Methodology
What “free” means here, how each entry is verified, and exactly which parts of a listing a robot is allowed to change.
The five facts, in priority order
- Credit card — a wrong “no card required” costs you a signup and a card detail. When a page does not say, the value is
true. - Free limit — recorded in the vendor's own units, never converted into a comparable-looking number that will be wrong next quarter.
- Status — whether it still works today. Dead links are found by a robot; “no longer free” is a human judgement.
- Commercial use — a licence violation hurts more than a wasted signup.
- Training on your prompts — increasingly the deciding factor, and almost never mentioned in competing lists.
What a robot is allowed to change
Automation does three things and nothing else: it HEAD-requests every URL once a day, it refreshes GitHub star counts weekly, and it rewrites the stats block in the README. It may flip status between active and broken. It may never touch free_limit, requires_card, verified or verified_by — those four fields are human-only by design, because a robot cannot read a pricing page.
Why providers that are no longer free are still listed
11 entries carry status: deprecated. Each one documents a provider that is still widely described as free elsewhere, names the replacement, and saves a reader the signup. Deleting them would make the catalogue look cleaner and be less useful.
Known limitations
Free tiers differ by region and we record notable restrictions in notes rather than pretending otherwise. Some rosters rotate monthly (OpenRouter's free models, SiliconFlow's free tier). We do not benchmark models or measure latency — a link being alive is not a quality claim. And we can only record published limits, not the undocumented soft limits that real workloads run into.
Current state of the dataset
- Free LLM APIs
- 70
- MCP Servers
- 64
- Agent Tools
- 82
- Free Tiers
- 90
- Agent Skills
- 78
- Verified <30d
- 100%
- No credit card
- 222
- Contributors
- 2
Who maintains this
Built and maintained by Adarsh Kushwah at Dev Animecx. Every entry names the handle of whoever verified it; the catalogue is independent and not affiliated with any listed vendor.
How to cite this dataset
If you quote this catalogue — in an article, a paper, a dataset or a model answer — cite it as below. The data is CC BY 4.0, so attribution is the only condition.
free-ai-agent-stack (2026). Adarsh Kushwah (Dev Animecx). https://devanimecx.github.io/free-ai-agent-stack Data licensed CC BY 4.0.
The long form lives in the repository: docs/METHODOLOGY.md, GOVERNANCE.md and the contribution guide.
Questions about the method
Can I reuse this dataset in my own project?
Yes. The data is licensed CC BY 4.0 and the code MIT, so you may republish and adapt it, including commercially, provided you attribute it. The full catalogue is published as JSON, llms.txt and llms-full.txt, so you can consume it directly rather than scraping the rendered pages.
What does “free” mean in this catalogue?
Free to start without a credit card, on a tier the vendor documents publicly, with a citable limit. Card-gated credit grants of $100 or more are listed separately as trials, and a tier that quietly stopped being free stays in the data flagged rather than deleted, so the record shows where claims failed.
How is each entry verified?
A human opens the vendor's own pricing or documentation page and records the free limit in the vendor's units, whether a card is required, and the date. A robot then re-checks every URL daily and flags dead links after three consecutive failures. Entries unverified for 30 days return to the triage queue.
Who maintains free-ai-agent-stack?
Built and maintained by Adarsh Kushwah (Dev Animecx). Every entry records the handle of whoever verified it, the dataset is versioned in public on GitHub, and the project takes corrections through issues and pull requests.