Refact.ai
Also known as: Refact, Refact.ai, refact.ai, SmallCloud AI, smallcloudai/refact
Open source autonomous coding agent, ranked first among open source agents on SWE-bench, that plans, edits, tests, and iterates inside your IDE, with self hosting and BYOK support.
Refact.ai is an open source autonomous coding agent that handles engineering tasks end to end from inside the editor. Created by SmallCloud AI, it plans a task, searches the repository for context, edits files, runs tests, and iterates until the work passes, rather than only suggesting the next line. It has ranked as the number one open source agent on the SWE-bench Verified leaderboard, autonomously solving 352 of 500 real world GitHub issues, with the full benchmark pipeline open source so results can be reproduced. In everyday use a developer describes a task in natural language and the agent carries it out.
The agent runs multi step and is capped at thirty autonomous actions per task, breaking work into steps, proposing and applying patches, running self tests, and revisiting earlier steps when something fails. It uses a Rust engine that keeps abstract syntax tree and vector indexes of the codebase current, and calls tools like symbol and semantic search and a debugging sub agent that runs and modifies scripts to find the cause of a failure. Because developers want oversight, Refact supports manual overrides: you can delegate a task and still preview and guide the process, review patches before they apply, and scope which files the agent can touch.
Refact runs in VS Code, JetBrains, Visual Studio, Neovim, and Sublime Text, pairing the agent with unlimited free code completion from a built in Qwen model and retrieval over the project. It integrates with GitHub, GitLab, Docker, PostgreSQL and MySQL databases, continuous integration pipelines, a browser tool, and the Model Context Protocol, so it can create a branch, implement an endpoint, run tests in a container, and open a pull request from one prompt. Models are open: bring your own key for OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter, or a local model, routed through an OpenAI compatible gateway at no extra platform cost.
Its strongest pitch is control. Refact is fully open source under a permissive license and can run entirely self hosted, on premises or in a private cloud, with zero telemetry so code never leaves the customer's servers, plus file level access controls, audit logs for data access, and fine tuning on the team's own codebase, a practical option for security conscious and regulated teams. Pricing is coin based: a free tier includes unlimited completions and agent access with a monthly coin allowance, paid plans start around ten dollars a month, and bringing your own key runs the agent at no coin cost beyond the provider's charges. The original open source repository is now a legacy archive, with development continuing in a successor repository.
Vendor details
Canonical URL
https://refact.ai
Category
Coding agent
Subcategory
Open source autonomous coding agent (self hosted, BYOK)
Funding status
Independent, created by SmallCloud AI, with Refact distributed as open source under a permissive BSD-3-Clause license. Specific venture funding is not disclosed in available sources. The project reports being the number one open source AI agent on the SWE-bench Verified leaderboard. Note that the original open source repository is now maintained as a legacy archive, with active development continuing in a successor repository, while the IDE products and pricing remain live.
Company status
independent
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
Integrates with GitHub and GitLab, PostgreSQL and MySQL databases, Docker, continuous integration pipelines, a browser tool, the Python debugger, and the Model Context Protocol, and runs as a plugin in VS Code, JetBrains, Visual Studio, Neovim, and Sublime Text. The agent accesses these resources and runs real operations autonomously, such as creating a branch, implementing an endpoint, running tests in a container, updating a database, and opening a pull request. It works with any large language model through bring your own key and an OpenAI compatible gateway.
In practice
You want an autonomous agent that finishes tasks, not just autocomplete. Refact plans, edits, runs tests, and iterates until it passes, capped at thirty actions per task, and ranks first among open source agents on SWE-bench.
You cannot let code leave your servers. Refact is open source and runs fully self hosted on premises or in a private cloud with zero telemetry, file level access controls, and audit logs for data access.
You want to control model cost and choice. Refact lets you bring your own key for OpenAI, Anthropic, Gemini, or a local model, running the full agent at no coin cost beyond your provider's charges.
Sources & related URLs
Related / legacy domains
Capability coverage
8.5 / 14 capabilities · 61%
| Integrations & Tool CallingIntegrates with GitHub, GitLab, PostgreSQL and MySQL databases, Docker, continuous integration pipelines, a browser tool, the Python debugger, and the Model Context Protocol, and acts by running real operations autonomously such as creating branches, running containers, updating databases, and opening pull requests, broad named integrations with genuine action. | Full |
|---|---|
| Workflow OrchestrationRuns a fully autonomous agent that plans a task, explores the repository, edits files, reproduces and fixes problems, and iterates until it passes, working multi step up to thirty actions with internal sub agents for planning and debugging, and can carry a task from prompt to a finished pull request, genuine end to end orchestration. | Full |
| Knowledge Grounding & RAGUses a Rust engine that keeps abstract syntax tree and vector database indexes of the codebase current and grounds the agent and completion in project context through retrieval, with symbol and semantic search tools and optional fine tuning on the team's own code, deep codebase understanding at the core of the product. | Full |
| Human Oversight & GuardrailsCaps the agent at thirty autonomous actions per task, lets developers delegate work while previewing and guiding the process, review and approve patch proposals before they apply, and scope which files the agent can touch, real oversight and control rather than a comprehensive runtime guardrail enforcement engine. | Partial |
| Security, Identity & GovernanceRuns fully self hosted on premises or in a private cloud with zero telemetry so code never leaves the customer's servers, and adds file level access controls, audit logs for data access, and no server side dataset storage, real data residency and access control, though without a public certification or single sign on matrix. | Partial |
| Observability & AuditabilityShows the agent's plan and steps in the editor for preview and guidance, runs self tests during a task, and provides audit logs for data access in self hosted deployments, real visibility into what the agent does, short of a comprehensive agent execution tracing and analytics platform. | Partial |
| Memory & State PersistenceLearns from interactions and feedback over time and organizes experience into a shared knowledge base, and grounds the agent in a codebase index and optional fine tuning, but these are feedback loop adaptation and knowledge grounding counted under other axes, not a distinct persistent agent memory or checkpoint product. | Unable to verify |
| Deployment & Data ResidencyIs open source under a permissive license and can be deployed on premises, as a hosted service, or on AWS, running entirely on the customer's own infrastructure or private cloud with a free local completion model, a genuine self host and data residency capability that keeps code inside the customer's boundary. | Full |
| Prebuilt Agents, Templates & PacksShips a single autonomous agent that teams customize around their tools, workflow, and coding style and fine tune on their code, which is configuration and customization of one agent, not a shipped library of prebuilt agents, templates, or packs, nor multi agent scaffolding to browse and remix. | Unable to verify |
| Triggers & Channel CoverageRuns as a plugin across VS Code, JetBrains, Visual Studio, Neovim, and Sublime Text, and is invoked through in editor chat and agent commands, real multi editor channel coverage that is developer invoked within the coding workflow rather than driven by broad external event triggers. | Partial |
| Model Flexibility & RoutingLets users bring their own key for OpenAI, Anthropic, Gemini, Grok, DeepSeek, OpenRouter, or a local model, ships its own completion models including a built in Qwen coder, and routes through an OpenAI compatible gateway that reaches many providers, genuine multi provider model flexibility and routing. | Full |
| APIs, SDKs & MCP ExtensibilityIs fully open source with a Rust agent that runs as a language server, exposes agentic tools for models to call, and supports the Model Context Protocol, a strong and extensible surface for building on and integrating the agent, though a dedicated public software development kit is not clearly documented as first class. | Partial |
| Testing, Debugging & OptimizationReproduces problems, runs a codebase's existing tests to avoid regressions, and uses a debugging sub agent that drives the Python debugger to modify and run scripts and find a fault's cause, iterating on failures, a strong test gated and self correcting loop rather than a standalone testing or quality product. | Partial |
| Browser & Computer UseExecutes shell commands, runs code and tests inside Docker containers, drives a browser tool to operate web applications, and reads and writes files in the workspace, real code execution and computer use within the development environment, though not general autonomous browser automation as the core product. | Partial |
Pricing
Free tier includes unlimited code completions, core agent access, and BYOK. Paid plans start around ten dollars a month. Extra coins cost one dollar per thousand with a five dollar minimum and never expire. Self hosting the open source edition is free.
Coin based usage plus a paid subscription. A free monthly coin allowance and unlimited completions come at no cost. Paid plans lift limits from around ten dollars a month, and extra coins cost one dollar per thousand with a five dollar minimum. Bringing your own model key runs the agent at no coin cost, so you pay only your provider.
Included quota
Free: unlimited code completions, core AI agent access, bring your own key support, and a monthly allowance of five thousand coins. Pro: higher limits and more coins from around ten dollars a month, with extra coins available at one dollar per thousand and a five dollar minimum. Coins never expire. The open source edition can be self hosted at no license cost, and bringing your own model key runs the full agent at no coin cost beyond the provider's charges. Enterprise adds on premises deployment, fine tuning, and custom terms.
What is public
Public: the free tier contents, the coin exchange rate and minimum, the Pro entry price, bring your own key at no coin cost, and the free open source self hosting option. Not public: exact per task coin usage and enterprise pricing.
Billing mechanics
A free tier plus a Pro subscription from around ten dollars a month, metered by coins for agent activity, with extra coins purchasable at one dollar per thousand. Bring your own key removes coin cost. The open source edition is free to self host.
Cost watchouts
The coin model meters agent activity, so autonomous multi step tasks consume coins faster than simple prompts, and deeper reasoning options are described as coin expensive. Bringing your own model key removes coin cost but shifts spend to the provider, billed by usage. Self hosting is free but you run and pay for your own infrastructure and any external model calls.
Variable cost rationale
The coin model explicitly charges for what the agent does, so cost scales directly with autonomous activity and multi step tasks, and deeper reasoning options are noted as coin expensive. Bringing your own model key removes coins but shifts to usage based provider billing, and self hosting shifts cost to infrastructure and model calls. In all paths, spend tracks usage rather than a flat seat, so exposure is high.
Additional watchouts
Coin spend scales with how much autonomous work the agent does, and reasoning heavy options cost more, so heavy users should model usage. Enterprise and self hosted deployments add infrastructure and model costs. The original open source repository is now a legacy archive.
Overage / add-ons
When the monthly coin allowance runs out, users buy more coins at one dollar per thousand with a five dollar minimum, and coins never expire. Bringing your own model key avoids coin consumption, with usage billed by the provider instead.
Sales call required
No — self-serve available
Free / trial
A genuinely free tier includes unlimited code completions, core AI agent access, a monthly allowance of five thousand coins, and bring your own key support. The open source edition is also free to self host.
Lowest paid plan
Pro from around ten dollars a month, above a free tier. Additional coins can be bought at one dollar per thousand with a five dollar minimum.
Commercial notes
Refact is priced to be one of the cheaper entries in the category, with a real free tier, low cost paid plans, and a self host path that is free apart from infrastructure and model calls. The coin model aligns cost with actual agent work, which favors light users and bring your own key setups, while heavy autonomous use raises coin spend.
Key ambiguities
How many coins a typical agent task consumes is not fixed, so real monthly cost depends on how heavily the autonomous agent is used and whether reasoning heavy options are called. Bring your own key removes the coin variable but replaces it with provider usage billing. Enterprise pricing is not public.
Cancellation / refund
Self serve plans can be changed or cancelled by the user, and purchased coins never expire. The open source self hosted edition carries no subscription. Enterprise terms are arranged directly.
Missing data
Exact coin consumption per task, Pro plan coin quotas above the free allowance, and enterprise pricing are not fully published.
Related vendors
- Cognition — Maker of Devin, an autonomous AI software engineer
- Aider — Open-source, model-agnostic terminal coding agent that works…
- Anthropic Claude Code — Anthropic's terminal and IDE-native coding agent with sub-agent…
- Augment Code — AI coding assistant evolving into an AI-native engineering platform
- Blackbox AI — AI coding assistant for code generation and developer workflows
- Blink — Agentic AI app builder that turns a natural language prompt into a…