# MCP platform for small companies: what to buy

> What an MCP platform for small companies needs, what it can skip, how pricing lands at 50 and 200 seats, and when to buy nothing.

**TL;DR** An MCP platform for small companies needs four things: one endpoint every AI client signs into, per-person OAuth sign-in, a handful of restrictions, and an audit log. Elaichi provides those on the Gold plan, and SCIM, custom roles and region choice come with it rather than as a separate enterprise tier. A team of five with no compliance obligation and three local MCP servers should buy nothing yet.

## The decision in front of you

It is Tuesday. Four people in sales are using Claude with their own HubSpot logins pasted in somewhere. Two engineers run local MCP servers configured in `~/.cursor/mcp.json` with API keys in plain text. Finance asked whether ChatGPT can read Stripe. You have a quarter to pick something, and the vendor pages are all written for a company with a platform team. Elaichi puts those calls behind one MCP address, each person signing in with their own account, and records every call.

An MCP platform for small companies is a narrower problem than the enterprise buyer's guides suggest. MCP, the Model Context Protocol, is how an AI client finds and runs tools from other systems. At 20 to 200 people you need four capabilities. This post names them, names what defers, and works through what per-seat and per-call pricing do to a 50-seat bill and a 200-seat bill.

Two kinds of claim show up below. General ones hold on any MCP platform. Elaichi-specific ones name its exact plans and prices. The general claims are the test to run on anyone's shortlist.

## What does a 50-person company actually need?

Four things, true of any MCP platform, not just this one: one endpoint, per-person sign-in, a few restrictions, and an audit log. Everything else on a feature matrix is optional at that size.

**One endpoint.** Elaichi serves MCP at `https://api.elaichi.ai/mcp`, the same string for every organization. There is no per-user URL, no per-team URL and no MCP server for anyone to create or revoke. Which organization and which person a client reaches comes from the OAuth grant behind the token, not from the address. On any platform, ask whether revoking one person means editing one record or every client config that person touched.

**Per-person sign-in.** Each member connects their client once and approves their own grant. An OAuth grant is the standing permission a person gives one named client on Elaichi's consent screen. Every call made with that grant acts as the person who approved it, and their permissions are looked up on each request rather than frozen into the token. A sales rep's Claude reads HubSpot as that rep. Nobody is sharing a service account. In Elaichi, removing or suspending a member revokes every live grant in the same transaction as the membership change. That is the number to ask any vendor for. "We support SSO" says nothing about how fast access actually stops.

**A few restrictions.** A restriction in Elaichi targets a role or one member, and names whole connectors or single tools. Three or four rules cover most of what a 50-person company worries about. Block the Stripe refund tool for everyone outside finance. Block Gusto's write tools outside HR. Restrictions are checked on Elaichi's side before a call leaves, on every path that runs a connected tool: `tools/call`, `execute_tool`, `read_tool`, `run_code`, toolbox execution and each step of a synthetic tool. They match the operation a tool performs rather than the label it is shown under, so renaming a tool does not slip it past a block. [MCP tool allowlist vs blocklist](/blog/allowlist-vs-blocklist-mcp-tools/) covers which of the two shapes to reach for first.

**An audit log.** Every tool call that runs through Elaichi writes a record containing one entry for each connected-tool call that reaches execution (a call refused earlier writes none). Elaichi keeps 90 days of it, readable in the console and exportable as CSV. The in-product audit log is not plan-gated. The general version of this test: ask whether the audit log is a paid add-on or whether it ships on the plan you would actually buy. A log you have to upgrade into after an incident is not a log.

## What can you skip until 200 seats?

SCIM, custom roles and a deliberate region choice. Not because they are bad, but because at 40 people the manual version takes less time than the automated version takes to configure.

**SCIM provisioning.** SCIM is the standard that lets your identity provider create and deactivate accounts in another system automatically. It earns its setup when joiners and leavers outpace an admin's attention, which is somewhere north of a hundred people. In Elaichi, SCIM deprovisioning suspends a member rather than removing them, and suspension revokes their grants. Suspending by hand does the same thing. On any platform, ask whether automated deprovisioning actually cuts off live access, or just archives the account record.

**Custom roles.** Elaichi ships eight system roles: Org Owner, Org Admin, People Admin, Team Admin, Member, Guest, Billing Admin and Auditor. Most of a 50-person company fits in Member, with two admins and one Auditor. A member holds exactly one role, so a custom role is how you combine two personas, and that need usually arrives with the second department that does not fit the pattern. Guest, Billing Admin and Auditor are free seats on Elaichi. Confirm that on any platform, since some bill every named account regardless of role.

**Data residency.** Elaichi has three regions, EU, US and APAC, chosen when the organization is created and not changeable later. That is the one deferral with a catch: pick it on day one. For an EU or US organization, the per-organization data store is pinned to that jurisdiction. APAC uses best-effort placement, which is not a residency guarantee. If your buyers are asking residency questions already, read [MCP data residency: self-hosted vs managed](/blog/self-hosting-mcp-for-data-residency/) before you sign anything.

SCIM, group-to-role mappings, custom roles and SSO over SAML or OIDC are all on Gold, the same plan a 30-person company buys. You defer configuring them. You do not defer paying for them. Several competitors put SCIM and SSO behind a separate "Enterprise" tier priced on request, a materially different bet for a 50-person company that might need them at 150.

Composio is one platform to check this on. Its pricing page lists SSO/SCIM and customer-managed keys under the Enterprise plan, priced on custom terms. Its plan comparison marks both as Enterprise features rather than ones included on Hobby or Pro ([Composio pricing](https://composio.dev/pricing), checked October 2026).

## What belongs on your shortlist at each size?

The four capabilities below are the whole test at 20 to 200 people. Everything else on a vendor matrix is a question for later.

| Capability | Needed at 20 to 50 | Needed at 50 to 200 | Where Elaichi puts it |
| --- | --- | --- | --- |
| One endpoint every client signs into | Yes | Yes | Every plan, including free Personal |
| Per-person OAuth sign-in | Yes | Yes | Every plan |
| Restrictions on a role or a member | Yes | Yes | Gold and Black |
| Audit log of tool calls | Yes | Yes | Every active plan, not plan-gated |
| Sharing a connection or toolbox | Often | Yes | Gold and Black |
| Custom roles | Rarely | Often | Gold and Black |
| SCIM provisioning and group mappings | No | Often | Gold and Black |
| Audit-log forwarding to Datadog | No | Sometimes | Black |

The column that matters on a competitor's page is the last one: which tier each capability sits behind. Two platforms can both tick restrictions and audit and still differ by a tier boundary that doubles your bill at 120 seats.

### Five questions to ask any vendor, not just this one

1. Is the audit log on the plan I'm buying, or does it require an upgrade?
2. When I suspend or remove a member, how long until their existing grants stop working: same request, same minute, next day?
3. Are SCIM, SSO and custom roles on the plan I'd buy at my current size, or on a separate enterprise tier priced on request?
4. Which roles are free seats, and which count against my billable headcount?
5. If I pin a shared credential to a team tool, what happens to that tool when the person who pinned it leaves?

A platform that cannot answer all five in its own docs, without a sales call, is not ready for a 50-person buyer. The same five questions worked through against one named vendor are in [this comparison of Composio alternatives](/blog/elaichi-vs-composio/).

## Which clients will your people actually use?

Claude, ChatGPT, Cursor and Microsoft Copilot, and probably all four at once. The cost of covering them should be one setup, not four.

With Elaichi, each points at the same `https://api.elaichi.ai/mcp` and signs in over OAuth. Claude Team and Enterprise Owners add it once from Organization settings, then Connectors, and each member then connects and authenticates individually. That path is in [Anthropic's own documentation](https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp) (checked October 2026). ChatGPT Business, Enterprise and Edu admins add a custom MCP app from Workspace settings, then Apps, then Create, give the endpoint, scan tools, complete the OAuth prompt, then publish. That path is in [OpenAI's developer mode page](https://help.openai.com/en/articles/12584461-developer-mode-and-mcp-apps-in-chatgpt) (checked October 2026). Cursor reads remote servers from `.cursor/mcp.json` in a project or `~/.cursor/mcp.json` in the home directory, with the server defined by a `url` field, per [Cursor's MCP docs](https://cursor.com/docs/mcp) (checked October 2026). Microsoft Copilot reaches Elaichi through a Copilot Studio agent that adds the endpoint as a tool with OAuth set to Dynamic discovery.

Elaichi keeps no list of approved clients. Any client that supports OAuth dynamic client registration, PKCE with S256 and a browser sign-in can connect. The whole-company version of this is in [how to roll out Claude and ChatGPT to employees](/blog/roll-out-claude-and-chatgpt-to-employees/).

One honest limit shapes how you think about client-side controls. No tool in Elaichi's MCP list names a connected app. The model finds a connected tool with `search_tools` and runs it with `read_tool` (a read) or `execute_tool` (anything), or with `run_code` under Code Mode. Search results say which of the two to use. So per-tool client permissions see Elaichi's connected tools as a few tool names, `execute_tool`, `read_tool` and `run_code`, and cannot tell one connected app from another. Gating `execute_tool` alone leaves `read_tool` and `run_code`. Per-app and per-tool control is set in Elaichi, as restrictions, not in the client's settings screen.

## What kinds of connector does the catalog hold?

Two kinds. Elaichi authors and runs most of them itself. The rest are native MCP connectors, where the vendor builds and runs its own MCP server and Elaichi handles sign-in, access and audit in front of it. Both kinds come through the same address, the same `search_tools`, `read_tool` and `execute_tool`, and the same restrictions and audit log.

The catalog runs to 650+ connectors. For a native MCP connector such as Linear or Datadog, the vendor runs the server and owns its tools; Elaichi does not write, test or fix them. What Elaichi adds is that calls through it are checked against the caller's restrictions and scopes before they leave, and every one that runs is written to the audit log with the person, the connector, the connection and the MCP client.

There is one more case. GitHub is not in Elaichi's catalog. An organization on Gold can add GitHub's own remote MCP server as its own remote MCP connector, which puts it under the same rules. Calls a client makes to a vendor's server directly, outside Elaichi, are not governed by it. That is the general gap to check on any gateway: does it govern everything a client connects to, or only what is in its catalog? [Vendor MCP servers vs single control plane](/blog/vendor-mcp-servers-per-saas-vs-one-plane/) works through when each shape wins.

## How do you share access without sharing credentials?

Through a toolbox. A small company usually has one person who holds the account that matters: the ops lead with the billing login, the founder with the Notion admin seat. Toolboxes are how that access reaches a team without the password going anywhere.

A stored toolbox is a named list of entries, each one a connector tool pinned to a connection. The person who pins a connection vouches for it: they must be able to use it, and Elaichi records them as that entry's delegator. Share the toolbox at the use level, and each grantee's call runs through the pinned connection under the delegator's authority, not through any connection of the grantee's. One person connects, everyone else runs the tools, and nobody else holds the credential.

Two consequences to plan for. Restrictions are still evaluated for whoever is calling, so a tool blocked for the Member role stays blocked inside a shared toolbox. And if the delegator leaves the organization, every entry they pinned resolves as unmet for everyone until someone with edit access re-pins it. [Offboarding AI access, contractors included](/blog/offboarding-when-the-agent-holds-access/) covers the order to do that in.

## How does per-user pricing compare with per-call at 50 and 200 seats?

Per-seat pricing is predictable and scales with headcount. Per-call pricing is cheaper when usage is light and unbounded when a developer's agent starts looping. At 50 seats the two are often close. At 200 they diverge, and the direction depends on who is using it, not how many.

**The seat math.** Gold lists at $15 per user per month in USD, or $10 per user per month billed annually. [The pricing page](/pricing/) shows a visitor their own country's price. Your monthly bill is that seat price times your billable seats, and headcount is the one figure you can count on day one. Guest, Billing Admin and Auditor are free seats, so a compliance reviewer and a bookkeeper cost nothing.

**The per-call break-even.** Divide the seat price you would pay by the per-call price the other platform publishes. That gives the number of tool calls per person per month at which the two bills are equal. Below that number, per-call is cheaper for that person. Above it, per-seat is. Then ask who on your team sits on each side of the line. An occasional user asking Claude about a deal makes a handful of calls a day. An engineer running an agent in Cursor can trigger hundreds in a single session: searching files, running tests, checking a ticket, re-running after a failed diff. In a 200-person company where thirty people run agents daily, the per-call bill stops being predictable from headcount. It becomes predictable from behavior you cannot see until the invoice arrives. Get each vendor's per-call price from its own pricing page before you run the division.

Elaichi's Black plan does bill usage on top of seats, at $0.02 a run minute and $5 a GB-month past a monthly allowance of 1,000 run minutes and 1 GB. That is for workflows and the Elaichi Agent, which a company at this size can skip. Gold has no usage billing. [MCP gateway pricing: per seat vs per call](/blog/mcp-gateway-pricing/) extends this model with more usage scenarios.

## When is buying nothing the right answer?

When you have fewer than ten people, no compliance obligation anyone has written down, and two or three MCP servers that one person set up and maintains.

A five-person engineering team running local MCP servers against their own repos is not a governance problem. Everyone already has access to everything. There is no auditor asking who ran what, no sales rep with a CRM credential that should expire, and no second department whose access should differ. A control plane would add a vendor and a monthly bill in exchange for solving a problem you have not got.

The signals that it has stopped being the right answer, in roughly the order they arrive:

- A second department starts using AI against a system the first department should not reach.
- Somebody leaves, and nobody can say which laptops still hold a working credential.
- A customer sends a security questionnaire asking which AI tools touch their data.
- An engineer pastes an API key into a config file that lands in a repo.
- Finance asks for read access to a system where a wrong tool call moves money.

One of those is enough. Two is late. [When you don't need an MCP gateway yet](/blog/when-you-dont-need-an-mcp-gateway/) argues the other side of this at length.

There is also a middle answer. Elaichi's Personal plan is free, with no end date: one member, 10 connections across at most 3 connectors, 1,000 connected-tool calls a calendar month. It cannot share anything, which is the point at which it stops being enough. One person can connect their own apps and see what the catalog does before a team buys. 14-day Black trial, no credit card to start. Buying Gold sets the paid trial to the remaining days rather than granting a fresh 14, so it is one continuous trial rather than two, while buying Black charges immediately. A workspace that buys nothing moves to the free Personal plan.

## Picking an MCP platform for small companies this quarter

Test four things, in this order, on whatever you shortlist. These are checks any MCP platform should pass, not an Elaichi-specific script:

1. Add the endpoint to Claude or ChatGPT yourself and complete the OAuth sign-in. If that takes longer than fifteen minutes, every one of your colleagues will need help.
2. Have a second person connect and confirm their calls act as them, not as you. Check it in the audit log.
3. Write one restriction, blocking a single write tool for the Member role, and confirm the refusal in a chat. In Elaichi a restriction change takes effect about two minutes, so give it that long before you decide it did not work.
4. Export or read the audit log for the calls you just made, and check it names the person, the tool, the connection and the client.

If all four pass, the platform does the job a company your size needs. If one fails, it will fail the same way in front of an auditor eighteen months from now.

The broader buying frame, including the questions that only matter above 200 people, is in [Enterprise MCP: a buyer's guide for IT teams](/blog/enterprise-mcp/). For what to configure in the first five days after you sign, [the first-week checklist](/blog/first-week-governance-checklist/) is the shorter read. If your shortlist still has four names on it, [best MCP gateways for company-wide AI access](/blog/best-mcp-gateways/) sorts them by what each one is built for.

## FAQ

### What is the minimum an MCP platform needs for a 50-person company?

Four capabilities: one MCP endpoint that every AI client signs into, per-person OAuth sign-in so each call acts as the person who made it, restrictions that can block a tool or a whole connector for a role, and an audit log that records who ran which tool through which connection. Elaichi provides all four, and its in-product audit log is available on every active plan, including the free Personal plan. Features such as SCIM provisioning and custom roles matter later, usually past 100 to 200 people.

### Do you need SCIM provisioning at under 100 employees?

Usually not. Below about a hundred people, inviting and suspending members by hand takes less time than mapping directory groups to roles. The thing SCIM buys you is automatic deprovisioning, and in Elaichi a suspension, whether made by SCIM or by an admin, revokes that member's MCP OAuth grants so their AI clients stop on the next call. An admin suspending someone manually gets the same result. Note that SCIM deprovisioning in Elaichi suspends a member rather than removing them from the organization.

### How much does an MCP platform cost at 50 seats versus 200?

Gold lists at $15 per user per month in USD, or $10 per user per month billed annually. See [the pricing page](/pricing/) for the price in your own country. Your bill is that seat price times your billable seats, so it scales straight with headcount. Guest, Billing Admin and Auditor are free seats and do not count toward the bill, and suspended members never count. Regional prices are lower in many countries.

### When should a small company skip buying an MCP platform entirely?

When it has fewer than about ten people, no written compliance obligation, and two or three local MCP servers that one person set up and maintains. If everyone already has access to everything and no auditor is asking who ran what, a control plane adds a vendor and a bill without solving a problem. The signals that this has changed: a second department needing different access, somebody leaving with credentials on a laptop, a customer security questionnaire about AI tools, or finance wanting read access to a system where a wrong tool call moves money.

### Can one MCP endpoint serve Claude, ChatGPT and Cursor at the same time?

Yes. Elaichi serves MCP at one address, https://api.elaichi.ai/mcp, and every client signs in there over OAuth. Which organization and which person a client reaches comes from the OAuth grant behind its token, not from the URL, so there is no per-client or per-team address to manage. Claude, ChatGPT, Cursor and Microsoft Copilot through a Copilot Studio agent all connect to the same string; what differs is only where each client's own settings take the URL.

### Can a team share one person's app connection without sharing the password?

Yes, through a toolbox, on the Gold or Black plan. One person connects their own account, pins it into a toolbox entry, and shares that toolbox at the use level. Each teammate's calls then run through the pinned connection under that person's authority, and nobody else holds the credential. Restrictions are still checked against whoever is calling. If the person who pinned the connection leaves the organization, the entry stops resolving for everyone until someone with edit access re-pins it.

## Read next

- [Composio alternatives for governed employee AI](/blog/composio-alternatives/) — How to compare Composio alternatives on per-person sign-in, tool-level restrictions, audit logs and pricing, and when Composio is still the right pick.
- [Runlayer alternatives: three questions to ask](/blog/runlayer-alternatives/) — Compare Runlayer alternatives on where connectors come from, how access rules are written, and whether every AI client signs into one address.
- [Automation platform vs MCP access: which job?](/blog/n8n-automation-vs-mcp-access/) — Scheduled workflows and ad hoc assistant calls are two jobs. How to split automation platform vs MCP access, and when one tool covers both.