The customizable workbench for putting AI agents to work.
Build software, research decisions, automate routines, and take on whatever comes next. Choose the agent and model for each job, connect the tools you already use, and run Recursive on hardware you control—then review the results from anywhere.
Coming soon for Windows & Linux
FAQs
Everything you need to put agents to work, from first request to reviewed result.
What is Recursive?
Recursive is a customizable workbench for putting AI agents to work. Start with software, then extend it to research, analysis, documents, operations, and whatever your tools make possible. Describe the outcome, choose the agent and model, and keep the plan, progress, decisions, and result together.
Bring Claude, ChatGPT, Gemini, Codex, Cursor CLI, OpenCode, local models through Ollama, or any OpenRouter model. Recursive gives each session the tools and context you configure while keeping your workbench and project data on hardware you control. No telemetry by default; remote access is end-to-end encrypted, and sync uses storage or Recursive instances you configure.
Recursive is for people who know the work — engineers, founders, analysts, designers, operators, and domain experts — and want agents to carry whole jobs forward without giving up direction or review. If you can describe an outcome and judge the result, you can direct an agent session.
- Agent sessions delegate whole jobs to durable agents while choosing the model, tools, permissions, and level of oversight.
- Tasks & plans turn ideas into tracked work with clear plans that you or another agent can execute.
- Routines run recurring agent work on a schedule or in response to events from connected tools.
- Plugins & MCP adapt the workbench with tools, skills, rules, interfaces, and Model Context Protocol servers.
- Run anywhere run the workbench on hardware you control, then reach it from desktop, mobile, or the web.
How does Recursive turn a request into reviewed work?
From request to reviewed result in four steps:
- Describe the outcome. Open a session with a feature, research question, document, operational task, or rough idea. Give the agent the context that matters.
- Choose how it runs. Pick the agent and model, connect the tools it can use, and set the permissions and review boundaries for the job.
- Let the agent work. Recursive keeps the plan, tool activity, questions, and progress visible while the session runs on hardware you control.
- Review the result. Inspect what changed, answer decisions in context, request another pass, or turn the next step into work for another agent.
Getting started
Is Recursive free?
Yes. The complete local product is free to run on your machine, with no account or card required. Pro is optional: $12 per month, or $99 per year (save 30%), for remote access and multi-device sync. See the pricing page for details or api/pricing.json for machine-readable pricing.
Which agents and models can I choose?
You choose the agent and model for each session or routine; Recursive does not automatically route work between models. Connect Claude, ChatGPT/Codex, Gemini, Grok, Kimi, GLM, OpenRouter, local models through Ollama or Osaurus, or another provider through a pluggable adapter.
What platforms are supported?
The packaged desktop app is available for macOS today (Apple silicon and Intel), with secure browser and mobile access through Pro. Windows, Linux, and server deployments are planned as Recursive expands onto more hardware you control. See the download page for the current build.
Privacy and control
What leaves the hardware running Recursive?
Recursive sends no telemetry by default, and anonymous diagnostics are opt-in. Workspace data leaves only when you send work to a connected model or tool, enable remote access, or configure sync. Remote access is end-to-end encrypted; sync uses storage or Recursive instances you choose.
Where does Recursive run?
On hardware you control. Today that starts with the macOS desktop app; Windows, Linux, and server deployments are planned. You choose the provider for each session, so only the requests you send to a connected AI service leave that hardware.
How is this different from Devin or Cursor?
Recursive is a plugin-first workspace for directing and reviewing agent work, not just an AI coding surface. Its privacy-first design runs locally by default and keeps your work under your control. It works with the models and providers you choose, then extends beyond coding through plugins, tools, automations, and channels.
Extending Recursive
Can I extend it?
Yes. Plugins (TypeScript + a manifest) add tools, skills, rules, themes, triggers, and channel integrations; anything speaking Model Context Protocol plugs in too. Start at the developer portal.
Is there an API?
Yes — every product operation is available over HTTP and documented in the OpenAPI specification, exposed through an embedded MCP server, and scriptable with the bundled recursive CLI. Start at the developer portal. Responses carry standard rate-limit headers, and the API follows a predictable URL versioning policy with six-month deprecation windows documented in the spec.