Integrations That Write Themselves
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
Ramp built an agentic system that autonomously builds and maintains third-party API integrations as customers request them, rather than relying on engineers to build each one. The agent researches a provider's API, writes and tests 'recipes' (single authenticated calls) chained into 'recipe books', asks for credentials via a secure widget that keeps them out of LLM context, and self-repairs when providers change their APIs. A companion 'integration factory' takes the demand signal from these customer-built integrations and produces first-party connectors autonomously, opening pull requests with evidence (API calls, responses, screen recordings) for human review instead of requiring engineers to trust the generated code directly. Security design assumes customer-supplied URLs are hostile: HTTPS-only, hostname allowlisting, no access to internal address space, immutable per-business recipes, and isolated stateless egress. Ramp reports 75 integrations shipped this way, cutting build time from weeks/months to hours or minutes and cost to under $25 per integration.
Table of contents
Users Building Their Own Custom IntegrationsOur Integration FactorySecurityImpactLearningsVisionThank You!Questions this post answers
How does Ramp prevent API credentials from being exposed to the LLM when building integrations automatically?
Credentials never enter the agent's context. Customers submit API keys through a secure widget, and when Ramp's internal factory needs test credentials it sends a secure link instead; either way the values go directly into encrypted storage and are only injected during runtime execution, making them unreachable for the LLM itself. Developers wiring AI agents into sensitive workflows can find more credential-safety patterns like this on daily.dev.
What security checks stop a customer-supplied integration URL from being abused for server-side request forgery (SSRF)?
Ramp requires the destination URL to use HTTPS and match an allowlisted hostname exactly or as a proper subdomain, resolves the hostname to reject private or internal address space, and fixes the destination permanently on the recipe at creation time so runtime input can only fill in body and query params, never redirect the call elsewhere. Teams hardening agent-driven API calls against SSRF can track patterns like this via daily.dev.
How much does it cost to build a new integration using Ramp's autonomous integration factory?
Ramp's internal integration factory can research, build, test, and open a pull request for a new first-party integration for under $25 and a short human review, compared to weeks of engineering time previously required. One example, a Pangram connector for AI detection, cost under $15 and was merged within hours. Anyone evaluating agentic build pipelines against manual integration work can follow real cost comparisons on daily.dev.