The best AI coding agent for most UK developers who work in the terminal is Claude Code, which comes with Claude Pro at £18 a month including VAT. Cursor is the strongest AI editor. GitHub Copilot's cloud agent is the pick for teams whose code lives on GitHub, from $10 (about £7.50) a month. Google Antigravity is the best free option, and Kilo Code is the one to choose if you want local, private models. I checked every price below on the official pages on 25 October 2026. This guide compares eight agents on where they run, what they cost a solo developer and a five-person team, how usage is limited, which models they use and whether they train on your code by default.
Key facts
- Best for terminal users: Claude Code, from £18 a month including VAT, or £15 a month billed yearly.
- Best AI editor: Cursor, free to start; cloud agents and non-Grok models from $30 (about £22.50) a month plus VAT.
- Best for teams on GitHub: Copilot cloud agent, from $10 (about £7.50) a month, or $19 (about £14.25) a seat on Business.
- Best free: Google Antigravity, free with weekly limits.
- Best for local and private models: Kilo Code, free and open source, with Ollama or LM Studio.
- Training: on personal plans, most of these agents can train on your code unless you switch it off.
On this page
- The picks: which AI coding agent should you choose?
- AI coding agents compared: the main table
- How to choose an AI coding agent
- What each costs a solo developer and a five-person team
- Privacy and training settings
- Running a coding agent with a local model
- Frequently asked questions
- Read next
- Sources
The picks: which AI coding agent should you choose?
Five agents stand out, each for a different kind of developer. The prices are the cheapest plan that includes the agent, as of 25 October 2026.
- Best overall for terminal users: Claude Code. It reads your project, edits files and runs commands from plain-English instructions, in the terminal, VS Code, JetBrains, a desktop app and the web. Anthropic shows UK prices in pounds with VAT included: Pro is £18 a month.
- Best IDE: Cursor. It is a full code editor built around agents, with local and cloud agents. My Claude Code vs Cursor guide compares the two in depth. The free plan is limited, and the $8 (about £6) Plus plan offers Grok models only. Claude and other models need Premium at $30 (about £22.50) a month, and they are billed at API list price on top.
- Best for teams on GitHub: GitHub Copilot cloud agent. You assign an issue to Copilot and it opens a pull request for review. It starts with Copilot Pro at $10 (about £7.50) a month, and Business costs $19 (about £14.25) a seat.
- Best free: Google Antigravity. The individual plan costs nothing and comes with weekly limits. Google AI Pro at £18.99 a month raises them.
- Best for local and private models: Kilo Code. The extension is free and open source, and its pricing page says you can "run local models with Ollama or LM Studio" and pay Kilo nothing.
Three more are worth knowing. OpenAI Codex is the obvious choice if you already pay for ChatGPT, and ChatGPT Free includes limited use. For the two-way choice, see Codex vs Claude Code. Devin suits teams that want to hand tasks to several agents at once. Meta Muse Code is the cheapest paid option at $5 (about £4) a month, but it trains on your code by default.
AI coding agents compared: the main table
Prices are the cheapest plan that includes each agent, checked on official pages on 25 October 2026. Pound prices with VAT are the companies' own UK prices. Dollar prices exclude VAT unless stated, and I have converted them at $1 = £0.75.
| Product | Where it runs | Cheapest plan with it (UK) | How usage is limited | Models available | Local models? | Trains on your code by default? | Best for |
|---|---|---|---|---|---|---|---|
| Claude Code | Terminal, VS Code and JetBrains, desktop app, web | Claude Pro, £18 a month including VAT (£15 a month billed yearly) | Allowance refills every five hours, plus a weekly cap, shared with Claude chat | Anthropic only: Opus 5.5, Sonnet 5.5, Haiku 5.5 | Not supported by Anthropic; works through Ollama | Yes on Pro and Max unless you switch it off; no on Team, Enterprise and the API | Terminal users handing over multi-step jobs |
| OpenAI Codex | Desktop app, open-source CLI, editor extension, web cloud tasks | ChatGPT Free (limited); Plus £20 a month (OpenAI does not say whether VAT is included) | Local messages per five-hour window; weekly limits may apply | OpenAI only: GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Astra, GPT-6 Luna | Yes, the CLI with Ollama or LM Studio | Yes on Free, Go, Plus and Pro unless you opt out; no on Business | People already paying for ChatGPT |
| GitHub Copilot cloud agent | GitHub's cloud, started from issues; editors, Copilot CLI, Copilot app | Copilot Pro, $10 (about £7.50) a month; Free excludes the cloud agent | Monthly GitHub AI Credits worth $0.01 each | Many, including Claude, GPT, Gemini and Grok | Copilot CLI can use Ollama models; the cloud agent cannot | Yes on Free, Pro and Pro+ since 24 April 2026 unless you opt out; no on Business and Enterprise | Teams whose code and issues live on GitHub |
| Cursor | Its own desktop editor (based on VS Code), CLI, cloud agents | Free (limited); Plus $8 (about £6), Grok models only | Weekly usage pool shared with Grok; third-party models at API list price | Grok on Plus; Anthropic, Google and others at API price from Premium | None listed on its pricing page | Yes unless Privacy Mode is on; Cursor does not say which is the default | Hands-on coding in an AI editor |
| Devin | Devin Cloud (browser, Slack, Teams), Devin Desktop (formerly Windsurf), CLI | Free (light quota); Pro $20 (about £15) | Daily and weekly allowance; extra usage at API prices | OpenAI, Claude, Gemini, SpaceXAI and open models, plus its own SWE-2 | No local option documented | Yes; paid users can opt out, Free users cannot | Teams handing tasks to several agents at once |
| Google Antigravity | Desktop app, CLI (agy), editor extensions | Free with weekly limits; Google AI Pro £18.99 a month | Weekly limits on the free plan; on Pro, refreshes every five hours up to a weekly cap | Gemini 3.8 Flash on free; Gemini 4 Argon and Claude 5.5 on Pro | Only through its SDK | Yes, through telemetry, unless you switch it off | Trying agentic coding for free |
| Kilo Code | VS Code, JetBrains, CLI, cloud agents | Free extension; models at provider rates plus 5% on top-ups | Pay per use, or Kilo Pass credits from $19 (about £14) a month | 500+ models through Kilo Gateway, your own keys or local models | Yes, through Ollama or LM Studio | Not stated plainly; it depends on the model provider, and local models keep code on your machine | Choosing your own models, including private local ones |
| Meta Muse Code | Terminal | Everyday, $5 (about £4) a month; no free plan | Prompts per five hours (10-50 on Everyday) | Meta's Muse Wonder; Wonder Full on Power ($50) and above | None documented | Yes, Contributor mode is on by default | Cheap, long-running agent jobs |
Two details in the table need care. Google does not say whether its £18.99 includes VAT. And GitHub, Cursor, Cognition, Kilo and Meta show only dollar prices, so check the total at checkout, where UK VAT may be added.
How to choose an AI coding agent
The right agent depends more on how you work than on benchmark scores. Five questions narrow it down.
- What do you already pay for? Every ChatGPT plan includes some Codex use, every paid Claude plan includes Claude Code, and Google AI Pro raises Antigravity's limits. Starting with the subscription you have costs nothing extra.
- Where do you want the work to happen? Claude Code, Codex and Muse Code are terminal-first. Cursor and Antigravity are editors. Copilot's cloud agent and Devin work in the cloud and hand back pull requests.
- How is usage limited? Claude Code and Codex use five-hour windows with weekly caps; my guide to Claude usage limits covers how Claude Code shares them with chat. Copilot counts monthly credits. Kilo Code charges per token. Heavy days hit window-based limits first, so check how each plan tops up.
- Can your code leave your machine? If not, use Kilo Code or the Codex CLI with a local model. See running a coding agent with a local model.
- Will it train on your code? On personal plans, most will unless you switch it off. For client code, use a business plan or change the setting first. See privacy and training settings.
If you would rather not write code at all, a coding agent may be the wrong tool. Our guide to what AI agents cost in the UK covers the other routes.
What each costs a solo developer and a five-person team
These sums use monthly billing and the plan a person would realistically start on, as of 25 October 2026. Anthropic's pound prices include 20% VAT. OpenAI also shows pound prices but does not say whether they include VAT, so I use them as shown; check the total at checkout. For dollar prices, I convert at $1 = £0.75 and then add 20% VAT, because most of these companies say tax may be added at checkout. If your business is VAT-registered, ask your accountant whether you can reclaim it.
| Product | Solo developer: plan and sum | Solo, a month with VAT | Five-person team: plan and sum | Team, a month with VAT |
|---|---|---|---|---|
| Claude Code | Claude Pro: £18, VAT included | £18.00 | Team Standard: 5 x £22.50 = £112.50, VAT included | £112.50 |
| OpenAI Codex | ChatGPT Plus: £20; OpenAI does not say whether VAT is included | £20.00 | Business: 5 x £18 = £90; OpenAI does not say whether VAT is included | £90.00 |
| Copilot cloud agent | Pro: $10 x 0.75 = £7.50, + 20% VAT | £9.00 | Business: 5 x $19 = $95 x 0.75 = £71.25, + 20% VAT | £85.50 |
| Cursor | Premium: $30 x 0.75 = £22.50, + 20% VAT | £27.00 | Teams: 5 x $30 = $150 x 0.75 = £112.50, + 20% VAT | £135.00 |
| Devin | Pro: $20 x 0.75 = £15, + 20% VAT | £18.00 | Teams: $80 + (5 x $40) = $280 x 0.75 = £210, + 20% VAT | £252.00 |
| Google Antigravity | Free (£0), or Google AI Pro: £18.99 | £0 or £18.99 | Gemini Enterprise: 5 x $30 = $150 x 0.75 = £112.50, + 20% VAT | £135.00 |
| Kilo Code | Kilo Pass Starter: $19 x 0.75 = £14.25, + 20% VAT (or £0 with local models) | £17.10 | Teams: 5 x $15 = $75 x 0.75 = £56.25, + 20% VAT, plus model usage | £67.50 |
| Meta Muse Code | Everyday: $5 x 0.75 = £3.75, + 20% VAT | £4.50 | No team plan listed: 5 x $5 = $25 x 0.75 = £18.75, + 20% VAT | £22.50 |
What the sums leave out matters as much as what they include.
- Yearly billing is cheaper. Claude Pro is £180 a year (£15 a month), and Claude Team is 5 x £18 = £90 a month billed yearly. ChatGPT Business is 5 x £15 = £75 a month billed yearly, with a two-seat minimum.
- Extra usage costs extra. Claude, Codex and Devin let you buy more usage at API rates when you hit a limit. Cursor bills Claude and other third-party models at API list price on top of the plan.
- Cheaper plans exist, with limits. Cursor Plus is $8 (about £6) but offers Grok models only and no cloud agents. ChatGPT Go is £7 with lighter Codex use.
- Devin's team sum follows its pricing page: "$80/month for team plan + $40/mo per full dev seat".
Privacy and training settings
On personal plans, most of these agents can use your code to train models unless you change a setting. Business plans generally do not. Change the setting before you open client or employer code.
| Product | Trains by default on the cheapest plan? | What to change | Plans without training by default |
|---|---|---|---|
| Claude Code | Yes, on Free, Pro and Max | Switch off Model Improvement in Claude's privacy settings | Team, Enterprise and the API |
| OpenAI Codex | Yes, on Free, Go, Plus and Pro | Opt out in ChatGPT's settings | Business, Enterprise and Edu |
| Copilot cloud agent | Yes, on Free, Pro and Pro+, since 24 April 2026 | Opt out under Privacy at github.com/settings/copilot | Business and Enterprise |
| Cursor | Yes, when Privacy Mode is off | Turn Privacy Mode on | Any plan with Privacy Mode on |
| Devin | Yes | Opt out on the Data Controls page (paid plans only; on Teams, only an admin can) | Enterprise |
| Google Antigravity | Yes, through telemetry | Switch off Enable Telemetry under Account in the app's settings | Gemini Enterprise |
| Kilo Code | Not stated plainly | Use a local model, or check your model provider's terms | Local models keep code on your machine |
| Meta Muse Code | Yes, Contributor mode is on by default | Switch Contributor mode off; it covers future content only | Standard mode and standard API models |
Training is only part of the picture. So is what the agent can reach on your own machine: my guide to NVIDIA OpenShell compares the sandboxes that limit it. Each cloud agent sends your prompts and the code it reads to the company's servers, mostly in the US. If that code includes customers' personal data, read our guide to AI agents and UK GDPR before you connect it. Each per-tool guide linked above covers that tool's data controller, retention and transfer terms.
Running a coding agent with a local model
Several of these agents can work with a model running on your own computer, so your code does not leave it. Local models are slower and weaker than the companies' cloud models, and they need plenty of memory.
- Kilo Code: supported. Choose Ollama or LM Studio as the provider, and Kilo charges nothing.
- Codex CLI: supported. Ollama's documentation gives
codex --ossorollama launch codex, and asks for a context window of at least 64k tokens. Codex cloud tasks do not work with local models. - Claude Code: works, but Anthropic does not support it. Ollama documents
ollama launch claude, or three environment variables pointing Claude Code athttp://localhost:11434. Ollama says hosted web search and advanced tool controls are not fully supported. Our step-by-step guide to running Claude Code with Ollama on a Mac covers the models that work and what you give up. - Copilot CLI: can find models in a running Ollama install through
/model, since 7 October 2026. The cloud agent itself always runs in GitHub's cloud. - Google Antigravity: only its Python SDK runs local models. The app and CLI use Google's cloud.
Memory is the real cost. Ollama recommends a context of 64,000 tokens or more for coding agents, which needs several gigabytes on top of the model. In practice, 32 GB of memory is the floor for the 19 GB coding models such as qwen3-coder, and 16 GB Macs can run only small models with weak results. Our local LLM memory calculator does the sums for your machine, and the Ollama guide covers installing it.
Frequently asked questions
What is the best AI coding agent in the UK in 2026?
For most developers who work in the terminal, Claude Code, which comes with Claude Pro at £18 a month including VAT. Cursor is the stronger pick if you want an AI editor, GitHub Copilot's cloud agent suits teams on GitHub from $10 (about £7.50) a month, and Google Antigravity is free with weekly limits. Prices checked 25 October 2026.
Claude Code vs Cursor: which should I choose?
Choose Claude Code if you want an agent that works in your terminal and editor on Anthropic's models, from £18 a month including VAT. Choose Cursor if you want a full AI code editor with cloud agents. Cursor is free to start, but Claude and other third-party models need Premium at $30 (about £22.50) a month plus VAT, and they are billed at API list price on top.
Claude Code vs Codex: which is cheaper in the UK?
Codex is cheaper to start, because ChatGPT Free includes limited Codex use and its CLI is free, open source and can run local models. For full use, Claude Pro costs £18 a month including VAT and ChatGPT Plus £20 a month, as of 25 October 2026; OpenAI does not say whether its pound prices include VAT. For a five-person team on monthly billing, Claude Team Standard costs £112.50 a month including VAT and ChatGPT Business £90 (5 x £18 a seat), as of 29 October 2026.
Is there a free AI coding agent?
Yes. Google Antigravity has a free plan with weekly limits, ChatGPT Free includes limited Codex use, Devin and Cursor have limited free plans, and Kilo Code is free and open source if you bring your own API keys or run a local model. Claude Code and Muse Code have no free plan, and Copilot Free excludes the cloud agent.
Read next
- Claude Code explained: UK pricing, install and how it compares
- OpenAI Codex explained: UK pricing and install on Mac and Windows
- GitHub Copilot coding agent: UK prices and how to assign issues
- Cursor review: AI coding agents on your desktop and in the cloud
- How to run Claude Code with Ollama and a local model on your Mac
Sources
- Claude plans and pricing, Anthropic (UK view) - Pro £18 monthly or £180 a year, Team £22.50 or £18 a seat, prices include 20% VAT - checked 25 October 2026
- ChatGPT pricing, OpenAI (UK view) - Go £7, Plus £20, Pro £89, Business £15 a seat billed yearly or £18 monthly; no statement on VAT - checked 25 and 29 October 2026
- Codex pricing and usage limits, OpenAI - Business $20 or $25 a user, five-hour limits, models - checked 25 October 2026
- GitHub Copilot plans - Pro $10, Pro+ $39, Max $100, AI credits, cloud agent by plan - checked 25 and 26 October 2026
- Plans for GitHub Copilot, GitHub Docs - Business $19 and Enterprise $39 a seat - checked 25 October 2026
- Updates to GitHub Copilot interaction data usage policy, GitHub - training on Free, Pro and Pro+ from 24 April - checked 25 October 2026
- Cursor pricing - Free, Plus $8, Premium $30, Super $100, Ultra $200, Teams $30+ a user - checked 25 and 26 October 2026
- Cursor data use - training when Privacy Mode is off - checked 25 October 2026
- Devin pricing, Cognition - Free, Pro $20, Max $200, Teams $80 plus $40 a full seat - checked 25 and 26 October 2026
- Google Antigravity pricing - free individual plan, business seats from $30 - checked 25 October 2026
- Google One plans (UK) - Google AI Pro £18.99 a month - checked 25 October 2026
- Kilo pricing - free extension, Kilo Pass from $19, Teams $15 a user, local models - checked 25 October 2026
- What is a Muse Code subscription?, Meta - plans from $5, Contributor mode default - checked 25 October 2026
- Claude Code integration, Ollama documentation - ollama launch claude, 64k context, caveats - checked 25 October 2026
- Codex integration, Ollama documentation - codex --oss, ollama launch codex, 64k context - checked 25 October 2026
- Training, model and local-model details for each tool come from my linked guides, which cite the official terms and documentation, checked 17-25 October 2026
