Every SaaS company suddenly has an "MCP server". Most articles about it are either a press release or a wall of JSON. This one is neither. Here is what the Model Context Protocol actually does, drawn out rather than described; what it looks like from the inside, using real output from the server we run ourselves; and which of the 300 tools in our database offer one — with what it costs, taken from their own documentation, and "not stated" where they don't say.
Start with the annoyance, not the acronym. You are in Claude or ChatGPT, and you ask something that needs your own data: what did we spend on software last quarter, which support tickets mention the new pricing, is this contractor's paperwork done. The model can reason perfectly well about all three. It just cannot see any of it.
The old fix was an integration per pairing. Every AI app had to build a connector for every tool, and every tool had to build one for every AI app. Three clients and three tools means nine integrations to write and — the part that actually kills projects — nine to maintain when an API changes. That is the mess on the top half of the picture below.
MCP, an open standard originally published by Anthropic, replaces that with one shared interface. The tool ships a server. The AI app is a client. Each side implements the protocol once. Three plus three, and the fourth AI client you adopt next year costs nothing to support.
Our illustration. The count is the point: integrations grow multiplicatively without a standard, and additively with one.
That is genuinely all the idea is. The reason it spread so fast is not technical elegance — it is that it let vendors stop guessing which AI assistant their customers would standardise on.
get_dossier(name). The model reads the list of tools and their descriptions, decides which to call, and the client executes it. The description text matters enormously: it is the entire instruction manual the model gets.An MCP session is short and boring, which is a compliment. Messages are JSON-RPC 2.0. The client says hello, asks what the server can do, and then calls things. There is no magic and no model involved on the server side at all — the server is a dumb, well-described API surface, and all the intelligence stays in the client.
Our illustration, using the real method names. If you have ever wondered what your AI client is doing in the two seconds after you connect a tool: this.
Two consequences follow from that shape, and they explain most of what you will experience in practice. First, tool descriptions are the product. The model picks tools by reading their text, so a server with vague descriptions gets ignored or misused, no matter how good its API is. Second, the result goes into the context window. A server that dumps 40KB of JSON per call will crowd out the conversation and cost you money. Good servers return compact prose. That is a design decision, and vendors vary wildly on it.
We are not a neutral bystander here. We publish an MCP server ourselves — the Honest Atlas MCP — and everything above comes from having built and maintained one rather than from reading the spec. Here is what it is, honestly.
The problem we built it for is narrow and real: AI assistants are confidently wrong about SaaS pricing. Ask any of them what a tool costs and you get a number that was true eighteen months ago, or a vendor's headline price with the per-seat multiplication quietly omitted. Meanwhile we sit on 300 hand-researched dossiers with the real pricing notes, the honest knock, and who should skip each tool. Publishing that as a webpage helps humans. Publishing it as an MCP server helps the place where people are increasingly making the decision.
Our illustration of our own server. 242 lines of Python, standard library only, no account, no telemetry.
It exposes four tools. This is the literal tools/list response, captured by piping JSON-RPC into the
server on the machine this page was built on:
$ echo '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' | python3 atlas_mcp.py
search_tools find tools by need, with honest one-liners + pricing
get_dossier the full honest dossier for one tool
compare_tools 2-4 tools side by side
list_categories orient a search
And this is a real answer, not a mock-up — get_dossier("Miro"), trimmed for length:
# Miro — honest dossier Category: Operations & Workflow · miro.com The honest knock: Per-member pricing charges for every member added to the workspace, so an org-wide rollout multiplies at full headcount — and the free plan's 3-board cap is workspace-wide, not per person, so real teams outgrow it in days. Model who truly needs to edit before you buy seats. — Source: The Honest Software Atlas by AIBuilder Marketplace. Hand-researched; no invented numbers. try_url links are affiliate links (commission possible, never changes the take).
The disclosure has to travel with the data. Our dossiers contain affiliate links. On a webpage the
disclosure sits at the bottom of the page and the reader sees it. Through MCP there is no page — the model receives
text and paraphrases it to the user. So the affiliate disclosure and the source attribution are appended to every
single tool response, and the server's initialize instructions tell the model to pass both through. If you
publish an MCP server with commercial links in it and you do not do this, your disclosure effectively disappears.
Compact beats complete. Our first version returned the full dossier JSON. It was accurate and useless: it ate the context window and the model started summarising away exactly the caveats we care about. Returning short, opinionated prose with the knock intact works far better.
Stale data beats no data. The server caches the dossier feed for 24 hours. If the fetch fails and the cache is expired, it uses the expired cache anyway rather than erroring. An assistant that says nothing is worse than one that says something slightly old with a date attached.
Zero dependencies was the right call. No pip install, no lockfile, no supply chain. One file you can read in ten minutes before you run it — which, given that you are about to let it talk to your AI assistant, is the minimum any of us should demand.
We went through all 300 tools in our database and checked, one by one, whether the vendor themselves publishes an MCP server. The rule was strict: it counts only if it is documented on their own domain. Community servers on GitHub — and there are hundreds, some excellent — do not count here, because you cannot hold a weekend project to an uptime promise. Several well-known tools failed that test: there are popular third-party MCP servers for Pipedrive, Aircall, Toggl Track and Kinsta, but nothing first-party we could find, so they are not in this list.
35 passed. Each card below states what the server actually does, and what the vendor says about plan and price — in their words, with a link to the page we took it from. Where they say nothing, we say "not stated".
MCP for 1Password Environments: create environments, list variable names and handle .env files — the secrets themselves never enter the model's context window.
Plan & price: Not stated in their own documentation.
Hosted remote server: read and update contacts, tags, custom fields, lists and automations. You copy your own MCP URL from your developer settings.
Plan & price: Their help centre states the MCP server is available on all plan tiers.
Server at mcp.apollo.io/mcp: search and enrich prospects, manage records, run sequences and read analytics.
Plan & price: Their docs: there is no additional cost to connect — your existing Apollo plan, permissions and credits still apply. Enrichment therefore still burns credits.
Your newsletter's content, subscribers, performance metrics, podcasts, automations, products, surveys, polls and referrals.
Plan & price: All beehiiv users, no upgrade required — but on a free plan it is read-only. Write actions require a paid beehiiv plan.
Server at mcp.brevo.com: 27 modules spanning contacts, email campaigns, CRM deals, SMS and WhatsApp, templates and account management.
Plan & price: Not stated in their own documentation.
Search, navigate and extract public web data without getting blocked. Runs fully managed or self-hosted.
Plan & price: Every new account includes 5,000 free requests a month; beyond that you go through your account manager.
Official trading MCP with 206 tools across market data, orders, positions and account. The 22 market-data tools work read-only without any API key.
Plan & price: Price not stated; read-only market data works without a key.
Read community analytics, member lists and posts; write actions such as inviting members or posting need your explicit approval each time.
Plan & price: Business plans and above, at no additional cost. Community admins only — regular members cannot use it.
`clay mcp` runs Clay's tools as a local server for coding agents; there is also a hosted endpoint.
Plan & price: Not stated in their own documentation.
Hosted server at mcp.clickup.com/mcp covering tasks, docs, time tracking, workspace search and chat. Delete operations are deliberately not supported.
Plan & price: All plans. Free: 50 calls a day; Unlimited and above: 300 a day on a rolling 24-hour window; with the Everything AI add-on you get the standard public API limits.
OAuth server for leads, activities and opportunities, with separate scopes for read, safe-write and destructive-write.
Plan & price: Not stated in their own documentation.
Server at mcp.databox.com/mcp with two core tools: ingest_data and query_dataset_with_ai — push data in, then ask questions about it in plain language.
Plan & price: Not stated in their own documentation.
Server at api.letsdeel.com/mcp: retrieve contracts, run worker-level operations and submit time-off requests. OAuth2 or a personal access token.
Plan & price: Not stated in their own documentation.
First-party hosted server: import media, run Underlord edits, create and manage compositions, and export.
Plan & price: Not stated on the page we were able to read.
Publish any Dify app as an MCP server — and consume other MCP servers as tools inside your workflows. It works both ways.
Plan & price: Not stated in their own documentation.
Generate speech, clone voices, transcribe audio and spin up outbound voice agents from Claude or Cursor.
Plan & price: The server is open source and free to run locally; every call spends credits from your ElevenLabs account.
Server at api.fireflies.ai/mcp: meeting transcripts, summaries, action items and speaker data, over OAuth.
Plan & price: Not stated in their own documentation.
MCP integration that lets Cursor or Claude connect securely to your Freshservice environment.
Plan & price: Early Access Program: beta, Enterprise plan only, and on request via your technical account manager or support.
Generate decks, docs and sites from a prompt or a template, check generation status, export to PDF/PPTX/PNG and read analytics.
Plan & price: Available on all Gamma plans; generations charge credits.
Read-only access to payroll and people data; you choose which categories of data the model is allowed to see.
Plan & price: Not stated in their own documentation.
Employee data, time off and tasks over a hosted endpoint, fenced in by the permissions of a Bob service user.
Plan & price: Not stated in their own documentation.
Exposes the Hostinger API: websites, VPS, domains and DNS, email marketing, subscriptions and payments, and e-commerce.
Plan & price: Not stated. You need an API token from hPanel, or OAuth for production use.
Server at mcp.jotform.com: create forms and read submissions, with an OAuth consent screen per user.
Plan & price: Not stated in their own documentation.
Server at app.lemlist.com/mcp: create, update, start and pause campaigns and sequences, manage leads, enrich emails and read team analytics.
Plan & price: Price not stated; enrichment operations (email finding, verification, phone enrichment) do consume credits.
Server at mcp.miro.com: read and build boards — frames, sticky notes, shapes, cards, tables, architecture and sequence diagrams, plus reading and resolving comments.
Plan & price: Enterprise. An admin has to enable the MCP server for the organisation before anyone can connect.
Read access across the travel and expense dataset: spend summaries, transaction detail and policy data. Write access is announced but not live.
Plan & price: Not stated in their own announcement.
Official Plesk extension exposing the Plesk and WP Toolkit REST APIs to ChatGPT, Claude or Gemini — server administration in plain language.
Plan & price: Free from the Plesk Extension Catalog. Plesk Obsidian for Linux 18.0.72 and above only; no support during the beta and not meant for production.
Server at mcp.usepylon.com: read and update issues, accounts and contacts — exactly the data you can already see in Pylon yourself.
Plan & price: Price not stated. You do need a Member or Admin role with 'MCP Access' enabled, and rate limits apply per tool.
Three tools — create_app_from_prompt, update_app_using_prompt and ask_question — letting an MCP client drive Replit Agent to build real apps.
Plan & price: Works with a Free, Core, Pro or Enterprise account. What the underlying Agent runs cost is not specified in that documentation.
Two servers: an API server for pods, endpoints, templates and volumes (needs an API key), and a docs server at docs.runpod.io/mcp that needs no authentication at all.
Plan & price: Not stated in their own documentation.
Server at mcp.shippo.com: compare carrier rates, buy labels, track packages, validate addresses and handle customs.
Plan & price: Works across all plan types. Reads — rates, address validation, tracking — cost nothing; buying a label charges your live account and asks for confirmation first.
Server at mcp.softr.io/mcp: 20 tools across databases, tables and records — read and write, over OAuth or a personal API token.
Plan & price: Not stated in their own documentation.
Hexa AI MCP at cloud.tenable.com/mcp: 90 structured tools over the Exposure Data Fabric — assets, vulnerabilities, dashboards and scans.
Plan & price: Requires a Tenable One Foundation or Tenable One Advanced licence. Paid only, and not available in FedRAMP Moderate environments.
First-party server at ai.todoist.net/mcp, plus a Todoist plugin that wires up the configuration for you.
Plan & price: Not stated for Todoist itself; their guide does note that Claude Code requires a paid Claude plan.
Query projects, manage tasks and navigate folders from Claude, ChatGPT, Copilot or Cursor.
Plan & price: Not stated in their own documentation.
| Partner | Category | Status | What they say about plan & price | Source |
|---|---|---|---|---|
| IT & Productivity | Beta | Not stated in their own documentation. not stated | source | |
| Growth & Revenue | Live | Their help centre states the MCP server is available on all plan tiers. stated | source | |
| Growth & Revenue | Live | Their docs: there is no additional cost to connect — your existing Apollo plan, permissions and credits still apply. Enrichment therefore still burns credits. stated | source | |
| Growth & Revenue | Live | All beehiiv users, no upgrade required — but on a free plan it is read-only. Write actions require a paid beehiiv plan. stated | source | |
| Growth & Revenue | Live | Not stated in their own documentation. not stated | source | |
| IT & Productivity | Live | Every new account includes 5,000 free requests a month; beyond that you go through your account manager. stated | source | |
| Financial Operations | Live | Price not stated; read-only market data works without a key. not stated | source | |
| Growth & Revenue | Live | Business plans and above, at no additional cost. Community admins only — regular members cannot use it. stated | source | |
| Growth & Revenue | Live | Not stated in their own documentation. not stated | source | |
| Operations & Workflow | Public beta | All plans. Free: 50 calls a day; Unlimited and above: 300 a day on a rolling 24-hour window; with the Everything AI add-on you get the standard public API limits. stated | source | |
| Growth & Revenue | Live | Not stated in their own documentation. not stated | source | |
| Operations & Workflow | Under active development | Not stated in their own documentation. not stated | source | |
| HR & People | Live | Not stated in their own documentation. not stated | source | |
| Operations & Workflow | Live | Not stated on the page we were able to read. not stated | source | |
| IT & Productivity | Live | Not stated in their own documentation. not stated | source | |
| Content & Media | Live | The server is open source and free to run locally; every call spends credits from your ElevenLabs account. stated | source | |
| Communication & Voice | Live | Not stated in their own documentation. not stated | source | |
| IT & Productivity | EAP beta | Early Access Program: beta, Enterprise plan only, and on request via your technical account manager or support. stated | source | |
| Operations & Workflow | Live | Available on all Gamma plans; generations charge credits. stated | source | |
| Financial Operations | Live (read-only) | Not stated in their own documentation. not stated | source | |
| Operations & Workflow | Beta | Not stated in their own documentation. not stated | source | |
| IT & Productivity | Live | Not stated. You need an API token from hPanel, or OAuth for production use. not stated | source | |
| Operations & Workflow | Live | Not stated in their own documentation. not stated | source | |
| Growth & Revenue | Live | Price not stated; enrichment operations (email finding, verification, phone enrichment) do consume credits. not stated | source | |
| Operations & Workflow | Live | Enterprise. An admin has to enable the MCP server for the organisation before anyone can connect. stated | source | |
| Financial Operations | Read-only | Not stated in their own announcement. not stated | source | |
| IT & Productivity | Very early beta | Free from the Plesk Extension Catalog. Plesk Obsidian for Linux 18.0.72 and above only; no support during the beta and not meant for production. stated | source | |
| Communication & Voice | Live | Price not stated. You do need a Member or Admin role with 'MCP Access' enabled, and rate limits apply per tool. not stated | source | |
| IT & Productivity | Beta | Works with a Free, Core, Pro or Enterprise account. What the underlying Agent runs cost is not specified in that documentation. stated | source | |
| IT & Productivity | Live | Not stated in their own documentation. not stated | source | |
| Operations & Workflow | Live | Works across all plan types. Reads — rates, address validation, tracking — cost nothing; buying a label charges your live account and asks for confirmation first. stated | source | |
| IT & Productivity | Live | Not stated in their own documentation. not stated | source | |
| IT & Productivity | Live | Requires a Tenable One Foundation or Tenable One Advanced licence. Paid only, and not available in FedRAMP Moderate environments. stated | source | |
| Operations & Workflow | Live | Not stated for Todoist itself; their guide does note that Claude Code requires a paid Claude plan. not stated | source | |
| Operations & Workflow | v2 in beta | Not stated in their own documentation. not stated | source |
Researched 2026-07-28. MCP is moving fast — check the source link before you commit to anything. Spotted a partner we missed, or a page that has changed? Tell us and we'll fix it.
First, the honest framing, because a lot of "best MCP server" content gets this wrong: for almost every tool here, the MCP server is not a product you buy. It is a door onto a subscription you already pay for. So ranking them by price is mostly meaningless. What is worth ranking is the door: can you open it on the plan you are actually on, how often can you use it, and will it let you write as well as read.
Our own count, from the 35 vendor pages linked above.
Both put the server on every plan including free. beehiiv is explicit that all users get it with no upgrade, with free plans read-only and writes requiring a paid plan — an honest, legible split. ClickUp is on all plans too and publishes its actual rate limits: 50 calls a day on Free, 300 on Unlimited and above. Published limits are rarer than they should be, and they are what tells you whether the thing is usable.
7 of the 35 are stated to work without a paid plan at all: ClickUp, Bright Data, Replit, ElevenLabs, Beehiiv, Plesk, Gamma — though "free plan" and "free to run" are different promises, and each row above says which one applies.
5,000 requests a month on every new account, stated in writing. For an agent that needs to read the live web that is a genuine working budget rather than a demo, and it is the only server in this list that quantifies a free tier that precisely.
Their documentation says plainly that there is no additional cost to connect and that your existing plan, permissions and credits still apply. That is exactly the sentence a buyer needs. Note the tail of it though: enrichment through MCP spends the same credits it always did, so an agent left to enrich a large list will cost you real money quietly.
Shippo separates free reads (rates, tracking, address validation) from actions that charge your live account, and asks for confirmation before buying a label. Circle keeps read tools always-allowed and routes every write through an explicit approval. When an AI is holding the pen, that distinction is worth more than any feature list.
These four are the ones where the plan genuinely decides. Miro's is Enterprise and an admin must switch it on. Tenable's Hexa AI MCP needs a Tenable One Foundation or Advanced licence. Freshservice's is an Early Access Program on Enterprise, by request. Circle's needs Business or above and admin rights, though it costs nothing extra on top. If you are evaluating any of these because of MCP, check your tier before you get attached.
21 of these 35 vendors publish nothing about plan or price for their MCP server anywhere we could find on their own site. That is not a criticism of the servers — several are excellent — but we are not going to invent a price to fill a column. If pricing matters to your decision, ask them, and treat the silence as a question to raise rather than a detail to assume.
An honest guide has to include the part the launch posts leave out.
"Has an MCP server" says nothing about quality. The gap between servers is enormous — a handful of well-described tools that return readable text versus 200 auto-generated endpoints that flood the context window. Tool count is a marketing number, not a quality signal. Bybit ships 206 tools; that is impressive engineering and also a lot of surface area for a model to get lost in.
Beta means beta. Nine of the servers here are labelled beta, early access or "under active development" by the vendors themselves. Plesk's own announcement says to expect bugs and not to use it in production. That is admirable candour, and you should take it literally.
Permissions are inherited, not designed. A server usually gives the model whatever the authenticated user can already do. That is the right default, but it means an over-permissioned account becomes an over-permissioned agent. Gusto's advice to run its server in an isolated session — away from other MCP servers — is the sanest security note any vendor in this list publishes, and it applies well beyond Gusto.
Read is solved; write is not. Most of the value shipped so far is retrieval. Navan is read-only with writes "coming"; Gusto is read-only by design; ClickUp won't delete. The genuinely agentic use cases are still early, and the vendors being cautious are, we think, the ones getting it right.
MCP (Model Context Protocol) is an open standard for connecting an AI assistant to a tool or data source. The AI app is the client, the tool ships a server, and the two speak one protocol. It means a tool builds one integration instead of one per AI app, and you can plug your own data into the assistant you already use.
The protocol itself is open and free. Whether a vendor's MCP server is free depends on the vendor. Of the 35 partner servers we checked, 14 state a plan or price in their own docs and 21 say nothing at all. Where a plan is stated, the MCP server is usually included with the subscription you already pay for rather than sold separately — the real limits are rate limits and whether writes are allowed.
Increasingly no. Remote servers are the norm now: you paste a URL into your AI client, approve an OAuth screen, and you're connected. Local servers still need a runtime — our own is a single Python file — but the setup is a copy-paste, not a project.
Treat it like any other integration: it inherits your permissions, so scope it. Several partners are deliberately conservative — Gusto's server is read-only, Navan's is read-only for now, ClickUp refuses delete operations, and Circle requires per-action approval for writes. Gusto also advises running its server in an isolated client session rather than alongside other MCP servers.
We publish one: the Honest Atlas MCP, a free server that gives any MCP client our hand-researched dossiers on 300 SaaS and AI tools — real pricing notes, the honest knock, who should skip it. It runs on your machine, needs no account and has zero dependencies.
How this page was made: we checked all 300 tools in our partner database against their own documentation on 2026-07-28, and cite the vendor page for every claim. Plan and price wording is theirs, not ours; "not stated" means we could not find it published on their own site. Illustrations are our own SVGs; partner logos are hosted on our own servers, not hotlinked. Some links are affiliate links — we may earn a commission if you sign up, at no extra cost to you, and it never changes what we write: a tool that is gated or expensive gets said so on this page. Review content lives in our dossiers, which this page links to rather than repeats. How we review →
New dossiers, cost-traps we found, and what actually changed in AI tooling — no hype, no sponsored-disguised-as-advice. Unsubscribe anytime.