The archive · AI & Models · Product decision · 2025-2026
Wildcard AI bets agents.json, OpenAPI for LLMs, becomes how agents call APIs
Wildcard AI bets agents.json, a stateless OpenAPI layer for LLMs, becomes the standard; Show HN Mar 2025, 1.3k stars, YC-backed
Wildcard AI
What the business is
Wildcard AI builds developer infrastructure for agent-API integration: the open-source agents.json specification (an OpenAPI layer that describes tools, flows and links for LLMs), a Bridge Python package that loads and runs the contracts, and a public registry of agents.json files; its stated plan is to charge API providers for white-glove onboarding onto the standard.
How it started
Wildcard AI started agents.json as an internal config file while building an agent with access to Gmail: putting the full API spec into the LLM's context gave poor accuracy in tool selection and tool calling, even with the output space cut to 5-10 endpoints, and the team concluded APIs are designed for developers, not LLMs. The file grew features like auth registration and tool search; when HubSpot CTO Dharmesh posted publicly about the concept of a specification that translates APIs for LLMs, Wildcard decided to open-source what it already had working. On the launch thread the founders described going through Y Combinator with a tool-search product, and the Show HN for agents.json went up on 2025-03-03.
What happened
The Show HN drew 212 points and 69 comments on 2025-03-03. The thread debated how agents.json differs from Anthropic's Model Context Protocol - stateless contracts any client can run versus stateful per-client connections - where it overlaps OpenAPI's Arazzo workflow spec, and how to license the tooling; Wildcard replied that its main opportunity is charging API providers for white-glove onboarding, kept the spec under Apache 2.0 and the Bridge package under AGPL, and said the long-term goal was to align with Arazzo. The launch post counted 10 vetted API integrations, some official, with Resend, Google Sheets, Slack and Stripe demos. By the 2026-09-05 snapshot the repository had 1.3k stars, 66 forks and 75 commits, a registry of available agents.json files, quickstart notebooks, and a roadmap still listing OAuth, loops, conditionals and failure handling as not yet built.
No ending yet — it is still running.
Background
Wildcard AI's agents.json began as an internal config file while the team built an agent with access to Gmail. Putting the full API spec into the LLM's context produced poor accuracy in tool selection and tool calling, even with the output space cut to 5-10 endpoints, and the team concluded that APIs are designed for developers, not LLMs - so it started describing how API calls should be chained into outcome-based tools, and used the file internally while adding auth registration and tool search.
Wildcard open-sourced agents.json 0.1.0 after HubSpot CTO Dharmesh posted publicly about the concept of a specification that translates APIs for LLMs - sounding like what the team already had working internally. The Show HN went up on 2025-03-03 and drew 212 points and 69 comments, mostly debating the spec's relationship to the stateful Model Context Protocol and to OpenAPI's Arazzo workflow spec; Wildcard said its plan was to charge API providers for white-glove onboarding onto the standard.
By the 2026-09-05 snapshot the repository had 1.3k stars, 66 forks and 75 commits, plus a registry of available agents.json files and quickstart notebooks for APIs such as Resend and Rootly. The spec remains at version 0.1.0, with OAuth, loops, conditionals and failure handling still on the roadmap, and the public materials reviewed here disclose no funding or revenue figures.
What has to be true
- The team's own Gmail agent proved that raw OpenAPI in context failed at tool selection even with 5-10 endpoints, so the problem was concrete rather than hypothetical
- OpenAI's Operator signaled a shift toward agents acting on the web, and APIs already run on backend infrastructure built for scale - safer ground than browser automation
- MCP's stateful per-client connections looked wrong for web products, leaving room for a stateless contract that could deploy on serverless infrastructure
- OpenAPI gave agents.json free distribution: providers already maintain a spec, adding one file beside it costs little, and the registry could grow by word of mouth
What can be applied
A standard wins by making adoption cheaper than the status quo: build on specs providers already maintain, stay stateless so it runs anywhere, and monetize the on-ramp, not the protocol
Aftermath
As of 2026-09-05 agents.json is still version 0.1.0 and explicitly a work in progress seeking community input, with the repository at 1.3k stars and a roadmap that lists OAuth support, loops, conditionals and failure handling as unfinished. Wildcard AI still maintains the spec, the Bridge package and the public registry, and the public materials reviewed here disclose no funding or revenue figures.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card