Monid Review: The 'OpenRouter for Agent Tools', Tested on Real Work

Monid is pitched as the OpenRouter for agent tools: one prepaid balance, one integration, and your agent buys whatever API it needs at the moment it needs it. I did not review it from the homepage. I wired it into Notis as the pay-per-use tool layer, then spent a morning this week pushing real jobs through it and reading the receipts. The router works. The metering is genuinely excellent. The catalogue and the ranking are where it gets interesting, and the governance is where I ended up writing my own.

What Monid actually does

Strip the marketing and it is three verbs. discover takes a plain-English query and returns ranked candidate endpoints with prices attached. inspect returns one endpoint's input schema and pricing breakdown. run executes it and debits your balance. That is the entire surface, exposed as a REST API, a remote MCP server and a CLI, all documented here.

Three tool slots buy a catalogue of any size, and your agent's context does not grow when the catalogue does. Anyone who has watched an MCP server dump two hundred tool definitions into a prompt knows why that matters.

One thing before you architect around it: the MCP surface authenticates by OAuth through a Monid account, not by API key, and the named clients in the MCP quickstart are Claude.ai, Claude Code, Codex and OpenCode. Fastest path in for one developer; a platform serving many users wants the REST API and a key.

The Monid catalogue is smaller than the headline

On 19 August 2026 the unauthenticated stats endpoint returned 1,608 endpoints across 54 providers. The marketing does not agree with it, or with itself: monid.ai renders "1300+ Tools across 13+ providers" in the page body while that page's own meta description says "1,800+ tools & APIs". Three numbers, one product. I took the API's, then paged through the public providers endpoint and added it up myself.

One provider, TikHub, accounts for 898 of those 1,608 endpoints: 56% of the catalogue from a single social-scraping supplier. The top five — TikHub, DefiLlama, Strale, BlockRun and Apify — are 76% of it, and the top ten are 82%. Those are the endpointCount fields in that providers listing, summed.

Break it down by the category tags in the public endpoints listing and it gets sharper: 301 Douyin, 154 TikTok, 64 Weibo, 42 Bilibili, 41 Zhihu, 38 Kuaishou. If your business does not touch Chinese short-video platforms or DeFi analytics, most of that headline number is not for you.

That is not a bait and switch, just a young marketplace: one anchor supplier plus a long tail. The tail is where the useful names live — Apify with 41 endpoints, Ahrefs with 36, plus Apollo, People Data Labs, Exa, ElevenLabs and Browserbase. Judge Monid on that tail: browse the catalogue first.

One real job, and the ranking picked the 26x option

Here is the test that told me the most. I asked discover for "read the text content of a web page url", and it answered in under a second with a ranked list.

Top result: Strale's /x402/url-to-text at $0.02376 a call. Third result: Context.dev's /web/scrape/markdown at $0.0009 a call. Same job, 26 times the price at the top of the list.

The tell sits in Monid's own catalogue before you spend anything, reproducible without an account. In the public endpoints listing, the $0.02376 top result describes itself as "Lightweight text extraction from URL. No JS rendering — just HTTP GET + HTML parse" and carries an empty tags array. The $0.0009 result underneath it is "Scrape any URL into clean, LLM-ready Markdown" and carries the verified tag. So the cheaper endpoint is the vetted one, and Context.dev's own page says that scrape handles JavaScript rendering for you, which is what a client-side-rendered page needs. It still came third. I ran both on the same URL and paid both.

To Monid's credit the ranking is auditable rather than mysterious: every discover result comes back with a scoreBreakdown object. The discover API reference does not document that field, so this part is my own observation from the responses I got — the components returned were query relevance, a verified-tag bonus and a pricing-predictability term. Price level is not one of them, and in my query the top result did not even collect the verified bonus. Read the price column yourself, and do not let an agent blindly take result number one.

The metering is the best part of the product

I expected this to be the weak spot. It was the opposite.

My wallet held $24.606698 before those two runs and $24.582038 after. The difference is $0.024660, which is exactly $0.0009 plus $0.02376. No rounding in the house's favour, and runs report cost in micro-dollars, so sub-cent calls are not quietly inflated.

Then I pointed the scraper at a domain that does not exist. HTTP 400 came back, the run reported zero cost, and the balance did not move. The "a provider failed, do I get charged" promise holds, checked against the wallet rather than the FAQ.

Latency was fine for interactive use: discover in 0.9 seconds, the two runs in 5.5 and 3.1 seconds by Monid's own timestamps.

What is not there yet

Three things stopped me from letting an autonomous agent loose on it.

No published terms or privacy policy. On 19 August 2026, /terms, /privacy, /legal and /privacy-policy all return 404 on monid.ai. That matters here, because the runs API returns both the input you sent and the output the provider returned for a past run — I pulled one of mine back an hour later and got the full scraped page. Payloads are stored, and no retention or training statement anywhere says for how long. For customer personal data in the EU, that is a blocker until they publish one.

No spend cap where you need one. The key management commands take a label, a key value, a JSON flag and a force flag — no budget flag, no per-key limit. The word budget does not appear in the docs, and the wallet API returns a single balance for your workspace, so that balance is the only ceiling. One agent stuck in a retry loop drains it for everything else on that workspace. The OAuth guide is explicit that tokens are scoped to a user, so a platform builder has to enforce per-user limits before the call, not after.

The price field has more shapes than the docs admit. The pricing guide documents two models, PER_CALL and PER_RESULT. The live catalogue also returns PER_UNIT_MATRIX on 16 endpoints, TIERED on 6 and PER_UNIT on 1. On those the top-level amount is not a per-call price: MiniMax's text-to-speech endpoint reports 60, which its own variants resolve to $60 per million characters, not $60 a call. A naive pre-flight check reading that field would block a cheap call or wave through an expensive one.

Two footguns while you write that check. Saperly's /provision-numbers costs $2 and provisions a real phone number, a recurring commitment created from a pay-per-call wallet. BlockRun's buy-a-number endpoint is $5.5011. In fairness, those are two of only four per-call endpoints priced at or above $1 in that same listing: across its 1,503 PER_CALL endpoints the median price is $0.0015 and 78% cost a cent or less.

How I run it in production

Notis exposes a curated slice of Monid as pay-per-use tools, and every guard rail sits on my side of the wire. Notis charges 20% over Monid's price. Any single pay-per-use call quoted above $1.20 to the user is refused outright, and the agent is told to state the price and ask for confirmation before trying again. The catalogue is an allowlist, not all 1,608 endpoints.

Prices are published per tool rather than hidden in a credit pack: reading a web page is about $0.004, and person or company enrichment starts at $0.12 a call. Both come out of the usage already included in the plan, which starts at $13 a month billed annually.

Should you use it

Use Monid directly if you are building your own agent, you want breadth without twelve vendor signups, and you are the one watching the balance. The ergonomics are the best I have used in this category.

Wait if you handle customer personal data under GDPR, or need contractual commitments. Ask for a DPA, a retention statement and documented rate limits first.

Do not build multi-tenant on it yet without your own per-user cap and your own allowlist. That is exactly the work I had to do, and it is not small.

Skip the integration entirely if what you want is the outcome rather than the plumbing — a page read, a company enriched, a lead researched — delivered into the tools you already use.

Monid is the first tool router I have tested where the money side is more trustworthy than the catalogue side. That is an unusual order to get right, and the harder half. Publish the legal pages, add per-key budgets, and this becomes infrastructure.