Metatext
Metatext is a free, open directory for the AI ecosystem. It compares large language models on price, benchmark score and context window, and indexes datasets, MCP servers, agent skills and AI jobs. No account required.
What is Metatext?
Metatext is a free, open directory that gathers the scattered pieces of the AI ecosystem into one site. Rather than moving between provider pricing pages, leaderboard sites, dataset hubs and tool registries, you compare large language models on price, performance and context window in a single table. Its organising idea is that every benchmark score appears next to the model's cost per million tokens, so the question becomes best value rather than highest score.
Seven sections make up the directory. LLMs is the price comparator, covering 89 providers and sortable by input and output cost, context window and open-weight status. Benchmarks aggregates four leaderboards — LMArena Elo across 192 models, LiveBench across 71, SWE-bench Verified across 49 and Aider Polyglot across 40 — each score shown against cost. Best LLMs turns that into five shortlists: cheapest, longest context, open-weight, coding and cheapest frontier. Models catalogues 7,967 language and NLP models with their specs, lineage and metadata. Datasets covers 5,413 entries across 179 tasks and 229 benchmarks, each linking to its source, paper and download. Tools indexes 19,622 Model Context Protocol servers, browsable by transport, runtime and author, and 30,701 agent skills, browsable by author and keyword. Jobs lists 244 open AI and machine-learning roles, 238 of them remote.
The intended readers are stated plainly: researchers, engineers and builders deciding which model to ship, which dataset to train or evaluate on, which MCP server or agent skill to plug into Claude or Cursor, or where to find an AI role. The site claims more than two million unique visitors from 89 countries.
Two things bound it. Metatext aggregates publicly available data and measures nothing itself, and its terms describe the content as provided as is, possibly incomplete or out of date, to be verified with the original provider before you rely on it. And it runs without accounts entirely: nothing is saved, nothing is billed, and the whole directory is readable from the first click.
What it does
- Compare large language models by input and output cost per million tokens, context window and provider
- Read benchmark leaderboards with every score paired to the model's price
- Browse ranked shortlists by use case: cheapest, longest context, open-weight, coding, cheapest frontier
- Search AI and NLP datasets by task, language or benchmark status, then follow through to the source, paper and download
- Explore Model Context Protocol servers by transport, runtime and author
- Explore agent skills by author and keyword, each one linking back to its repository
- Filter AI and machine-learning job openings by category and remote status
When to use Metatext / When not to
A quick filter to help you decide if Metatext is the right fit.
When to use Metatext
- AI and ML engineers choosing which model to ship, who need the cost per million tokens sitting next to the benchmark score
- AI researchers and NLP scientists hunting for a dataset to train or evaluate on, filtered by task, language or benchmark status
- Developers building on Claude or Cursor who need to find an MCP server or an agent skill and trace it back to its source repository
- Founders, CTOs and product managers arbitrating the cost of an LLM feature before they commit to a provider
- Engineers and students job-hunting in AI, with 244 open roles listed of which 238 are remote
When not to use Metatext
- Teams that need programmatic access: Metatext publishes no API, so none of this data can be pulled into a pipeline
- Anyone who needs first-party measurements, since every score is aggregated from third-party leaderboards such as LMArena, LiveBench, SWE-bench and Aider
- Buyers who want contractual guarantees: the terms disclaim accuracy and ask you to verify each figure with the original provider
- Users who expect to save comparisons, filters or price alerts, because there are no accounts and nothing persists between visits
- Non-English speakers, as the interface and every listing are English only
How to use Metatext
A typical end-to-end flow, from setup to results.
- Open the site in a browser — there is nothing to install and no account to create
- Start from the LLMs section if your question is about cost: it compares models from 89 providers by price per million tokens
- Sort or filter on input cost, output cost, context window or open-weight status to narrow the field
- Switch to Benchmarks to check how the shortlisted models actually score, with the price shown alongside each result
- Pick the leaderboard that matches your workload — LMArena for human preference, LiveBench for reasoning, SWE-bench Verified or Aider Polyglot for code
- Use Best LLMs if you would rather start from a need than from a model: cheapest, longest context, open-weight, coding, cheapest frontier
- Go to Datasets to find training or evaluation data, filtering by task, language or benchmark status
- Follow a dataset entry through to its source, paper and download rather than trusting the summary
- Visit Tools to pick an MCP server or an agent skill for Claude or Cursor, filtering by transport, runtime, author or keyword
- Check Jobs for AI and ML roles, and optionally subscribe to the newsletter to hear when things change
Pros & Cons
Pros
- Completely free and account-free: the entire directory opens on the first click, with no personal data required
- Cost and performance are treated as one question, with every benchmark score sitting next to a price per million tokens
- Unusual breadth for a single site: pricing, benchmarks, models, datasets, MCP servers, agent skills and jobs
- Serious volume behind the claim — more than 50,000 MCP servers and agent skills combined, 5,413 datasets, 7,967 models
- Every dataset and tool entry links back to its source repository, paper or download, so anything can be checked
- The underlying data is presented as open, and listings are described as refreshed on a continuous basis
- A short, readable privacy policy over a genuinely small data footprint: no accounts, no user content
Cons
- No public API, so nothing here can be pulled automatically into a script, a dashboard or a procurement process
- Every score is aggregated from third-party leaderboards; Metatext measures nothing itself and inherits their biases and their lag
- The terms disclaim accuracy outright and push verification back onto the reader and the original provider
- The publisher is barely identifiable: a trade name only, with no full legal entity, no postal address, no named founder or team and no social presence
- Unmodified starter-kit template pages are still online and contradict the product — /pricing advertises plans at 9.99, 19.99 and 29.99 US dollars with placeholder bullets reading Feature 1 to Feature 7, /faq describes a 14-day trial and invoices, /docs is the starter kit's own documentation, and every page header offers Sign In and Sign Up although the privacy policy states that accounts do not exist
- Nothing persists without an account: no saved comparisons, no stored filters, no price alerts
- English only, and the same site is served on two domains, metatext.ai and metatext.io, with no redirect between them
Pricing & Plans
Metatext is provided free of charge in its entirety. The terms of service state that the directory requires no account and is provided free of charge, and the site's own FAQ confirms that browsing models, pricing, benchmarks, datasets, tools and jobs is free. There is consequently no paid plan, no entry price and no billing currency, and no free trial applies since nothing is chargeable. An unlinked template page at /pricing displays plans at 9.99, 19.99 and 29.99 US dollars per month with placeholder feature labels; it is left over from the starter kit the site was built on, is contradicted by the terms, and has not been treated as a genuine price.
- Free — full access to the whole directory
- with no account
- no time limit and no paid tier above it
Data, GDPR & hosting
A consolidated view of how Metatext handles your data.
GDPR overview
There is no GDPR implementation to report, and that is worth stating plainly: the term GDPR appears nowhere on the site — not in the privacy policy, not in the terms, not on any page collected. No Data Processing Agreement is published or offered, no subprocessor list is maintained, no Article 27 EU representative is designated and no Data Protection Officer is named. What the privacy policy does offer is narrower but real: browsing needs no account and no personal data, Google Analytics and Resend are named as the third parties involved, IP addresses are not stored, and access to or deletion of a newsletter address can be requested at any time by email. The silence is not a claim of non-compliance; the site simply never addresses the question.
Who owns the data?
The terms draw a clear line. Metatext claims the Metatext name, the site design and the compilation of the directory itself; the underlying data, model names and trademarks stay with their respective owners. Nothing you read on the site becomes yours, and nothing catalogued becomes Metatext's. On the visitor's side there is almost nothing to own: browsing requires no account and no personal information. The only personal data handled is pseudonymous analytics and, if you subscribe, the newsletter address you provide, which the publisher states it neither sells nor shares. Your IP address is used to approximate location but is explicitly not stored.
Reuse rights
The terms grant no reuse licence. Metatext presents itself as an informational directory aggregating publicly available data, and it reserves the compilation, the name and the design for itself, so republishing the directory as a whole is not covered. Individual facts — a price, a benchmark score, a dataset reference — belong to the providers and repositories the site links back to, and their own licences govern what you may do with them. In practice the site expects you to read, compare and follow the source link rather than harvest: acceptable use explicitly rules out heavy automated scraping that degrades the service, and no API exists. The terms also insist that figures be verified with the original provider before you rely on them, which places the burden of any downstream use squarely on the reader.
Data retention & training
Hosting summary
Metatext declares no hosting jurisdiction. The privacy policy names the third parties involved — Google Analytics for audience measurement and Resend for newsletter delivery — but says nothing about where anything is stored, and no country or region is mentioned anywhere on the site. Independently of what the publisher states, the domain resolves to an IP address geolocated in the United States on Amazon's network; that address is flagged as anycast, so it identifies a point of presence rather than a place where data lives and should not be read as a hosting location. In practice the surface is very small: there are no user accounts and no user-generated content, so the only personal data that could be stored anywhere is a newsletter email address held by Resend, plus pseudonymous analytics held by Google. Anyone with a documented hosting requirement will have to ask the publisher directly, because the site provides no answer.
Things to keep in mind
Risks and trade-offs to weigh before adopting Metatext.
- Prices and benchmark scores move constantly, and the terms admit the listings may be incomplete, out of date or inaccurate — a procurement decision taken on a stale figure is the reader's risk, not the publisher's
- Leaderboard scores are easy to over-trust: a single Elo number or coding percentage says nothing about how a model behaves on your data, your prompts or your latency budget
- Optimising for the cheapest cost per million tokens can quietly degrade a product, because price per token ignores retries, context bloat and the quality gap that makes a cheap model more expensive in practice
- Template pages left online actively mislead: one advertises invented plans between 9.99 and 29.99 US dollars, another promises a fourteen-day trial, and the header offers Sign In and Sign Up although the privacy policy says accounts do not exist
- The publisher is hard to hold to account — no legal entity, no postal address, no named team — which matters if a figure you relied on turns out to be wrong
- No GDPR statement, no data processing agreement, no subprocessor list and no declared hosting jurisdiction, which is awkward for anyone who has to document their supply chain even for a read-only site
- Google Analytics runs by default, and opting out is left entirely to the visitor through a browser add-on or cookie controls
Setup & Integrations
Technical difficulty
None worth the name. Metatext is a public website: you open it and the whole directory is there — no installation, no configuration, no API key, no account. There is nothing to connect and nothing to provision, and no mobile app or browser extension to add. The only optional step is entering an email address to receive the newsletter. The real effort is interpretive rather than technical: knowing which benchmark matches your workload, and remembering the terms' own advice to confirm any figure with the original provider before acting on it.
Deployment
Supported languages
Behind Metatext
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
Is Metatext free to use?
Do I need an account?
What exactly can I compare on Metatext?
What is an MCP server, and why does Metatext list them?
How current is the pricing and benchmark data?
Does Metatext offer an API?
Can I rely on the prices and scores for a purchasing decision?
What data does Metatext collect about me?
Should you pick Metatext?
Metatext is a decision tool rather than an execution tool: it helps you pick a model, a dataset, an MCP server or a skill, and runs none of them. Judged on that job it is unusually good value, because it is free, opens without an account, and does the one thing most leaderboards refuse to do — it puts the price next to the score. The catalogue is large enough to justify the visit on its own, with more than fifty thousand MCP servers and agent skills, 5,413 datasets and 7,967 models, and almost every entry links back to the repository, paper or provider it came from, so little has to be taken on faith.
The weakness is the publisher rather than the product. Metatext is a trade name with no full legal entity, no postal address, no named team and no social presence, and the site never mentions the GDPR, publishes no data processing agreement and declares no hosting jurisdiction. Its own terms disclaim accuracy and hand verification back to you. More confusingly, several unmodified template pages from the SaaS starter kit it was built on are still online and contradict the real product: invented pricing plans, a promised fourteen-day trial, sign-in buttons for accounts that do not exist. None of it looks malicious, but a hurried reader can mistake scaffolding for the offer.
Read it, then, for what it is — a fast, free, well-organised map of a market that moves weekly, maintained by someone who has not said who they are. Use it to shortlist and to orient yourself, and confirm any number destined for a budget with the provider that publishes it.
- Choosing a selection results in a full page refresh.
- Opens in a new window.