low-cost terminal agent on Grok 4.7, well behind the leaders on current tests
Best for Trying a terminal agent cheaply at $8 a month, or free with limitsYou give up Solves about 38% of terminal tasks against about 65% for Claude Code
60ChecklistFree - 10Kimi Code CLIMoonshot AI
Open-source terminal agent for Kimi members; behind the leaders and opting out of training takes an email
Best for Kimi Plus members who want an open-source agent on Kimi K3You give up Needs a paid Kimi plan, and the usage quota is not published
50ChecklistNo public price - 11JunieJetBrains
JetBrains' agent for its IDEs and the terminal; a natural pick if you already use IntelliJ or PyCharm
Best for JetBrains IDE users, free if they bring their own model keysYou give up Not scored yet: its last independent result is older than 90 days
Not scored yetFree - 12CursorCursor (Anysphere, part of SpaceX)
AI code editor with a cheap in-house model; default model unstated and trains on code unless Privacy Mode is on.
Best for People who want an AI code editor and are happy to choose the model themselvesYou give up Trains on your code unless Privacy Mode is on; default model not named
Not scored yet$8/mo - 13Factory DroidFactory
Agent for terminal, IDE and cloud machines aimed at teams; no current independent result
Best for Teams willing to pay for an agent that also runs on cloud machinesYou give up No free plan listed, from $20 a month; not scored yet
Not scored yetFrom $20/mo - 14GitHub Copilot (including Copilot CLI and cloud agent)GitHub
Cheapest paid coding agent at $10/month, CLI on the free plan; no independent test, and personal plans train by default.
Best for Anyone wanting a paid coding agent for the least money, or a free one to tryYou give up Free users can't pick models, and Claude and Codex agents need Pro+ at $39
Not scored yet$10/mo - 15ClineCline Bot Inc.
Free open-source coding agent; you pick the model and pay the provider. Quality depends on the model you choose.
Best for VS Code users who already pay for a model and want a free open agentYou give up Quality depends on your model; a recent security flaw means updating to 3.0.30
Not scored yetFree - 16Kilo CodeKilo Code
Open agent for VS Code, JetBrains and CLI at provider prices; owned by Anaconda since July 2026
Best for VS Code and JetBrains users who want an open agent at model makers' own ratesYou give up 5% fee on credit purchases; update past a security flaw fixed in 7.4.1
Not scored yetFree - 17OpenAI Codex CloudOpenAI
Codex tasks that keep running in OpenAI's cloud, on Plus and up; no independent test of the cloud side yet.
Best for ChatGPT Plus or Business users who want coding tasks to run unattendedYou give up Not on Free, Go or an API key alone; cloud tasks use more of your plan allowance
Not scored yet$20/mo - 18Google AntigravityGoogle
Free coding agent app on Gemini Flash; Claude 5.5 needs AI Pro at $19.99/month. Change the data setting first.
Best for Trying a free coding agent app with several agents running at onceYou give up Free plan runs on weaker Gemini Flash; top Claude models need AI Pro at $19.99
Not scored yetFree - 19Devin Desktop (formerly Windsurf)Cognition
Agentic code editor from Cognition, renamed from Windsurf on 2 Jun 2026; included in Devin plans
Best for Devin subscribers who also want a code editor at no extra costYou give up Not scored yet: no independent result in the last 90 days
Not scored yetFrom $20/mo - 20WarpWarp
Terminal with a built-in coding agent; the free plan works with your own model keys
Best for People who want a modern terminal with an agent, free with their own model keysYou give up Not scored yet; cloud agents are billed separately from the plans
Not scored yetFree
| # | Name | Score | What placed it | Price | Best for | You give up |
|---|---|---|---|---|---|---|
| 1 | piEarendil · Open source | 90Our checklist | #4 app by AI usage through OpenRouter, a model marketplaceWhat it rests onTheir figure: OpenRouter #4 app by tokens (385B, data to 2026-10-17); npm 16.3M last month; GitHub 116.1k stars. checklist: job 2 (NTU, arXiv 2610.00917, 2026-10-01: pi v0.84.4, the newest release from Aug 28 to Sep 4, scored 0.603 on a 63-task Terminal-Bench 4 subset with GPT-6 Astra, the best of the harnesses tested), model 1 (no default model), value 2, privacy 2, limits 2 (v1.3.0 Oct 15) = 9 Report a mistake | Free (open source); you pay for the model | Developers who want a small, free agent they can shape around their own model | No default model; you bring and pay for your own key or subscription |
| 2 | OpenCodeAnomaly · Open source · our review | 80Our checklist | 7th of 8 we list on a coding-agent test (Artificial Analysis)What it rests onTheir figure: AA Coding Agent Index v1.5 53.6 (GLM-5.3); GitHub 216.4k stars. review 2026-10-11, unchanged: job 1, model 1, value 2, privacy 2, limits 2 = 8; AA v1.5 still 53.6 with GLM-5.3 today Report a mistake | Free (open source); optional Zen pay as you go | Developers who want a free agent and the choice of 75+ model providers | Scores below Codex and Claude Code; several free Zen models may train on prompts |
| 3 | OpenHandsOpenHands (All Hands AI) · Open source | 80Our checklist | #13 app by AI usage through OpenRouter, a model marketplaceWhat it rests onTheir figure: OpenRouter #13 app by tokens (80.6B, data to 2026-10-17); GitHub 91.2k stars. checklist: job 2 (NTU, arXiv 2610.00917, 2026-10-01: OpenHands SDK 1.44.1, the newest release from Aug 28 to Sep 7, scored 0.571 on a 63-task Terminal-Bench 4 subset with Claude Opus 5, above Claude Code at 0.492), model 1 (no default model), value 2, privacy 2, limits 1 (the openhands-ai app has had no release since 2026-07-09; the SDK is active) = 8 Report a mistake | Free (open source); cloud Individual free, models at cost | Teams who want to self-host an open agent, or try it free in the cloud | Free cloud tier allows 10 conversations a day; the app has had no release since July |
| 4 | Claude CodeAnthropic · Closed · our review | 70Our checklist | Ranked #1 of 23 on a coding-agent test (Terminal-Bench 4.0)What it rests onTheir figure: AA Coding Agent Index v1.5 68.4 (Sonnet 5.5 max); Terminal-Bench 4.0 64.8% (Opus 5.5), #1 of 23. review 2026-10-11, unchanged: job 2, model 2, value 1, privacy 1, limits 1 = 7; still top on AA v1.5 and Terminal-Bench 4.0 Report a mistake | Pro $20/mo; Max from $100/mo | Developers who already pay for Claude and want the strongest current coding model | No free plan; costs far more per task than Codex, which scores close |
| 5 | Devin (Devin CLI with Fusion)Cognition · Closed | 70Our checklist | 2nd of 8 we list on a coding-agent test (Artificial Analysis)What it rests onTheir figure: AA Coding Agent Index v1.5 68.1 (Fusion: Opus 5.5 + GPT-6 Luna). checklist: job 2 (AA v1.5 68.1, #2 row), model 2 (Fusion lead Opus 5.5, arena WebDev #1), value 1 ($20), privacy 1 (no training unless opt-in; export not stated), limits 1 (quota not published) = 7 Report a mistake | Free (limited); Pro $20/mo; Max $200/mo | Teams that hand tickets from Slack or Linear to an agent that returns pull requests | Usage quotas are not published, so it is hard to know how far $20 goes |
| 6 | Muse CodeMeta · Closed | 70Our checklist | 3rd of 8 we list on a coding-agent test (Artificial Analysis)What it rests onTheir figure: AA Coding Agent Index v1.5 68.0 (Muse Wonder Full, Power plan); 63.5 (Muse Wonder). checklist: job 2 (AA v1.5 63.5 with Muse Wonder, in $5 plan), model 1 (Muse Wonder arena WebDev #15), value 2 ($5), privacy 1 (standard not trained; contributor tier is), limits 1 (10-50 prompts per 5h) = 7 Report a mistake | Everyday $5/mo; High $15/mo; Power $50/mo | Coders on a tight budget who still want a strong agent | Only 10 to 50 prompts per five hours; a contributor tier trains on your data |
| 7 | Google Antigravity CLIGoogle · Closed · our review | Not scored | 4th of 8 we list on a coding-agent test (Artificial Analysis)NotesTheir figure: AA Coding Agent Index v1.5 63.8 (Gemini 4 Argon high). checklist (AI Pro): job 2 (AA v1.5 63.8 with Gemini 4 Argon), model 2 (Argon arena WebDev #9, preliminary), value 1 ($19.99), privacy 1 (opt-out in settings), limits 1 (entry limits on Pro) = 7 Report a mistake | Free; Google AI Pro $19.99/mo for Gemini 4 Argon | Anyone who wants to try a terminal coding agent for free, or who used Gemini CLI | Pricing page doesn't name models or limits; some Gemini CLI features are missing |
| 8 | OpenAI Codex CLIOpenAI · Open source · our review | 70Our checklist | 5th of 8 we list on a coding-agent test (Artificial Analysis)What it rests onTheir figure: AA Coding Agent Index v1.5 62.9 (GPT-6.1 Sol xhigh); Terminal-Bench 4.0 58.2%. review 2026-10-11, unchanged: job 2, model 2, value 1, privacy 1, limits 1 = 7; AA v1.5 62.9 and Terminal-Bench 4.0 58.2% today Report a mistake | ChatGPT Plus $20/mo; Pro from $100/mo; or API key | ChatGPT Plus subscribers who want a terminal coding agent at no extra cost | Default model unnamed; Plus limits vary widely and personal plans train by default |
| 9 | Grok BuildxAI · Closed | 60Our checklist | 6th of 8 we list on a coding-agent test (Artificial Analysis)What it rests onTheir figure: AA Coding Agent Index v1.5 56.3 (Grok 4.7 xhigh); Terminal-Bench 4.0 37.6%. checklist: job 1 (AA v1.5 56.3; Terminal-Bench 4.0 37.6% vs 64.8% top), model 1 (Grok 4.7 arena WebDev #16), value 2 ($8), privacy 1 (opt-out setting), limits 1 (shared weekly pool) = 6 Report a mistake | Free (limited); Plus $8/mo; Premium $30/mo | Trying a terminal agent cheaply at $8 a month, or free with limits | Solves about 38% of terminal tasks against about 65% for Claude Code |
| 10 | Kimi Code CLIMoonshot AI · Open source | 50Our checklist | Lowest of 8 we list on a coding-agent test (Artificial Analysis)What it rests onTheir figure: AA Coding Agent Index v1.5 51.9 (Kimi K3). checklist: job 1 (AA v1.5 51.9 with Kimi K3), model 1 (Kimi K3 arena WebDev #14), value 1 (needs Kimi Plus or higher), privacy 1 (opt-out by email), limits 1 (quota not published) = 5 Report a mistake | Not confirmed | Kimi Plus members who want an open-source agent on Kimi K3 | Needs a paid Kimi plan, and the usage quota is not published |
| 11 | JunieJetBrains · Closed | Not scored | Ranked #6 on recent real coding tasks, best agent app there (SWE-rebench)NotesTheir figure: SWE-rebench #6, 61.8% (May-Jul 2026 window), top agent product on that board. not scored: SWE-rebench window ended 1 Jul 2026, older than 90 days Report a mistake | Free with your own API keys; paid JetBrains AI plans | JetBrains IDE users, free if they bring their own model keys | Not scored yet: its last independent result is older than 90 days |
| 12 | CursorCursor (Anysphere, part of SpaceX) · Closed · our review | Not scored | Ranked #10 on recent real coding tasks (SWE-rebench)NotesTheir figure: SWE-rebench #10, CursorAgent 51.7% (May-Jul 2026 window). not scored: SWE-rebench window ended 1 Jul 2026, older than 90 days; no current independent agent test Report a mistake | Free; Plus $8/mo; Premium $30/mo | People who want an AI code editor and are happy to choose the model themselves | Trains on your code unless Privacy Mode is on; default model not named |
| 13 | Factory DroidFactory · Closed | Not scored | Ranked #10 of 142 on a retired coding-agent test (Terminal-Bench 2.0)NotesTheir figure: Terminal-Bench 2.0 77.3% (GPT-5.3-Codex), #10 of 142, archived board. not scored: only result is on the archived Terminal-Bench 2.0 board with an older model Report a mistake | Pro $20/mo; Plus $100/mo; Max $200/mo | Teams willing to pay for an agent that also runs on cloud machines | No free plan listed, from $20 a month; not scored yet |
| 14 | GitHub CopilotGitHub · Closed · our review | Not scored | Ranked #4 in a test of AI code review (Martian)NotesTheir figure: Martian online code review #4, 60.9% F1 (reported 6 Oct 2026). not scored: fresh independent results cover code review only, not its coding agent Report a mistake | Free; Pro $10/mo; Pro+ $39/mo; Max $100/mo | Anyone wanting a paid coding agent for the least money, or a free one to try | Free users can't pick models, and Claude and Codex agents need Pro+ at $39 |
| 15 | ClineCline Bot Inc. · Open source · our reviewSecurity note: see above | Not scored | About 70,500 GitHub stars: popular, which isn't the same as goodNotesTheir figure: GitHub 70.5k stars (adoption, used for ordering only). not scored: no independent result for Cline in the last 90 days Report a mistake | Free (open source); pay for model use | VS Code users who already pay for a model and want a free open agent | Quality depends on your model; a recent security flaw means updating to 3.0.30 |
| 16 | Kilo CodeKilo Code · Open sourceSecurity note: see above | Not scored | About 27,600 GitHub stars: popular, which isn't the same as goodNotesTheir figure: GitHub 27.6k stars (adoption, used for ordering only). not scored: no independent result for Kilo Code in the last 90 days Report a mistake | Free (open source); Kilo Pass from $19/mo | VS Code and JetBrains users who want an open agent at model makers' own rates | 5% fee on credit purchases; update past a security flaw fixed in 7.4.1 |
| 17 | OpenAI Codex CloudOpenAI · Closed · our review | Not scored | No outside test result foundNotesTheir figure: no independent result found; same Codex agent as the CLI. not scored: no independent test of Codex running in the cloud Report a mistake | ChatGPT Plus $20/mo; Pro from $100/mo | ChatGPT Plus or Business users who want coding tasks to run unattended | Not on Free, Go or an API key alone; cloud tasks use more of your plan allowance |
| 18 | Google AntigravityGoogle · Closed · our review | Not scored | No outside test result for the app foundNotesTheir figure: no independent result for the app found. not scored: AA tests the Antigravity CLI and SDK, not the desktop app by name Report a mistake | Free; Google AI Pro $19.99/mo | Trying a free coding agent app with several agents running at once | Free plan runs on weaker Gemini Flash; top Claude models need AI Pro at $19.99 |
| 19 | Devin Desktop (formerly Windsurf)Cognition · Closed | Not scored | No outside test result foundNotesTheir figure: no independent result found. not scored: no independent result for Windsurf in the last 90 days Report a mistake | Included in Devin Pro $20/mo | Devin subscribers who also want a code editor at no extra cost | Not scored yet: no independent result in the last 90 days |
| 20 | WarpWarp · Closed | Not scored | No outside test result foundNotesTheir figure: no independent result found. not scored: no independent result for Warp in the last 90 days Report a mistake | Free; Build $20/mo | People who want a modern terminal with an agent, free with their own model keys | Not scored yet; cloud agents are billed separately from the plans |
Nothing matches all of that. , or try everything.
How we ranked this
Order. Scored ones first, highest score first; on a tie, the better independent result. Unscored ones follow, by their public signal.
How we score this list. Our checklist: independent evidence that it does the job (from the last 90 days), the model under it, value, privacy and reliability. The full rules are on How we score. Open "How it was scored" on any row to see the working.
What we did not do. We did not run any of these ourselves. Every number comes from the public source linked on its row, checked on 19 Oct 2026.
Spot a mistake? A wrong rank, price or link: tell us and we will check it against the source. Corrections are dated in News.
Reviews are drafted with help from AI models, including Anthropic's Claude. Model scores come only from outside leaderboards and independent evaluations, by a written rule. Claude Code is made by Anthropic, whose Claude models help draft our reviews.