Buy Me a Coffee
Card payment in any amount, one time or monthly.
Support APIForge
APIForge is free and built by one person. Its AI tools run on free-tier models, which come with a request limit and answers you should check. Funding pays for better models, higher limits and faster responses.
Why there is an AI limit
The assistant, payload generator, test cases, TypeScript generation and AI-prompt generation run on free-tier language models (Groq-hosted open models today). Free tiers cap how many requests a service can send, so each visitor gets 40 AI requests per rolling 5 hours, counted across all of those tools.
The rule-based quality score, the parser, the proxy and Roast My API's scoring are not AI and do not count toward the 40. Roast text uses AI optionally and falls back to a deterministic version.
Free models are slower under load and less accurate than frontier models. They can miss edge cases in a schema, invent details that look plausible but are wrong, or refuse. Check generated types and tests before you ship them.
Illustration
AI requests left34 / 40
At the limit0 / 40
AI buttons stop until older requests age out of the 5-hour window. Scoring, parsing and the request proxy keep working.
What support pays for
- More accurate models
- Frontier-class models miss fewer edge cases in schemas and invent fewer details, so generated types and tests need less checking.
- Higher limits
- Paid model access comes with higher request limits, which is what allows the 40 requests per 5 hours rule to be raised.
- Faster responses
- Paid capacity is not shared with every free-tier user, so answers arrive sooner when traffic is high.
Ways to help
UPI
Scan the code or copy the ID into any UPI app.
Star the repo
A star costs nothing and helps other developers find the project.
Report a bug or suggest a rule
Wrong score, bad AI answer, or a check the quality score is missing. An issue with a small spec is enough.
Contributors1
People who fund APIForge or have sent code to it. GitHub contributors are added automatically.
Entries tagged Example are placeholders that show how the list looks. They are not real people and are not in the count.
Maintainer1
Sponsor1
- Ada Example@ada-exampleSponsorExamplePlaceholder entry that shows how a sponsor looks.Sponsor since Sep 2026
Supporters2
- Sam SampleSupporterExamplePlaceholder entry that shows how a supporter looks.Supporter since Sep 2026
- Priya Placeholder@priya-placeholderSupporterExampleSupporter since Oct 2026
Contributors2
- Jonas Demo@jonas-demoContributorExamplePlaceholder entry that shows how a code contributor looks.Contributor since Aug 2026
- Mei Mockdata@mei-mockdataContributorExamplePlaceholder entry for a rule suggestion.Contributor since Oct 2026
- Become a supporterYour name goes here
Questions
The AI features run on free-tier language models, and free tiers cap how many requests a service can send. Each visitor gets 40 AI requests per rolling 5 hours across the assistant, payload generator, test cases, TypeScript generation and AI-prompt generation. The cap keeps one person from using up the shared allowance for everyone else.
The models behind the free tier are smaller and slower under load than frontier models. They can miss edge cases in a schema, invent details that look right, or refuse a request. Read generated types, payloads and tests before you ship them, and run them against the real API.
Higher limits cost money, so they depend on funding. Support pays for access to more accurate models, higher limits and faster responses. There is no date or target for any of that, and the current limit stays until the cost of raising it is covered.
When you use an AI feature, the relevant part of your spec is sent to the model provider (Groq today) to produce the answer. The APIForge server keeps request metadata such as the feature used, the model, latency and errors. It does not write your spec text to a database. Do not paste specs you are not allowed to share with a third-party model.
Yes. APIForge is open source. Clone the repo, copy .env.local.example to .env.local, and set GROQ_API_KEY to your own Groq key. Your key sets your own limits, and the model can be changed with the optional GROQ_MODEL variable.
Open an issue on GitHub with the feature you used, the spec or a trimmed version of it, and what the model got wrong. A wrong answer caused by a missing rule is often fixable in the prompt or in the rule-based checks, which do not use the AI quota.