AI coding agents for people who are not developers

Prices and features checked: 2026-11-09. Prices in the currency the vendor shows, checked on the date shown; we do not convert.

Verdict by user type

A coding agent is not autocomplete. Autocomplete finishes the line you are typing; an agent takes a goal ("clean this spreadsheet export and flag duplicates"), writes the files, runs them, reads the error and tries again. That loop is what makes the tools useful to analysts, operations people and marketers who have never shipped software, and it is also what makes them dangerous, because the loop runs commands on your machine.

Our criteria, stated first: the price and what the entry plan includes, read from the vendor page on 2026-11-09; whether the agent runs in your terminal, your editor, or a cloud sandbox; which models you can choose; the documented permission model; the training and retention statement; and the licence where the tool is open source. None of the seven products has an affiliate programme in our catalogue; ChatGPT and Claude are listed there with "none". Then the evidence section, which gives both sides of the only randomised trials that exist, because the top results for this query are written by and for developers and quote only the favourable number.

We have not timed these tools ourselves. The three worked examples below are reproducible project briefs you can run in any of the seven, not logged results.

You are…	Pick	Why
A non-developer who already pays for Claude or ChatGPT	Claude Code (Pro, USD 17–20/mo) or Codex (Plus and above)	Included in the subscription; terminal, desktop and web surfaces; no new vendor
Someone who wants a free start with a familiar editor	GitHub Copilot Free, then Pro at USD 10	Agent mode and the CLI are included on Free; the pull-request coding agent starts at Pro
A person who wants an editor built around the agent	Cursor (Free, then Plus USD 8)	Agent on every tier; usage pool refreshes weekly; privacy mode
A technical analyst who wants a free, open tool and will pay a model provider directly	Aider (Apache 2.0)	Terminal pair programmer, works with "almost any LLM, including local models"
A team that wants agents sandboxed on its own infrastructure	OpenHands / Agent Canvas (MIT)	Docker sandboxes, bring your own model, runs Claude Code and Codex too
Someone searching for Windsurf	Devin (Cognition)	windsurf.com/pricing now redirects to devin.ai/pricing; Free, Pro USD 20, Max USD 200

Who this is for

Not professional developers. They have their own benchmarks and their own arguments. This page is for the analyst who cleans the same export every Monday, the operations lead who wants a script that reconciles two systems, and the marketer who needs an internal dashboard and has been quoted six weeks by IT. Search demand says the group exists: "ai coding agents" is searched about 8,100 times a month in the US, up 238 percent year on year, "coding ai" 49,500 and "best ai for coding free" 590 (DataForSEO, retrieved 2026-11-09).

For this group the right question is not "which agent writes the best code" but "which agent lets me get a small, checkable tool working without handing my laptop to a process I do not understand". That reframes the comparison around permissions and sandboxes, which is where the products differ most.

Comparison: price, where it runs, models, permissions

Tool	Entry price (checked 2026-11-09)	Where the agent runs	Model choice	Permission model as documented	Training and retention	Licence
GitHub Copilot	Free USD 0; Pro USD 10; Pro+ USD 39; Max USD 100 per user/mo; 1 AI credit = USD 0.01, Pro includes USD 15 of credits	IDEs (VS Code, JetBrains, Xcode, Vim, Zed, Eclipse), CLI, GitHub.com; coding agent opens pull requests	Anthropic Claude, OpenAI GPT, Google Gemini, xAI Grok, Microsoft MAI and others per plan	Not stated on the plans page	Individuals: GitHub "may use Copilot interaction data" to train, with opt-out; Business/Enterprise under a DPA	Proprietary
Claude Code	Pro USD 17/mo annual (USD 20 monthly); Max USD 100; Ultra USD 200; Team USD 20/seat annual	Terminal, VS Code, JetBrains, desktop app, web (claude.ai/code)	Claude models; terminal and IDEs "also support third-party providers"	Hooks run shell commands before or after actions; permission wording not on the overview page we read	Individual plans: training opt-out; Team and Enterprise: none by default	Proprietary
OpenAI Codex	Included in ChatGPT Plus ("Yes") and Pro ("Expanded"); limited on Free and Go; price amounts did not render	CLI, IDE, and "Codex Web", the cloud agent	OpenAI models	Not stated on the pages we could read	Not stated	CLI: Apache-2.0
Cursor	Free USD 0; Plus USD 8; Premium USD 30; Super USD 100; Ultra USD 200; Teams USD 30+/user	Editor, CLI, cloud agents from Premium	Plus includes Grok models only; third-party models "at API list price"	"Auto-run, browser, and network controls" listed as Enterprise	Privacy mode: code "not used for training by us or our model providers"; SOC 2, ISO 27001, ISO 42001 listed	Proprietary
Devin (formerly Windsurf URL)	Free USD 0; Pro USD 20; Max USD 200; Teams USD 80 plus USD 40 per dev seat	Cloud agent plus Devin Desktop	Not stated	Not stated; VPC deployment on Enterprise	Not stated	Proprietary
Aider	Free software; you pay your model provider	Terminal; watch mode inside any editor	"Almost any LLM, including local models"	Not stated; commits changes automatically with generated messages	Not stated (local tool; data goes to the provider you choose)	Apache-2.0
OpenHands (Agent Canvas)	Free software; cloud and enterprise offered without published prices	Self-hosted web control centre; Docker sandbox optional	"Use with any LLM"; runs Claude Code, Codex, Gemini and other ACP agents	No-sandbox mode gives the agent "full access to your filesystem"; Docker sandboxes offered	Not stated	MIT

Sources: ^{[1][2][3][4][5][6][7][8][9]}.

Two notes on the table. The GitHub plans page lists "Access to 3rd party agents (Claude Code and Codex)" in the Pro bullets while its comparison grid marks that feature for Pro+ and Max only; we report both and suggest you check before paying for it ^[1]. Cursor's pricing page on 2026-11-09 described its usage pool as "shared across Cursor and Grok" and listed Grok models as the included allowance ^[5]; if you want Claude or OpenAI models inside Cursor, the page says they are billed at API list price.

Three reproducible projects for a non-developer

Each brief fits in one message, produces a checkable result, and touches nothing outside a folder you create for it. Run them in a new empty directory, with the agent's permission prompts on.

1. Spreadsheet cleaner. "In this folder is export.csv. Write a Python script that trims whitespace, normalises email addresses to lower case, flags rows with duplicate emails in a new column, and writes clean.csv. Print how many rows were flagged. Do not modify export.csv." Check: open both files; the row count must match and the flagged count must be what you see with a filter.

2. Report script. "Read sales.csv (columns: date, region, product, amount). Produce report.md with totals by region and by month, and the top five products, as Markdown tables. Add a README that explains how to rerun it." Check: pick one region and verify its total in your spreadsheet.

3. Internal dashboard. "Build a single-file HTML dashboard that loads report.json and shows totals by region as a bar chart and a sortable table. No build step, no server; it must open by double-clicking." Check: it opens, it sorts, the numbers match project two.

Each brief is deliberately bounded: one input file, one output, no network, no credentials. That is the shape of task where the evidence below says assistance helps most, and where the damage from a wrong action is a file you can delete.

The evidence on speed and quality, both sides

Two randomised trials dominate this debate, and they point in opposite directions because they measure different people doing different work.

The favourable one. Peng and colleagues recruited 95 professional developers and asked them to implement an HTTP server in JavaScript; the group with GitHub Copilot finished 55.8 percent faster C^[10]. It is an arXiv preprint with a single artificial task and recruited freelancers, which is why we grade it C; it is also the number every vendor quotes.

The unfavourable one. METR's 2025 trial gave 16 experienced open-source developers 246 real tasks on repositories they knew well, randomised to allow or forbid early-2025 AI tools. With AI allowed they were 19 percent slower. They had expected to be 24 percent faster and, after the study, believed they had been 20 percent faster B^[11]. Small sample, very experienced developers, mature codebases, preprint: grade B. The result that should survive the caveats is the gap between felt and measured speed.

What this means for a non-developer. Neither trial studied you. The closest evidence comes from outside coding: in a field experiment with 758 consultants, GPT-4 raised speed and quality on tasks inside its capability and lowered correctness on tasks outside it B^[12]; in a quasi-experiment with 5,172 customer-support agents, generative-AI assistance raised productivity about 14 percent on average, with gains concentrated among the least experienced and near zero for the most experienced A^[13]. If that pattern transfers, a non-developer building bounded tools is close to the best case: low baseline skill, tasks well inside the model's reach, and output that can be checked against a spreadsheet. The METR developers were the opposite case. The honest summary is that the agent probably helps you more than it helps an expert, and that you are also less able to spot when it is wrong, which is why the projects above come with a check.

A meta-analysis of 106 human–AI experiments adds the structural warning: combinations performed worse than the better of human or AI alone on decision tasks and better on creation tasks A^[14]. Writing a script is creation; deciding whether its output is right is a decision. Keep the second one.

Security and data rules

• Sandbox first. OpenHands says the no-sandbox mode gives the agent "full access to your filesystem" and offers Docker sandboxes instead ^[9]; Codex Web and Cursor's cloud agents run remotely; Claude Code runs in your terminal with hooks you can set ^[3]. Whatever you use, work in a fresh folder and keep the permission prompts on until you understand what the agent asks for.
• Credentials. Never paste an API key, a database password or a customer export containing personal data into a brief. For the three projects above, use a sample file you have anonymised.
• Training. Copilot's individual plans let GitHub use interaction data for training unless you opt out ^[1]; Claude's individual plans are opt-out and its Team and Enterprise plans train on nothing by default ^[2]; Cursor's privacy mode states code is not used for training by Cursor or its providers ^[5]. Codex, Devin, Aider and OpenHands did not state it on the pages we read; for local tools, the statement that matters is your model provider's.
• Residency. None of the seven stated EU data residency on the pages we read. Anthropic's pricing page offers US-only inference as an API option at 1.1x price ^[2]. If residency is a requirement, a self-hosted tool with a provider that states a region is the only path today.
• Automation bias. People under-check automated aids, and the remedy is accountability and training rather than better interfaces A^[15]. Read the diff. Run the check. Every time.

Who should not use a coding agent

• Anyone who cannot check the output. If the script's result cannot be verified against something you already trust, you have built an oracle, not a tool.
• Anyone working on production systems or data they are not allowed to copy. Start with samples.
• Anyone who will measure success by feel. The METR developers felt 20 percent faster and were 19 percent slower B^[11].
• Anyone who wants a chat assistant for writing and research rather than files and commands; see our comparison of AI assistants for work.

In our Future Demand Radar, the "AI agents and delegated work" category scores 59.7 out of 100 in the baseline scenario (band 55.2–64.1) and is classed "build now"; our baseline scenario expects coding agents to remain the most mature sub-segment of that category, which is a forecast, not a fact. For agents that act in your business systems rather than in a folder, see the best AI agents for work; for running coding agents on your own server, open-source AI agents covers OpenHands and the licences.

What to do next

If you already pay for Claude or ChatGPT, open the coding agent that is included and run project one this week; the marginal cost is zero and the output is a CSV you can check. If you do not, GitHub Copilot Free gives you agent mode and the CLI at no cost, and Aider gives you a free, open tool if you are willing to put a few dollars on a model provider's account. Keep each project in its own folder, keep the permission prompts on, and write down how long it took before you decide you were faster. Then read how to manage AI agents for the delegation routine, run the Tool Finder to see these prices next to the assistants and automation tools, and subscribe to the briefing below to receive the three project briefs above as a starter prompts pack you can paste into any of the seven.

Frequently asked questions

Which AI coding agent is best for a non-developer?

The one included in a subscription you already pay for: Claude Code with Claude Pro (USD 17 to 20 a month) or Codex with ChatGPT Plus. If you pay for neither, GitHub Copilot Free includes agent mode and the CLI at no cost. Start with a bounded project in an empty folder and keep the permission prompts on. Prices read 2026-11-09.

Is there a free AI coding agent?

Yes. GitHub Copilot Free includes agent mode and the Copilot CLI with limited usage; Cursor and Devin have free tiers; Aider (Apache-2.0) and OpenHands (MIT) are free software, though you pay a model provider for the calls or run a local model. Claude's Free plan does not include Claude Code; Codex is limited on ChatGPT Free.

Do AI coding agents make you faster?

The two randomised trials disagree. Peng et al. 2023 found 95 freelance developers finished a single task 55.8 percent faster with Copilot (preprint, grade C). METR 2025 found 16 experienced open-source developers were 19 percent slower with early-2025 AI tools on their own repositories while believing they were 20 percent faster (preprint, grade B). Neither studied non-developers; evidence from other fields suggests novices gain most and are least able to catch errors.

Are AI coding agents safe to run on my laptop?

Only inside limits you set. Work in a fresh folder, keep permission prompts on, never paste credentials or personal data into a brief, and prefer a sandbox: OpenHands offers Docker sandboxes and warns that its no-sandbox mode gives the agent full access to your filesystem; Codex Web and Cursor's cloud agents run remotely. Read every diff before you accept it.

Do coding agents train on my code?

It depends on the plan. GitHub Copilot's individual plans let GitHub use interaction data to train unless you opt out; Claude's individual plans are opt-out and its Team and Enterprise plans train on nothing by default; Cursor's privacy mode states code is not used for training. Codex, Devin, Aider and OpenHands did not state it on the pages we read on 2026-11-09; for local tools, your model provider's policy is the one that applies.

What happened to Windsurf?

On 2026-11-09, windsurf.com/pricing redirected to devin.ai/pricing, which lists Devin plans from Cognition (Free, Pro USD 20, Max USD 200, Teams USD 80 plus USD 40 per developer seat) and does not mention the Windsurf name. We report what the page states and nothing more.

Sources and evidence

1. GitHub Copilot plans (Free, Pro, Pro+, Max; AI credits; coding agent availability; data use) — GitHub (2026), vendor page A github.com/features/copilot/plans (retrieved 2026-11-09)
2. Claude pricing (Pro, Max, Ultra, Team; Claude Code inclusion; training policy) — Anthropic (2026), vendor page A claude.com/pricing (retrieved 2026-11-09)
3. Claude Code overview (surfaces, hooks, third-party providers) — Anthropic (2026), vendor page A code.claude.com/docs/en/overview (retrieved 2026-11-09)
4. ChatGPT pricing (Codex availability per plan; price amounts did not render) — OpenAI (2026), vendor page B chatgpt.com/pricing/ (retrieved 2026-11-09)
5. Cursor pricing (tiers, usage pool, privacy mode, certifications) — Cursor (2026), vendor page A cursor.com/pricing (retrieved 2026-11-09)
6. Devin pricing (redirect target of windsurf.com/pricing) — Cognition (2026), vendor page A devin.ai/pricing (retrieved 2026-11-09)
7. Aider repository and documentation (Apache-2.0; supported models) — Aider AI (2026), vendor page A github.com/Aider-AI/aider (retrieved 2026-11-09)
8. OpenAI Codex CLI repository (Apache-2.0; plan inclusion) — OpenAI (2026), vendor page A github.com/openai/codex (retrieved 2026-11-09)
9. OpenHands / Agent Canvas repository (MIT; sandbox options; requirements) — OpenHands contributors (2026), vendor page A github.com/All-Hands-AI/OpenHands (retrieved 2026-11-09)
10. The Impact of AI on Developer Productivity: Evidence from GitHub Copilot — Peng S, Kalliamvakou E, Cihon P, Demirer M (2023), RCT C doi.org/10.48550/arXiv.2302.06590 (retrieved 2026-11-09)
11. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — Becker J, Rush N, Barnes E, Rein D (METR) (2025), RCT B doi.org/10.48550/arXiv.2507.09089 (retrieved 2026-11-09)
12. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — Dell'Acqua F, McFowland E, Mollick ER, et al. (2023), RCT B doi.org/10.2139/ssrn.4573321 (retrieved 2026-11-09)
13. Generative AI at Work — Brynjolfsson E, Li D, Raymond L (2025), observational A doi.org/10.1093/qje/qjae044 (retrieved 2026-11-09)
14. When combinations of humans and AI are useful: A systematic review and meta-analysis — Vaccaro M, Almaatouq A, Malone T (2024), meta-analysis A doi.org/10.1038/s41562-024-02024-1 (retrieved 2026-11-09)
15. Complacency and Bias in Human Use of Automation: An Attentional Integration — Parasuraman R, Manzey DH (2010), systematic review A doi.org/10.1177/0018720810376055 (retrieved 2026-11-09)