Free Tool-calling agent stack
An agent loop on a free LLM API that publishes tool-calling support, deployed on a free host.
2 slots filled from the catalog, verified rows first; see how we verify. Other templates: Chat app, RAG / docs Q&A, Image tool, Voice agent, Fine-tune and serve a model.
Your stack
Free Tool-calling agent stack: OVHcloud AI Endpoints Free Tier, Cloudflare Workers Free Plan
- Cards required
- OVHcloud AI Endpoints Free Tier
- Programs to qualify for
- None
- Tightest published LLM bound
- 576,000 requests/day on OVHcloud AI Endpoints Free Tier (Qwen3.8-27B, by requests per minute, sustained)
- Free GPU time
- Not part of this template
The catches
- OVHcloud AI Endpoints Free Tier: Needs a saved payment method (the US site also requires a $0.99 purchase); billing starts automatically once the credit is used up; anonymous API use is capped at 2 requests per minute.
- Cloudflare Workers Free Plan: 10 ms CPU time per invocation and 100,000 requests a day; over the limit, further requests fail until the daily reset at 00:00 UTC.
- LLM: ovhcloud.com
- Host: cloudflare.com
Frequently asked questions
- Where do the picks come from?
- From the aifree.dev catalog: LLM, embedding and speech slots use the models each offer serves free with their published rate limits; the image slot uses free image tiers; the GPU slot uses the published free GPU quotas; vector database and hosting slots use a short curated list of free plans and startup programs that the nightly lint checks are still live; a listing that is not public yet is left out.
- Why are some picks startup programs?
- Some infrastructure is free only through a startup program. The vector database and hosting slots list free plans first and startup programs after; a program pick shows its eligibility checklist (incorporated, company age, funding) so you can see at once whether you qualify, and you can replace it with a free plan if you do not.
- How are the totals computed?
- Only from stored data: cards required come from each offer, the tightest LLM bound from published rate limits, GPU hours from the vendor quota file, and the catches are each offer’s own "the catch" line. Nothing is estimated.
- Can I change a pick?
- Yes. Every slot has a "replace this pick" selector with up to eight alternatives; the URL names every pick, so the link you share rebuilds exactly that stack even after the catalog re-ranks.