One key instead of a dozen: what Monid does, and how many tools are really in there?

Anyone who has wired external data into an agent keeps the same mental folder: a key for the search API, a key for the scraper, a key for company enrichment, a key for the image generator. Each with its own response shape, its own billing, and its own way of charging full price for an empty result.

Monid offers to collapse that folder into one key. It shipped on 16 September, MIT licence, written on Deno. The authors describe the idea themselves: OpenRouter, but for tools.

We installed and checked it, because writing about a tool from its README is not on.

Three verbs, and the first two are free

discover takes the job and ranks the whole catalogue against it across every provider at once, returning candidates with price, live health and observed latencies. inspect returns one endpoint's full contract. run executes and is billed.

The point of that split is that the endpoint gets picked at call time instead of being pinned in code six months ago to whichever vendor happened to get integrated back then.

Where the real engineering is

Not in the catalogue. In how the money is counted.

Every connector declares its own usage model right in the definition: a flat charge per call, a charge per returned result, or a rate per unit such as a second of video. The engine settles that against the raw response envelope, before any output mapping. What gets billed is what came back over the wire.

One consequence is rare in paid APIs: a vendor error, an unmatched company, an unresolved person all complete as data and settle at zero. You do not pay for nothing.

A second decision of the same calibre: the endpoint description is the product, because it is what discover ranks. In the authors' example the search description ends by admitting that this endpoint returns snippets only and that full text lives in a neighbouring one. An endpoint that names its own successor is worth more than any amount of parameter documentation.

What we checked by hand

git clone --depth 1 https://github.com/monid-ai/monid.git
cd monid
deno task check && deno task test

Replay tests instead of live calls are the right call: you cannot fire 340 endpoints for real on every commit, yet response parsing and cost accounting do need checking.

The next command, though, fell over.

error: Uncaught (in promise) Error: The filename, directory name, or
volume label syntax is incorrect. (os error 123)
    at versionOf (scripts/lib.ts:20:20)

The cause is one line in scripts/lib.ts: the repository root is taken as new URL("../", import.meta.url).pathname. On Windows a file URL's .pathname gives /C:/Users/... with a leading slash, and join() then assembles the non-existent \C:\Users\.... The fix is also one line, using a standard library function:

import { fromFileUrl, join } from "@std/path";

export const REPO_ROOT = fromFileUrl(new URL("../", import.meta.url));

Note that check and test do not catch this: the tests never go through that file, so a green build on Linux says nothing about Windows.

How many tools are actually in there

With the catalogue working, we counted it.

This is not a deception, but the difference is worth knowing. What is open is the connector layer: the format plus the providers already described in it, while the full catalogue lives on the hosted service and cannot be verified from the repository. The open part looks sensible: Apify with 46 endpoints, Ahrefs with 36, Firecrawl, Exa, and the video generators Kling and MiniMax. A third of the catalogue is Surf alone, with 105 crypto-analytics endpoints.

On money the documentation says less than one would like: the usage models are described, but the platform's markup and whether a free tier exists are not disclosed.

Monid or MCP: these are different things

The question everyone asks who already runs a stack of MCP servers.

That last row is more interesting than it looks, and we took it apart in another piece the same day: every MCP server you connect writes its schemas into the first call to the model, and the opening context grows. A gateway that hides hundreds of endpoints behind three verbs changes that arithmetic in your favour. Monid has an MCP endpoint of its own for exactly this, so you can connect it as an ordinary server.

A sensible split: MCP for what lives with you (files, git, your database, internal services), a gateway for what lives with vendors and costs money.

Who this is for right now

Worth a look if you are building an agent that needs varied external data and you are tired of parsing five response shapes and reconciling five bills. Particularly if the set of sources is not known in advance, which is when discover earns its keep. Worth waiting if you have one or two sources already wired in. Worth joining if you run an API and want a channel to agents: a declarative connector plus a pull request is less work than maintaining an MCP server of your own.

And the caveat we attach to every week-old project: two hundred-odd stars, incomplete pricing documentation, and a dependency on the hosted service for the most interesting part. Before production, check your own scenario by hand and keep a direct vendor call as a fallback on the critical path.

Versions and figures as of 17 September 2026: Deno 2.9.6, 21 providers and 340 endpoints in the open repository. The project is young and moving fast, so check the current state of the catalogue and its pricing before planning around them.