Refact.ai

Documented integrations span GitHub and GitLab for source control, PostgreSQL and MySQL databases, Docker containers, shell commands, the Python debugger, browser tools, and MCP servers the agent consumes, with CI/CD pipelines connected; the vendor describes the agent connecting to a customer's GitHub, Docker and PostgreSQL to mimic developer behavior and complete tasks autonomously, such as creating a branch, implementing an endpoint, running tests in a container, updating a database and opening a pull request from a single prompt. Support covers 25+ programming languages.

Sourceplugins.jetbrains.com Refact listing, refact.ai and github.com/smallcloudairead 2026-08-30

The agent plans a task, searches and analyses the repository for context, edits files, reproduces problems, runs tests and iterates until the work passes, executing up to thirty autonomous actions per task and revisiting earlier steps on failure, with internal sub-agents for planning and for debugging that drive the Python debugger to modify and run scripts; it carries a task from a single prompt through branch creation, implementation, containerized test runs and an opened pull request. The vendor reports ranking first among open-source agents on SWE-bench Verified, autonomously resolving 352 of 500 real GitHub issues, with the benchmark pipeline published as open source.

Sourcegithub.com/smallcloudai, refact.ai and refact.ai/blog/2025/open-source-sota-on-swe-bench-verified-refact-airead 2026-08-30

A Rust engine runs inside the IDE as a language server and keeps abstract syntax tree and vector database indexes of the codebase current, exposing symbol and semantic search as tools the model calls, with retrieval-augmented generation grounding both agent work and completion in project context; the vendor states the agent understands company context and standards by analyzing documentation and codebase, and enterprise deployments can fine-tune models on the organization's own code. Context sources named on the product page span workspace, codebase, databases, files, documentation and web.

Sourcegithub.com/smallcloudai and refact.airead 2026-08-30

Tool execution is mediated by confirm rules in each mode's configuration: safe reads default to auto and destructive or sensitive operations to ask, and when a tool requires confirmation execution pauses until the user approves or rejects it; patch, create, update, remove, move and undo style operations require confirmation. Shell commands and integrations carry allow, ask and deny rules, and the agent is capped at thirty autonomous actions per task. The documentation is the project wiki that docs.refact.ai redirects to.

Sourcegithub.com/JegernOUTT/refact/wiki Agent-Tools, Integrations-Shell-Commands and Integrationsread 2026-09-29

Access can be restricted to particular files or projects so private or confidential code is withheld from the assistant, the vendor states it never stores code server side and documents privacy settings, self-hosted deployment runs on premise or in a private cloud with zero telemetry, and audit logging for data access is documented for self-hosted deployments; after the announced retirement of hosted Refact Cloud there is no vendor-side service retaining customer code. No security attestation or certification, SSO, SAML or organization-level role-based access control is documented.

Sourcedocs.refact.ai, refact.ai and refact.ai/blog/2026/refact-cloud-is-shutting-downread 2026-08-30

Each chat's trajectory, the chronological agent and chat history including tool execution, is persisted locally under.refact/trajectories and can be searched by the agent, reasoning is shown step by step in the editor, and proposed patches are inspectable before they apply; self-hosted deployments document audit logs for data access. These are local session records on the developer's machine, with no audit log of agent actions, telemetry export or run analytics for a team.

Sourcegithub.com/JegernOUTT/refact/wiki Chat-System and Memory-and-Knowledge, and refact.airead 2026-09-29

Durable project memory lives in the project-scoped.refact/ directory: typed task memories (decision, spec, gotcha, risk, handoff, progress, postmortem, brief) preserve structured task knowledge across sessions for project-scoped recall, and a knowledge graph of facts and relationships has staleness tracking and a cleanup pass that removes stale or duplicate nodes. Scope and a purge path are stated; the files sit in the project on disk. The VecDB and AST indexes are separate from this memory and ground the agent in the codebase.

Sourcegithub.com/JegernOUTT/refact/wiki Memory-and-Knowledge, reached through the docs.refact.ai redirectread 2026-09-29

The product is open source under a permissive license and local-first by design, running from the IDE with a local Refact engine and whichever providers or local runtimes the customer configures, with no hosted account or managed inference service required; on-premise and private cloud deployment keep code on the customer's own servers with zero telemetry, and local runtimes including Ollama, LM Studio and vLLM are supported. The vendor announced on 30 April 2026 that hosted Refact Cloud is being retired, stating the project continues as open source, local-first, BYOK and community maintained, and that local configuration and project data remain under the customer's control with requests sent only to configured providers and integrations.

Sourcerefact.ai/blog/2026/refact-cloud-is-shutting-down and refact.airead 2026-08-30

An extensions marketplace discovers and installs three kinds of item, skills, commands and subagents, from built-in embedded sources, built-in GitHub sources and user-added GitHub sources, with the source index bundled in the engine. The catalog includes ready subagents a customer can install.

Sourcegithub.com/JegernOUTT/refact/wiki Marketplace, reached through the docs.refact.ai redirectread 2026-09-29

The engine's scheduler runs cron-style scheduled tasks that fire into the agent: durable jobs persist per project in.refact/scheduled_tasks.json and survive restarts, use 5-field cron expressions with jitter, catch up missed one-shot runs and expire recurring jobs after 30 days by default. Invocation otherwise spans VS Code, JetBrains, Visual Studio, Neovim and Sublime Text plugins, a standalone CLI and TUI, and an in-browser GUI served by the local daemon.

Sourcegithub.com/JegernOUTT/refact/wiki Scheduler-and-Cron and Home, reached through the docs.refact.ai redirectread 2026-09-29

The agent runs on bring-your-own-key or local model providers configured on the customer's own machine, with at least one provider required in Provider Setup before use; named BYOK providers are OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter, Groq and xAI, and named local runtimes are Ollama, LM Studio, vLLM and custom OpenAI-compatible endpoints, with default models selectable per surface. Code completion ships on a bundled Qwen2.5-Coder model. Following the retirement of Refact-issued model credits announced 30 April 2026, BYOK or a local runtime is the only path to running the agent.

Sourceplugins.jetbrains.com Refact listing and refact.ai/blog/2026/refact-cloud-is-shutting-downread 2026-08-30

The local engine exposes a documented v1 HTTP API covering chat, code, tools, AST, retrieval, git, checkpoints, integrations, knowledge and voice, and the GUI and IDE plugins are thin clients of that daemon, so other programs can drive Refact the same way. The engine and plugins are open source, and MCP servers the customer configures extend the agent.

Sourcegithub.com/JegernOUTT/refact/wiki HTTP-API and Home, reached through the docs.refact.ai redirectread 2026-09-29

The agent reproduces a reported problem, runs the codebase's existing tests to avoid regressions, and iterates when they fail, with a dedicated debugging sub-agent that drives the Python debugger to modify and run scripts and isolate the cause of a fault; a standalone AI Code Review tool analyses code and proposes fixes.

The vendor publishes SWE-bench Verified results with the full benchmark pipeline open sourced for reproducibility. Running the project's tests verifies the agent's own output, the SWE-bench pipeline is the vendor's own evaluation rather than a customer-facing harness, and no evals, regression testing or agent-behavior debugging surface is exposed to customers.

Sourcerefact.ai/blog/2025/open-source-sota-on-swe-bench-verified-refact-ai and refact.airead 2026-08-30

Refact includes a built-in Chrome tool: the agent launches Chrome or Chromium, connects to an existing DevTools endpoint or reuses a session for the chat, opens and switches tabs, navigates, inspects and interacts with pages, extracts data and collects screenshots, for debugging web apps and checking UI behavior. The homepage carries a customer account of the agent operating a WordPress admin interface to fix a plugin fault. Shell, Docker and the Python debugger are programmatic routes rather than browser use.

Sourcegithub.com/JegernOUTT/refact/wiki Browser-Automation and refact.airead 2026-09-29