Recurse

Recurse is a platform for focused AI applications that can run as standalone tasks or inside clients such as Codex, Claude, and ChatGPT. Its catalog packages carefully designed workflows into small tools for jobs including invoice generation, keyword research, podcast creation, image conversion, email validation, resume building, and HubSpot cleanup. Rather than asking one general assistant to improvise every process, a Recurse app defines the job and the expected output. A single account and wallet work across the catalog, so people can choose a narrow app when a project needs a PDF, audio file, image, dataset, or other concrete result. The same tool can also continue inside a conversation. A user can review the first output, ask for a narrower result, and turn the findings into another project asset without changing applications. That combination of structured work and conversational follow-up is especially useful when a deliverable needs several rounds of judgment rather than one unattended run.

The platform also lets developers build and publish their own agent applications. A typical project contains an agent configuration, a prompt, Python tools, and its dependency files. The configuration describes input fields, optional defaults, the output schema, and the tools the agent may call. Recurse validates successful results against that declared output shape. A coding agent can read the Recurse skill, create these files, test the project with sample inputs, and prepare it for deployment. This favors specialist workflows that can be inspected and repeated over one-off prompts whose structure lives only in a chat transcript.

Deployment uses a command line interface. The developer installs the Recurse SDK, runs a project locally with sample input, and deploys it as an MCP service. Recurse supplies the hosted runtime for the agent and its tools, while the developer can choose more CPU or memory when an application needs it. Once deployed, the MCP connection makes the app callable from a supported AI client, so a user can request the task, inspect the result, and ask follow-up questions without leaving the conversation where the surrounding work already lives.

Recurse can attach secrets to a deployment without placing them directly in project files. A developer saves a key under a name, then maps it to an environment variable during deployment. The official workflow shows this pattern for services such as GitHub and lists a broad set of potential systems, including Slack, Stripe, Notion, Shopify, HubSpot, Gmail, Google Drive, OpenAI, Anthropic, Linear, Salesforce, databases, monitoring tools, and deployment platforms. These names indicate services an agent can be equipped to use, though each app still needs the corresponding tool code and permission.

For people who don't want to build an app, the public gallery is the easier entry point. Recurse's own tools include a PDF invoice generator from job notes, keyword research that compares volume and ranking difficulty, text to podcast output with an MP3 and transcript, PNG to SVG conversion, email validation, and duplicate HubSpot contact review. The common thread is a defined deliverable rather than an open-ended chat. Quality still depends on the app and its inputs, but the narrower workflow gives users a clearer basis for judging whether the result is ready to keep.

The product fits developers who want to turn an agent workflow into a reusable MCP service, as well as teams that want to call specialist tools from an AI client they already use. It is less suited to someone looking for a no-cost general assistant or a conventional visual automation canvas. Custom app creation involves project files, Python tools, a CLI, testing, and an understanding of inputs and outputs. The platform handles runtime and connection concerns, but it doesn't remove the need to design a good prompt, secure external credentials, check sample results, and decide what compute a task needs.

Pricing is based on use rather than a subscription. Gallery users can start with $1 in a shared Recurse wallet, and model plus compute costs are deducted when an app runs. For custom agents, Recurse publishes zero markup on model input and output tokens and a secure execution rate of $0.013 per active minute. It says there are no idle charges, while non-preemptible direct runs have three times the CPU and memory charges. This model makes short experiments accessible without a separate plan for each tool, though costs vary with model choice, runtime, and resources, so frequent or computationally heavy use requires closer tracking than a fixed monthly bill. Only active requests are billed for serverless execution, which prevents a deployed but idle specialist app from accumulating runtime minutes.