Six support platforms publish an MCP server, a CLI or an API that an AI agent such as Claude Code, Cursor or ChatGPT can use, but only two document a path where the agent starts the account and then configures the AI agent itself: Fin (now part of Salesforce), through its Fin CLI, and Lorikeet, through a keyless sign-up endpoint plus a product MCP server. Plain, Front and Freshdesk run MCP servers built for working the queue, not for setting up an AI agent, and Zendesk documents an MCP client, which points the other way.

This guide compares what each vendor's own documentation says on 30 October 2026: whether your agent can sign you up, which credential it uses, and what it can change once connected.

Key takeaways

• Check the direction first. An MCP server lets your agent reach the platform. An MCP client lets the platform's AI reach other systems. Zendesk's August 2026 release is a client.

• Sign-up and setup are different jobs. Fin's CLI can create a new workspace from the terminal. Lorikeet's endpoint lets an agent start the sign-up, but the person finishes it in their own browser and adds their own card.

• Read the credential model. OAuth per person (Plain, Front, and the interactive MCP flows) ties every action to a named user. An API key (Freshdesk, the Fin CLI's token path) gives its permissions to whoever holds it.

• Region limits are real. The Fin CLI supports US workspaces only, and Intercom's MCP server does not yet support AU-hosted workspaces.

• Ask what the agent can write. Tool lists differ: some only add internal notes, some send customer replies, some edit the AI agent's workflows.

Which tools let an AI agent set up AI customer support?

Lorikeet lets an AI assistant start a 30-day free trial over plain HTTP with no API key: a POST to its agent sign-up endpoint emails the person a sign-up link, and after they finish in their browser, its OAuth-protected MCP server lets Claude, ChatGPT or Claude Code build workflows and run simulations, as described in its agents.md and auth.md. Fin documents a Fin CLI built for coding agents, with an auth signup command that creates a new workspace. Plain's MCP server gives each person's AI client the same permissions as their Plain user, with no API keys needed. Front's MCP server authenticates agents as a specific Front user over OAuth.

How we chose and checked these tools

We included a platform only if its own public pages document at least one of these: a remote MCP server an outside AI client can connect to, a command-line tool for coding agents, or an endpoint an agent can use to start a sign-up. We then read each vendor's developer docs, help centre and pricing page on 30 October 2026 and recorded what they say about authentication, region limits, what the agent can write, and the free trial. Where a vendor does not state something, the table says "Not published". We read Plain's vendor round-up as a pointer to sources, then checked every fact on the vendor's own page.

Lorikeet publishes this guide and is one of the tools listed. We have tried to describe every product, including our own, using only what each vendor publishes, and we link to the source for each claim.

Entries are ordered by how far an agent can get on its own, from sign-up to a configured AI agent, not by overall product quality.

Comparison: MCP, API and agent-led setup

1. Fin (now part of Salesforce): the most complete agent-led setup today

Fin, which Salesforce completed its acquisition of on 10 September 2026 and which was formerly Intercom, documents the longest chain from nothing to a configured AI agent. The Fin CLI (@intercom/cli) is described as built for developers and AI coding agents such as Claude Code, Cursor or Codex. Its auth signup command creates a brand-new Intercom workspace, intercom setup installs the Messenger and populates the help centre in one step, and intercom fin manages procedures, guidance, audiences, data connectors, simulations and workflows. A --plan flag lists what the CLI would do before it does it.

Separately, the Intercom MCP server gives AI clients 14 tools for conversations, contacts, companies and help centre articles, plus five Fin Operator tools for workspaces that have Operator, including one that applies Operator's proposed changes to live configuration. A third surface, the Fin Agent API MCP server, lets your agent call Fin as a tool rather than configure it.

Best for: US-hosted teams that want a coding agent to stand up a help centre, Messenger and AI agent in one session.

Limits: the CLI supports US workspaces only, tokens expire after about 14 days with no automatic refresh, and the OAuth browser flow is not yet supported. The MCP server's note tool cannot send customer-visible replies, and AU-hosted workspaces are not supported. Pricing starts at $29 per seat a month for Essential, with Fin at $0.99 per outcome and a 14-day trial with no credit card.

2. Lorikeet: agent-started sign-up and a product MCP server for complex workflows

It documents two separate doors. The first is sign-up: per its agents.md, an AI agent calls GET on the agent sign-up endpoint to learn what to ask, confirms the person's work email and name, and sends one POST. The person gets an email with a link a link valid for 7 days; they create the workspace and add a card at checkout to start the 30-day free trial, and nothing is charged until the trial ends. The agent never receives a credential and never pays. Personal mailboxes are rejected.

The second door is configuration. The product MCP server, listed in its server card, uses OAuth 2.1 with scopes from manage:read to manage:write and lab:write. Once connected from Claude, ChatGPT or Claude Code, an agent can build and edit workflows, manage knowledge, test tools and run simulations. Its engineering post says every MCP action is logged with the agent's stated reason, and that deleting a tool in use by a workflow fails. Firm360 runs analysis against the MCP server every two weeks to rank knowledge base gaps with draft answers, per its customer story.

Best for: regulated or complex support teams that want an agent to build and test multi-step workflows on top of Zendesk, Intercom, Salesforce or another helpdesk.

Limits: an agent cannot finish sign-up alone, because the person must open the email and add a card. The ChatGPT connection needs a Team, Business or Enterprise workspace. Plans on the pricing page start at $2,100 a month (Start) or $5,100 a month (Scale), paid annually, plus $0.99 or $0.90 per resolved chat or email ticket, which is more than a seat-priced helpdesk for a small team.

3. Plain: per-person MCP access for B2B support queues

Plain's MCP server docs say the server has the same permissions as your user, with no API keys needed, and that replies to threads appear as though they come from that user. Plain's own round-up lists 43 tools as of 28 October 2026, 29 read and 14 write, covering threads, tasks, knowledge, customers, tenants and help centre articles. Each person completes OAuth once in their own client, so an admin's connection does not hand admin access to anyone else.

Best for: B2B teams who want engineers and account managers to pull support context into Claude, ChatGPT or Cursor under their own role.

Limits: Plain says its documented list has no tool for building workflows, so configuring the Customer Agent through MCP is not on offer today; the GraphQL API is the fallback. Plain's docs also note that Claude Code does not yet refresh OAuth tokens, so they recommend the mcp-remote helper. The pricing page lists plans from $45 per user a month and a 7-day free trial.

4. Front: user-scoped MCP with admin-set limits

Front's developer docs describe an MCP server at mcp.frontapp.com/mcp where agents authenticate as a specific Front user, with permissions matching that user's role. Admins choose read, write and send permissions on the OAuth app, and Front advises requesting the narrowest set: read for a research agent, read and write for a drafting copilot, and send only if the agent will send messages. Claude and ChatGPT can be installed from official directories without creating your own app.

Best for: shared-inbox teams that want drafting copilots with a clear ceiling on what an agent can send.

Limits: Front says it does not support Dynamic Client Registration, so your AI assistant must send an OAuth client ID and secret, and some popular assistants are not yet compatible. Front's docs do not describe agent-led sign-up or AI agent configuration over MCP. The pricing page offers a 14-day trial of the Professional plan with no credit card.

5. Freshdesk: MCP generally available, with metered actions

Freshworks' help article says its MCP server became generally available on 10 September 2026 and lets tools like Cursor and Claude connect to Freshdesk to do things such as create tickets, using public APIs. It gives setup steps for Claude Code, Claude Desktop, Cursor, Microsoft Copilot Studio and VS Code.

Best for: existing Freshdesk teams that want ticket work from an AI client without building an integration.

Limits: authentication is by API key only, custom domains are not supported for the MCP endpoint, and each successful tool call counts as an action against plan allowances and rate limits. The article covers specific plan versions only. The pricing page mentions a 14-day free trial.

6. Zendesk: an MCP client and a broad REST API

Zendesk's 10 August 2026 announcement made its MCP client generally available: admins connect Zendesk to public or private MCP servers and choose which tools AI agents, auto assist and custom agents can use in action flows. That helps Zendesk's AI reach your systems. It is not a server your own agent connects to. For agent-led setup, the route is the Zendesk API reference, which covers ticketing, help centre, AI agents, messaging and voice.

Best for: teams already on Zendesk who want their Zendesk AI agent to call tools in Stripe, GitHub or private systems.

Limits: the MCP client needs Suite Growth or Support Team and above, and Zendesk does not document a vendor-maintained MCP server or agent sign-up flow on the pages we checked. The pricing page describes a 14-day trial on Suite Professional.

What does an AI customer support platform with an API or MCP actually need to support?

At minimum it needs standard authorization your agent can complete. The MCP authorization specification says authorization is optional, but where it is supported, MCP servers must implement OAuth 2.0 Protected Resource Metadata (RFC 9728) so a client can find the authorization server, and clients and servers should support Dynamic Client Registration (RFC 7591). Dynamic registration is why Claude can add some servers by URL alone, and why Front needs a client ID and secret first. Beyond auth, look for tools that write the things you would otherwise click through: workflows, knowledge, test runs.

What should a human still approve when an agent sets up support?

Anything that reaches a customer or spends money. OWASP's LLM06 Excessive Agency names excessive functionality, permissions and autonomy as root causes and recommends human approval for high-impact actions. In support setup that means a person should:

• enter payment details and accept the contract (the agent sign-up flow above requires this by design);

• approve OAuth scopes, starting read-only and adding write access only for the job;

• review workflow changes and simulation results before an AI agent goes live on real customers;

• own escalation rules, refund limits and any compliance sign-off.

Use dry runs where they exist, such as the Fin CLI's --plan flag, and test in a separate environment before switching on.

How to choose

• You want an agent to do everything, US-hosted, on a seat-priced helpdesk: start with Fin's CLI.

• You keep your helpdesk and need multi-step, tested workflows in a regulated business: Lorikeet's sign-up endpoint and MCP server fit, at a higher entry price.

• You want staff to work the queue from their own AI client: Plain or Front.

• You are staying on Freshdesk or Zendesk: use Freshdesk's server, or Zendesk's API and MCP client.

For the sign-up side in more depth, see AI customer support you can sign up for from ChatGPT or Claude; for a same-day build, AI support agents you can set up yourself in a day; for security review, how a CISO or CTO approves an AI support agent; and for the customer-facing side of agent traffic, WebMCP vs public endpoint vs llms.txt. To try the agent sign-up flow yourself, start a 30-day free trial.