Monid Review: One Skill File, 2,499 Tools, and Five Different Tool Counts

Give an agent one line and it wires itself to 2,499 tools. That is Monid’s pitch, and the line is literal:

set up https://monid.ai/SKILL.md

The repo README puts it more bluntly: “Are you an AI agent? Fetch https://monid.ai/SKILL.md and save it to your skill directory. It does the rest.”

That is a real architectural claim, not a tagline. It also sits next to five different tool counts on the same property, a repository that is seven weeks old, and a billing rule most pay-per-call platforms do not offer. All three are worth knowing before you point an agent at it.

Disclosure: the link to Monid in this article is an affiliate link. If you sign up through it, this site earns a share of what you spend. That does not change what follows — the numbers below were measured, and the criticisms are the ones the platform’s own pages volunteer.

What Monid actually is

The repo describes itself as “OpenRouter for agent tools.” The comparison is precise. OpenRouter normalises hundreds of language models behind one API key; Monid normalises hundreds of third-party data and media APIs behind one base URL and one key.

The catalogue spans web search and scraping, people and company enrichment, social platforms, reviews, market and prediction data, and generative media — video, image, voice. Some named integrations from the homepage: Apollo, Exa, Ahrefs, Firecrawl, Browserbase, ElevenLabs, People Data Labs, DeFiLlama, AgentMail, Apollo, MiniMax.

There are three ways to attach it to an agent, all sharing one registry and one balance:

The three verbs, and why two are free

Monid exposes three operations to an agent:

1. discover — ranks the whole catalogue against a described job, across every provider at once. Each candidate comes back with its price, live health, and observed p50 and p95 latency, plus hints naming a cheaper or better-fitting endpoint.
2. inspect — returns one endpoint’s full input schema and contract.
3. run — executes the call and debits the balance.

discover and inspect are free. Only run is billed. That is a deliberate design choice: the agent is meant to shop.

And the agent does the shopping. The homepage says it plainly — “Tell your agent what to do. It picks the tools itself.” The endpoint is chosen per call, not pinned in code months earlier. The README frames the difference against the usual workflow: instead of integrating one vendor that happened to be available at build time, the API is selected at call time against everything currently available, ranked by price and measured latency.

The billing rule worth reading twice

Most metered platforms bill you for the attempt. Monid’s engine settles usage on the raw HTTP response envelope, before any output mapping — the repo’s phrase is “what is billed is what came back over the wire.”

The consequence is stated in the README, in the docs, and again in the tool catalogue:

“A vendor’s non-2xx response is data, not an exception: the run completes and settles at zero usage.”

“A vendor error, an unmatched company, an unresolved person: each of those completes as data and settles at zero.”

So a failed lookup, an empty result, an unmatched person record, or a 5xx from upstream all cost $0. On a platform whose whole premise is paying per call across 89 providers you did not vet yourself, that rule is doing a lot of work.

A unit is defined per connector — a flat charge per call, a charge per returned result, or a rate per unit such as a thousand characters or a second of video. The pricing you pay is the endpoint’s declared model, not a flat platform rate.

What the platform admits about its own tools

This is where the catalogue gets more honest than most. Individual tool entries carry explicit Strengths and Trade-offs sections, and some of the trade-offs are unflattering:

• MrScraper — “every call scrapes the target site live in a real browser, so responses take tens of seconds (typically 10-60s, some scrapers 100s+) and can fail when the target site blocks; failed and empty runs are not charged.”
• cloro — “assistant answers take 30-60 seconds and occasionally minutes, so they run async; answers vary between calls.”
• Litescrape — “app-store and ads endpoints are upstream alpha and their fields may change.”
• Kling — “per-second pricing that climbs steeply at 4K ($0.42/s), output links expire after 30 days, generation is asynchronous (1-5 minutes) with no cancel.”
• Recall.ai — “billed per second of bot time including the waiting room.”
• OpenAI GPT Image — “one PNG per run, up to 2 minutes at the highest quality.”

If you are building an agent that assumes sub-second tool latency, this catalogue tells you upfront that assumption is wrong. That is a useful thing for a marketplace to publish.

Five different tool counts, one property

Here is where a reader should slow down. Monid publishes its own size differently depending on where you look — all read on 2026-10-16:

Two explanations fit, and both are probably partly true. The partner kit’s copy is dated — its comparison table says “read from each program’s public page on 15 September 2026,” and its example run receipts are dated 14-15 September 2026. And the counts measure different scopes: the affiliate kit specifies 1,700+ data tools, while the full 2,499 includes non-data categories such as media generation and voice/phone.

The provider spread (72+ → 89+) is easier to reconcile, and the README explains why. Connectors are declarative definitions added by pull request, and Apify actors can be scaffolded automatically — deno task apify:scaffold <actorId> reads the actor’s published input schema from the Apify API and generates the endpoint schema for review. Public commit history shows batches of 40-50 endpoints landing in single pull requests. The README header simply was not rewritten as the catalogue grew.

The practical takeaway: do not trust any of the five numbers. The CLI can enumerate what is actually there — deno task catalog providers and deno task catalog endpoints.

What a human still has to do

The “one line” framing is accurate about the agent’s side, but a person is still in the loop:

• Create the Monid account and top up the prepaid balance (the site advertises $1 in free credit to start; the affiliate page mentions a $5 top-up running real workloads).
• Generate the single API key and put it in the agent’s environment.

After that the agent calls discover and inspect freely and spends from the balance on run. One human gate, one shared balance, instead of N vendor portals.

Two things the sources do not state, and you should ask: whether a prepaid balance expires, and what the refund policy is. Neither appears anywhere in the material reviewed.

The repository is seven weeks old

Measured through the GitHub API on 2026-10-16 for monid-ai/monid:

• 4,300 stars, 567 forks, 168 watchers (MIT licence, TypeScript)
• Created 2026-08-26 — roughly seven weeks old at the time of writing
• 10 contributors, 57 commits — the top two accounts hold 72% of commits
• Commit timeline: 2 (Aug) → 52 (Sep) → 3 (Oct)
• Latest tag: catalog-v0.0.4. No v1.0.

Read that carefully, because the numbers cut both ways. 4,300 stars on a seven-week-old project is real traction. But 168 watchers against 4,300 stars is under 4% — watchers are the stronger signal, since they asked to be notified. A bus factor of two means the catalogue’s direction currently depends on two people. And a tag line that never left 0.0.x tells you the project does not yet consider itself stable.

None of that is disqualifying for a tool registry — connector definitions are additive, and a broken one is a pull request away from being fixed. It does mean you should not treat the catalogue’s contents as guaranteed.

There is also no stated uptime SLA and no data-freshness SLA. The engine does fail closed: the README notes that “a run-time breach of a declared contract is its own error class rather than a corrupted result,” which is the right behaviour. But an error class is not a guarantee.

Who this is for

A good fit if you are building an agent that needs to reach many external data sources, you do not want to manage N vendor accounts and N keys, and you would rather have the agent pick a tool at runtime by price and measured latency than hardcode one vendor. The zero-charge-on-failure rule removes most of the downside risk of letting an agent choose tools you did not personally vet.

A poor fit if you need a contractual SLA, you depend on a specific vendor’s exact schema, or you need deterministic sub-second latency. Live scrapers in this catalogue are explicitly 10-60+ seconds. And if your workload is narrow — one or two APIs you already use — the single-balance abstraction buys you little.

How to evaluate it yourself

The repository’s own quickstart is the cheapest honest test, and it runs offline:

git clone https://github.com/monid-ai/monid.git
cd monid
deno task check && deno task test   # types + 188 replay tests, zero network
deno task catalog providers          # what is actually there
deno task catalog inspect 'exa#search'

The 188 replay tests run with no network, which means you can verify the connector framework’s behaviour before spending a cent. Then ask the catalogue the questions this article could not answer from public pages: does your balance expire, what is the refund policy, and what does a run that returns HTTP 200 with an empty payload actually bill.

The bottom line

Monid’s central mechanism is legitimate and specific, not marketing: one key, one balance, call-time tool selection, and failed calls settle at zero. The catalogue publishing its own tools’ latency and failure modes is a good sign.

The catalogue’s headline number is not yet trustworthy across the platform’s own surfaces, the project is seven weeks old with a bus factor of two, and there is no SLA. That is a normal shape for something moving this fast — and a reason to verify the specific endpoints you plan to depend on rather than the number on the homepage.

Start with the free credit, run discover on a real task, and read what it returns for price and latency before wiring anything into production.

Try Monid →

Sources: monid.ai homepage, /tools, /tutorial and /affiliates (scraped 2026-10-16); the monid-ai/monid README and repository audit via the GitHub REST API (2026-10-16); tool catalogue entries for MrScraper, cloro, Litescrape, Kling, OpenAI GPT Image and Recall.ai. Tool and provider counts are as published on those pages on that date.