Pomerado

The repository is an npm package you run on Node 24.21 or later. You ask for an integration either from the command line or from your coding agent: Codex, Claude Code or Cursor can call the pomerado tools directly, so the agent describes what it needs and oversees the build while Pomerado does the browser work. During minting, Pomerado signs in (handling saved credentials and 2FA codes), maps the filters and fields it needs, runs the task once for real, and asks clarifying questions when the site is ambiguous, for example which of several customers an invoice list should cover, or whether to include a comment thread. Guardian reviews each action and the finished source as it goes.

What comes out is a tested integration exposing one or more tools, published either to your own machine through the standalone host or to Pomerado Cloud. Agents then call it over MCP or REST. The hosted service is where the self-healing runs: when a call fails because the site changed, maintenance reruns the failing step, works out what moved, patches that step, tests it and publishes the fix, then completes the original call so the agent gets its result rather than an error.

• •Mint From a Description - An integration is built from a natural-language description of the task, with no HAR capture, no recorded session and no selectors written by hand.
• •Deterministic, Not Reasoned - The output is source code that runs the same steps every time, which is why it is faster and cheaper per call than a browser agent re-reading the page.
• •Works Behind Logins and 2FA - Sign-in is handled during both minting and running, including sites that require two-factor codes or one-time passwords.
• •Guardian Review - Each browser action and the finished integration source are reviewed during minting rather than trusted blindly.
• •Self-Hostable MIT Core - The minter, harness, prompts, Guardian and a standalone MCP host are all in the MIT repository, and run with no Pomerado account.
• •Built From Your Coding Agent - Codex, Claude Code and Cursor can call Pomerado's own tools to find, build and run integrations, with the agent answering the minter's clarifying questions.
• •Self-Healing on Cloud - The hosted service detects a broken integration, repairs it and finishes the original call. This is a Cloud capability, not part of the open source core.

Teams wiring agents into systems that have no public API: EMRs, payer and prior-authorisation portals, ERPs, procurement and vendor portals, HR and benefits systems, government filings and internal tools. It is the direct counter-position to the browser agents already in this directory, and the comparison is worth making explicitly: if an agent performs the same task against the same site many times a day, compiling the interaction once is a better trade than reasoning through it every run. The free tier at 100 tool calls a day and 10 new integrations a month is enough to test that claim on a site you care about.

Treat the benchmark on the homepage with caution. The claims of 2x faster, 10 to 1,000 times cheaper and 30% more accurate come from 30 tasks across 25 business apps with one attempt each, run by the vendor, with no published harness; they are the vendor's own numbers, not an independent result. The bigger caveat is age. The public repository was created on 5 October 2026, so the project is weeks old, and 1,454 stars is the smallest traction signal among the open source tools in this category. Nobody yet knows how well self-healing holds up over a year on sites that change often, and a site operator who objects to automated access is a risk this approach does not remove. If you need something settled and proven in production, wait and watch this one.