Refact.ai
Refact.ai is an open-source, local-first AI coding agent for VS Code and JetBrains. It plans, writes, debugs and ships code from your IDE, using model providers you choose or fully local runtimes, with on-premise deployment available.
What is Refact.ai?
Refact.ai is an autonomous, open-source AI coding agent that runs from your IDE rather than replacing it. A single refact binary hosts a resident daemon, a terminal interface, an in-browser GUI and per-project workers; the VS Code and JetBrains plugins are thin clients of that daemon. Three uses sit on top of it: an agent that takes a task from planning to deployment, a project-aware chat, and real-time autocompletion built on the Qwen2.5-Coder model with retrieval-augmented generation over your repository.
Context is the selling point. The agent reads the workspace, the codebase, connected databases, files, documentation and the web, then acts through integrations — GitHub, GitLab, Bitbucket, Docker, PostgreSQL, MySQL, MCP servers, a shell and a Chrome-driven browser. The documentation adds agent modes, sub-agents, worktrees, a task planner, persistent memory, context compression and a cron scheduler. More than twenty-five programming languages are supported, among them Java, Python, JavaScript, TypeScript, Rust, C, C++, PHP, Ruby and SQL, across VS Code, the JetBrains family and beta plugins for Visual Studio, Neovim and Sublime Text.
The publisher is Small Magellanic Cloud AI Ltd, a British company started in 2021 by Oleg Klimov — previously at OpenAI — with Oleg Kiyashko and Vlad Guber. Its blog claims strong public benchmark results: top open-source agent on SWE-bench Verified in June 2025, 60.0% on SWE-bench Lite, and 93.3% on the Aider Polyglot benchmark in thinking mode.
One thing dominates any current assessment. On 30 April 2026 the company announced that Refact Cloud is shutting down: hosted accounts, managed inference, credits, subscriptions and cloud team features all go, with the final date still unpublished. The product itself continues as an open-source, local-first, bring-your-own-key project maintained by its community, and the documentation has moved to a GitHub wiki on a community fork. What you adopt today is a local engine you configure and control, not a managed service.
What it does
- Hand the agent an engineering task and let it plan, execute and deploy it end to end
- Chat about the project inside the IDE, with the repository supplying the context
- Complete code in real time using retrieval-augmented suggestions drawn from your own codebase
- Wire the agent into GitHub, GitLab, Docker, PostgreSQL, MySQL, MCP servers, a browser and a shell
- Point the engine at any model you like, from hosted providers to fully local runtimes, with your own key
- Fine-tune models on the company codebase and coding style
- Deploy on-premise, as SaaS, on AWS or inside your own VPC
When to use Refact.ai / When not to
A quick filter to help you decide if Refact.ai is the right fit.
When to use Refact.ai
- Software engineers who want to hand over a whole task — planning, coding, testing, deployment — instead of accepting line-by-line autocompletion
- Teams bound by strict code-confidentiality rules, who need on-premise deployment so that source never leaves their own servers
- Developers who insist on choosing the model themselves, whether a hosted provider such as Anthropic, OpenAI or DeepSeek, or a local runtime like Ollama, LM Studio or vLLM
- Engineering organisations that want models fine-tuned on their own codebase and stack, running across multiple GPUs
- Non-developers building working prototypes by describing what they want, as several testimonials on the homepage describe
When not to use Refact.ai
- Anyone shopping for a fully managed cloud service: Refact Cloud is being retired, taking hosted accounts, managed inference and cloud billing with it
- Users unwilling to hold their own provider API keys or run a local model runtime, since model billing now sits with whichever provider they configure
- Mobile-first users: there is no iOS or Android application, only IDE plugins and a local engine
- Buyers who need contractual commercial support, which gives way to GitHub issues, discussions, pull requests and Discord once the transition completes
- Organisations that require written compliance guarantees: the site claims no GDPR conformity and publishes no DPA, no subprocessor list and no retention period
How to use Refact.ai
A typical end-to-end flow, from setup to results.
- Install the plugin in your IDE: VS Code from the Visual Studio Marketplace, or JetBrains from the plugin repository
- Let the plugin start the local refact-lsp engine — there is no separate backend to deploy
- Skip account creation: the normal setup path needs no hosted Refact login
- Add at least one provider in Provider Setup, either a hosted service with your own API key or a local runtime such as Ollama, LM Studio or vLLM
- Pick your Default Models: one strong chat or agent model, a faster chat model, a completion model, plus an embedding model for semantic search
- Enable only the integrations you want the agent to reach, such as GitHub, GitLab, Docker, PostgreSQL, MySQL, MCP, shell or browser
- Describe a task in plain language and let the agent plan and execute it, previewing and steering the steps as it goes
- For a company evaluation, request a demo through the contact form or apply for the two-week enterprise trial
- For self-hosting, start from the refact-self-hosting repository and deploy on-premise, as SaaS, on AWS or in your VPC
- If you were a Refact Cloud user, update the plugin and engine, configure a provider, export anything still held in the cloud, and move credentials and settings into your local configuration
Pros & Cons
Pros
- Genuinely open source and self-hostable, including for individuals rather than enterprises only
- Local-first architecture: the engine runs on your machine and your code need never leave it
- Free choice of model provider or local runtime, so you are not locked to one vendor's catalogue
- Real on-premise deployment, which the site's own comparison tables show Cursor and GitHub Copilot lack
- Fine-tuning on a company codebase, plus a broad set of tool and database integrations
- Public benchmark results claimed on SWE-bench Verified, SWE-bench Lite and Aider Polyglot
- Works inside the IDE you already use instead of asking you to adopt a new editor
Cons
- Refact Cloud is closing: hosted accounts, managed inference, credits, subscriptions and cloud team features all disappear, with no final date published
- The homepage pricing block and the fair use policy still describe the paid cloud tiers the shutdown notice retires, so the site contradicts itself
- Documentation now lives on the GitHub wiki of a community fork, and docs.refact.ai simply redirects there
- No GDPR statement, no DPA, no subprocessor list, no retention period and no minimum age anywhere on the site
- The privacy policy is very short and undated; the terms date from January 2023
- No postal address is published, and there is no support email — help runs through Discord and GitHub
- Bring-your-own-key shifts both the cost and the data policy onto whichever provider you pick
Pricing & Plans
A permanent free plan is offered. The cheapest paid entry point advertised on the homepage is the Pro plan, from USD 10.00 per month, with Enterprise priced on request. This figure needs a caveat: the announcement of 30 April 2026 states that active Refact Cloud subscriptions are being terminated and that no paid cloud subscription layer will remain, while the pricing block and the fair use policy have not been updated to match. The June 2026 documentation is blunter still — Refact does not sell model access, and billing, quotas and rate limits belong to whichever provider or local runtime you configure.
- autonomous agent with limited daily usage (8 requests a day
- 240 a month)
- in-IDE chat with 32k context
- unlimited completions powered by Qwen2.5-Coder
- code-aware vector database
- 40 agent requests a day (1
- 200 a month)
- unlimited in-IDE chat with 64k context
- additional code completion models
- LLM fine-tuning on the company codebase
- multi-GPU optimisation
- code privacy
- priority support and onboarding
- Two-week enterprise trial available through the trial form
- for self-hosted or SaaS evaluation
- the 30 April 2026 notice states that active Refact Cloud subscriptions will be terminated and paid invoices from the last 30 days refunded
Data, GDPR & hosting
A consolidated view of how Refact.ai handles your data.
GDPR overview
No page collected on the site mentions the GDPR at all. Small Magellanic Cloud AI Ltd is a British company and its terms are governed by the laws of England and Wales, yet nothing names a data protection officer, an Article 27 EU representative, a data processing agreement, subprocessors or a retention period. The privacy policy is short, carries no last-updated date, and asserts that no data is collected server-side; deletion requests go by email to info@smallcloud.tech. The terms date from 9 January 2023 and the cookies policy from 27 December 2022, both older than the product they cover. A single SOC 2 reference appears in a February 2025 blog post about an AWS Inferentia2 deployment, not as a product certification.
Who owns the data?
The terms of use, last updated on 9 January 2023, say plainly that code the plugin suggests in your editor becomes your intellectual property. Small Magellanic Cloud AI Ltd receives no rights over code or other intellectual property you develop while using the service, and grants you a non-exclusive, perpetual, royalty-free licence over any suggestion you accept. The service itself and its materials stay the property of the company or its licensors. Responsibility for what you produce is yours, including spotting cases where generated code reproduces known works or breaches their licences. Operational data — provider settings, chat trajectories, project knowledge, local indexes — stays in local directories on your own machine.
Reuse rights
You may reuse the code the agent produces without asking permission: the terms assign it to you outright and grant a perpetual, royalty-free licence over the suggestions you accept. The limits sit elsewhere. You may not resell, disassemble or reverse-engineer the plugin, the models or anything else supplied, and you may not use the service's own materials commercially without written consent. The fair use policy forbids feeding prompts and results into the training of another AI model — it calls this model scraping — and rules out spam, fraud, phishing, malware and unauthorised access. Checking that generated code does not reproduce licensed works remains your responsibility, not the vendor's.
Data retention & training
Hosting summary
No hosting country or region for user data is published anywhere on the site. The privacy policy states flatly that no data is collected server-side, that transfers are encrypted with SSL/TLS and encrypted at rest, and that access is role-based, monitored and limited to named personnel working on the project. The June 2026 documentation goes further: operational data — provider settings and credentials, chat trajectories, task metadata, project knowledge, local indexes, usage summaries and integration configuration — lives in local user and project directories on your own machine. What leaves it goes only to the providers, endpoints and integrations you enable. On-premise deployment is presented as ensuring code never leaves your servers, with SaaS, AWS and VPC offered as alternatives. The company is British and its terms fall under the laws of England and Wales; the marketing site itself resolves to a Google Cloud address in the Netherlands, which says nothing about where user data sits.
Things to keep in mind
Risks and trade-offs to weigh before adopting Refact.ai.
- The hosted cloud is closing with no final date announced: anyone relying on Refact Cloud must migrate, export their data and reconfigure before the switch-off
- The site contradicts itself on price — the homepage still sells a Pro plan the shutdown notice retires — so budget on the documentation, not the marketing block
- Bring-your-own-key moves the data policy to your chosen provider: the guarantees that then apply are theirs, not Refact's, and you have to read them yourself
- An agent with shell, browser, database and repository access can act destructively if left unsupervised; preview the steps rather than trusting them wholesale
- Delegating whole tasks erodes familiarity with your own codebase over time, and a developer who stops reading generated code loses the ability to review it
- You carry the legal risk on generated code: the terms make you responsible for spotting reproductions of known works and licence breaches
- Thin legal documentation — no GDPR statement, no DPA, no subprocessor list, no retention period — leaves regulated organisations without the paperwork they need
Setup & Integrations
Technical difficulty
Basic setup is easy: install the VS Code extension or JetBrains plugin and the local engine starts itself, with no backend to deploy and no account to create. One extra step is now unavoidable — configuring a provider with your own key, or a local runtime such as Ollama, LM Studio or vLLM, then choosing default models. Each integration you enable needs its own configuration. Self-hosting, multi-GPU fine-tuning and migration away from Refact Cloud demand real infrastructure skills. Call it straightforward for an individual developer, moderate for a team, and a project for an enterprise rollout.
Deployment
Integrations
Behind Refact.ai
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Refact.ai.
Frequently asked questions
Do I need a Refact account to use it?
Which models and providers can I use?
How is Refact.ai billed?
What is happening to Refact Cloud?
Does my source code leave my machine?
Which IDEs and languages are supported?
Can I self-host or deploy on-premise?
Is there a mobile app?
How do I ask for my data to be deleted?
Who is behind Refact.ai?
Should you pick Refact.ai?
Refact.ai is one of the more genuinely open propositions in AI-assisted development. It is open source, self-hostable by individuals as well as companies, indifferent to which model you point it at, and it works inside the IDE you already use rather than asking you to move house. The agent reaches beyond autocompletion into planning, execution and deployment, and the benchmark results the company publishes place it among the stronger open-source agents. For a team that has to keep source code on its own infrastructure, the on-premise option and the fine-tuning story are hard to match elsewhere.
The catch is timing. The commercial layer is being dismantled: the notice of 30 April 2026 retires hosted accounts, managed inference, credits, subscriptions and cloud team features, and the final date is still unannounced. Documentation has moved to the GitHub wiki of a community fork. Support will run through issues and Discord rather than a contract. Meanwhile the homepage still advertises a Pro plan at USD 10 a month, so the site tells two different stories at once, and the cost of running the agent now lands with whichever model provider you choose.
The legal documentation is thin by any standard — no GDPR statement, no data processing agreement, no subprocessor list, no retention period, no postal address, and privacy terms that predate the current product. That matters more for a tool that reads your entire repository than for most.
Judge it accordingly. If you want a managed service, look elsewhere for now. If you want a local engine you control, inspect and fine-tune, Refact.ai remains a serious option — provided you accept a community-maintained future and do your own compliance homework.
- Choosing a selection results in a full page refresh.
- Opens in a new window.