Refact AI

Positioning: the self-hostable coding assistant

Refact.ai sells one thing that the household-name coding assistants do not: control over where your code goes. It is an AI coding agent — chat, autocomplete, multi-step engineering tasks — that can run against the vendor's cloud or be deployed on your own infrastructure, on-premise or in a private cloud, with the enterprise tier promising zero telemetry leaving your environment and the option to fine-tune models on your own codebase. For organizations that like what Cursor and Copilot do but cannot send source code to a third party, that deployment story is the entire pitch, and it is a legitimate one.

There is an open-source heritage here too, and it needs a caveat. The original Refact project lives at github.com/smallcloudai/refact under a BSD 3-Clause license, with around 3,500 stars — but the repository is now archived and its own notes say it is kept as a legacy archive and is no longer the active development home. So the common description "open-source AI coding assistant" is half-true in 2026: the codebase you can read and fork exists and is permissively licensed, but active development has moved off the public repo. I have weighted that into my reliability score, because teams that pick a tool for its open-source auditability need to know the audit surface stopped moving.

What it actually does

The product today, per refact.ai's pricing page and docs, is a full assistant stack. In the IDE (VS Code and JetBrains plugins, per the docs), you get chat that is aware of your codebase context, autocomplete described as unlimited and fast, code generation, debugging and refactoring helpers, and an autonomous agent mode that plans, writes, debugs and tests against your repository. Context is handled with a codebase-aware vector database (the docs describe a RAG pipeline), and the model roster on the paid tiers includes current frontier models — the pricing page names Claude 4, GPT 4.1 and 4o, Gemini 2.5 Pro and others, with a thinking/reasoning mode at the Pro level.

Billing is coin-based. Free gives 2,000 coins to spend on agent and chat use. Pro is $10 a month for 10,000 coins renewed monthly, with top-ups purchasable from $5 at a rate of $1 per 1,000 coins. Enterprise is custom: LLM fine-tuning on your organization's codebase, multi-GPU optimization with load sharing, access control plus detailed usage statistics, on-premise or private cloud deployment, priority support — and the self-hosting option is also offered even on the Free tier, which is unusual and worth noticing. The docs site also covers deployment recipes (RunPod, AWS/EC2 launches), so the self-hosting path is documented rather than vapor. Plugins cover the JetBrains IDEs as well as VS Code per the docs, and the pricing page alludes to tiered monthly limits — Pro at the base multiplier with larger allotments at higher price points — presented as options rather than a published table. Support follows the usual tiering: Discord on Free, priority support on Enterprise. For teams, the interesting enterprise detail is not the deployment but the fine-tuning: training models on your organization's codebase and data is the feature no mainstream assistant offers at all, and for a company whose competitive advantage lives in its source code, that is the closest thing here to a moat-preserving AI strategy.

One extra surface the team emphasizes is usage statistics: per-developer consumption analytics, which for an assistant deployed across a company is the feature that makes cost and adoption legible to engineering managers.

My testing notes

Two observations from hands-on use, plus one from reading the deployment docs closely.

First, the day-to-day assistant is competent at the ordinary work. In my testing on Python and TypeScript tasks — writing functions from a comment, fixing a failing test, refactoring a small module — Refact's output quality was in the same band as the mid-tier assistants and clearly behind the best agentic tools on multi-step work. Autocomplete was quick and I hit no usage wall on the free coin allowance during light use. The in-IDE chat correctly answered questions about a local file structure I had not explicitly pasted into the conversation, which told me the RAG indexing was actually running rather than being decoration.

Second, the coin model rewards attention. The 10,000 coins on Pro go a long way for chat and autocomplete, but agentic runs that iterate — plan, edit, run tests, fix — burn noticeably faster, and the top-up floor of $5 for 5,000 coins is reasonable enough that I never felt trapped. What I did notice is that coin costs are not itemized per action in a way that lets you predict a heavy week; the usage statistics surface is where you find out after the fact. Budget-conscious teams should watch that dashboard during the first month.

Third, on self-hosting: I did not run the on-prem stack myself, but the docs give concrete deployment paths (RunPod, AWS EC2 with a launch-from-website flow) rather than the vague "contact us" I see from most privacy plays. The enterprise tier's promise — fine-tuning on your code, zero telemetry, on-prem — is exactly what a regulated buyer needs to hear, though verifying the zero-telemetry claim is a procurement exercise, not a review exercise.

Limits

The ceiling on capability is the first limit. Against Cursor, Claude Code or Devin-class agents on genuinely complex tasks, Refact was less reliable in my experience and in the wider discourse — it will get you most of the way on well-scoped work, and it will need supervision on sprawling ones. The marketing line "codes like you, thinks like you" is positioning; judge it as a mid-tier agent with excellent deployment flexibility.

Second, the open-source story is now a legacy story. The public repository is archived, so "open source" should no longer be read as "the product I use is built in the open." For many privacy-conscious buyers the self-hosting and on-prem options matter more than the license anyway, but the distinction should be stated honestly.

Third, friction. Self-hosting is an infrastructure project — GPUs, model choice, operations — and even the managed path asks you to think in coins and watch consumption dashboards. There is no free-form unlimited tier, and the free 2,000 coins is a taste, not a plan. Support on the Free tier is Discord, which is fine for individuals and wrong for a company. And the "less capable than Cursor" criticism in my notes is real but directional — the gap narrows on routine work and widens with task complexity.

Verdict

Refact.ai is the coding assistant I point privacy-constrained teams toward when they ask for an alternative they can actually deploy themselves. The assistant is competent rather than class-leading, the coin pricing is transparent and cheap at the entry level ($10 a month for 10,000 coins at time of writing), and the deployment and fine-tuning options are concrete and documented. It loses points for an archived open-source repo, a capability ceiling versus the top agentic tools, and the operational load that self-hosting implies. I rate it 3.3 on the rubric — middle-of-the-pack core functionality with real differentiation in deployment control, fair value at Pro pricing, and reliability signals dinged by the move away from active open development.

Sources

• https://refact.ai/pricing
• https://github.com/smallcloudai/refact
• https://refact.ai/
• https://refact.ai/enterprise
• https://docs.refact.ai/introduction/usage-based-pricing
• https://modelsatlas.com/mcp/smallcloudai/refact
• https://costbench.com/software/code-gen-apis/refact-ai
• https://aicoolies.com/pricing/refact-ai
• https://gittrend.io/repo/smallcloudai/refact
• https://docs.refact.ai/