Pylon MCP server, its tools and how to connect

Pylon, the B2B support platform at usepylon.com, runs its own MCP server at https://mcp.usepylon.com, or https://mcp.eu.usepylon.com for a workspace hosted in the EU. You add that address to Claude, ChatGPT, Cursor or another AI tool and sign in with your Pylon account, since it takes no API key. The AI tool then works as you, through 109 tools for issues, accounts, contacts, the knowledge base, triggers and more. One of them replies to your customers, and a reply can’t be unsent. You need a Member or Admin seat on Professional, Advanced or Enterprise, and Pylon sells every one of those plans through its sales team.

Everything here comes from Pylon’s docs, help centre, pricing page and changelog, Claude’s and OpenAI’s setup guides, and the GitHub repos of two community-made servers, read on 10 November 2026. Each fact links to where we read it.

We make Deacon, an AI agent that replies to customers from your docs through a chat on your website, and you can sign up and start on its free plan yourself. Its MCP server does a job Pylon’s own help article suggests for Claude Code, getting from a customer’s report to a fix in your code, and the end of this page shows how.

How to connect Pylon’s MCP server

Start in Pylon. Open Settings › AI Controls › MCP Server and copy your workspace’s address from Installation, as Pylon’s docs say. Pylon’s integration page says that’s also where the server is configured and turned on. Every example in Pylon’s help article uses the US address, so if your workspace is in the EU, put https://mcp.eu.usepylon.com in its place in the steps below.

Then add it to your AI tool. The first four ways are the ones that help article gives.

In Claude Code, run this, then run /mcp and pick pylon to sign in.

claude mcp add pylon --transport http https://mcp.usepylon.com/

In Claude’s desktop app or on the web, Pylon has you add a custom connector named Pylon with the address https://mcp.usepylon.com/, leave OAuth Client ID and OAuth Client Secret blank, and press Connect next to it. Its article puts that in Admin Settings, which is the route on Claude’s Team and Enterprise plans. There, Claude’s help centre says, an Owner adds it under Organization settings › Connectors, and each person then presses Connect in Customize › Connectors. On Pro or Max you add it yourself, in Customize › Connectors with + Add and then Add custom connector. A free Claude account can have one custom connector.

In ChatGPT, search the Apps tab for Pylon and install Pylon’s own app. That app leaves out the four REST API tools, which a custom connector you add yourself includes, according to Pylon’s docs.

In Cursor, add Pylon to ~/.cursor/mcp.json, then open Cursor Settings › Tools & MCP and press Connect next to it.

{
  "mcpServers": {
    "pylon": {
      "url": "https://mcp.usepylon.com/"
    }
  }
}

Codex isn’t in Pylon’s article, but Pylon’s integration page says the server works with any MCP-compatible client, and Pylon’s sign-in settings let a tool register itself, so there’s no client ID to make first. These two lines, in the form OpenAI’s Codex docs give, add it and open the sign-in.

codex mcp add pylon --url https://mcp.usepylon.com/
codex mcp login pylon

Each time, your browser opens on Pylon. You log in, authorise the connection, and the browser sends you back to the AI tool, connected.

The seat and plan Pylon’s MCP server needs

Only Members and Admins can connect, since they’re the roles with MCP Access, according to Pylon’s roles page. A Viewer or an Integration User who tries gets an “Authorization failed” error, and Pylon’s help article says the fix is an admin upgrading their seat in Settings › Users. If your workspace uses custom roles, MCP Access is one of the Developer permissions a role can have, and a teammate on that role needs it switched on to connect.

Pylon’s pricing page ticks the Pylon MCP server on all three of its plans, Professional, Advanced and Enterprise, and Pylon pricing has what each one costs. Some groups of tools need more than the plan.

• Tasks, projects, account notes and call recordings need Account Intelligence, an add-on, and Pylon’s pricing FAQ says it needs at least 50 customer accounts.
• The nine feature request tools need the Feature Requests plan feature, and the pricing page lists Pylon’s feature request features on Advanced and Enterprise, not Professional.
• The AI agent tools, which test Pylon’s own support agents against past issues, need AI Agents and Backtesting, and the pricing page says backtesting uses credits on every plan.
• The two audit log tools need an Admin, and below Enterprise they only search the last 30 days.

You can’t sign up and try it on your own. Every plan on the pricing page leads to Book a demo or Talk to sales, its FAQ says “There is a minimum seat count for new customers”, and a free trial is a demo organisation your sales rep sets up. Pylon says it’s “built for B2B support teams rather than solo operators”.

Every tool in Pylon’s MCP server

Pylon’s docs list 109 tools, 65 that read and 44 that change something. Each one runs as you, so it can only see and change what you can in Pylon. Pylon’s help article adds that the tools available can vary by organisation and by feature flag, so one listed here may not show up in your workspace. Every tool also needs the Pylon permission for what it does, and the last column adds anything more Pylon asks for.

For everyday support work, start with search_issues to find conversations, get_issue_messages for a whole thread, send_issue_note for a note only your team sees and send_issue_reply for the customer. Pylon’s changelog for 9 November says MCP search now reads an issue’s body text as well as its title. update_issues makes the same status, assignee, team or tag change on up to 100 issues in one call.

If you fix bugs with a coding agent, use link_product_issue and search_issues. link_product_issue ties a Pylon issue to a Linear, Jira, GitHub, Asana or Shortcut ticket, and every search_issues result lists its linked engineering tickets and their state. Pylon’s help article suggests asking Claude Code to “debug this” with a Pylon link and open a pull request with the fix.

What to watch for when you set it up

A reply goes straight to the customer

send_issue_reply sends on the issue’s own channel, whether that’s Slack, email or chat, and Pylon’s docs say it can’t be unsent. On an email issue it replies to everyone on the thread. Pylon tells AI tools to show you the exact message and get your approval before sending, and most AI tools ask you to confirm each one as well. Keep that prompt switched on, because it’s the only human in the loop between the AI tool and your customer.

It only signs in through a browser

The server takes OAuth and nothing else. Pylon’s help article says it doesn’t support API keys or machine-to-machine credentials, and a client that can’t finish the browser sign-in “cannot connect directly to the Pylon MCP server today”. So a scheduled job, a GitHub Action or an agent running headless can’t use it. Pylon points those at its REST API instead.

The sign-in also runs out. Depending on the AI tool, a session lasts a few hours or several days, and then the tools start failing or return nothing. Sign in again the way you first did, with /mcp in Claude Code, codex mcp login pylon in Codex, or the Connect button in Claude’s connector settings or Cursor Settings › Tools & MCP. Pylon says that’s normal, and to contact its support if you’re reconnecting several times a day.

It can only do what your Pylon role can

Everything runs with your own permissions. Pylon’s roles page says Members can’t delete records or publish knowledge base articles, so a Member’s AI tool can save a draft with save_draft_article but can’t publish it. An Admin’s connection can delete, because an Admin can.

Smaller catches, from deletes to old tool names

• delete_feature_request, delete_tag and delete_macro delete for good, and deleting a tag takes it off every issue, account and article it was on. The last two arrived on 2 November, according to Pylon’s changelog. Macros are Pylon’s canned responses, and none of the macro tools sends one.
• Merging two contacts with save_contact permanently deletes one of them, and it can’t be undone.
• A trigger made with save_trigger starts switched off, so you can check it in Pylon before it runs.
• When search_issues comes back empty for an account or a person’s name, read its warnings. Pylon says an unmatched name only means the lookup found no record, and there may still be issues.
• Pylon has folded many tools into the save_ ones in the table. Old names such as create_contact, create_trigger and create_project no longer work, and create_task and update_task still run but no longer show in the tool list, as Pylon’s docs set out. A prompt, script or guide written before the change may name tools you won’t find.
• Calls are limited per tool and per organisation, and a tool that hits its limit returns error -32000 with the seconds to wait. Pylon doesn’t give those limits as numbers.

What Pylon’s API does that the MCP server doesn’t

Pylon’s server can call the API itself, so it can do most of what the API does. Its four REST API tools find any operation in Pylon’s REST API with search_api_operations, read its arguments with get_api_operation, and run it with call_api_read or call_api_write, always as you (MCP docs). Reads and writes are separate tools, so you can have your AI tool ask before writes only. Writes apply at once, can be visible to customers and have no idempotency key, so when you can’t tell whether one went through, read the record before trying again. A read that comes back over 200 KB is an error until you narrow it with a JMESPath select. A write whose response is that big still goes through and returns a short note, so don’t send it again. MCP vs API covers the difference for any product.

Pylon lists the operations the server won’t run, among them importing issues, messages, projects and tasks, sending customer portal invitations, setting a contact’s portal role, and updating users, which covers changing their roles and deactivating them. For those you need the API itself.

The API is also the way in without a browser. It takes a Bearer token (authentication), and only an Admin can make one, in Settings › Integrations & dev › API tokens. A token never expires. It can have its own name and a role you pick, or act as a teammate or a bot, and its changes show under that name. You can also hold it to up to 20 IP addresses or ranges. US workspaces call https://api.usepylon.com and EU ones https://api.eu.usepylon.com. Each endpoint allows 30, 120 or 300 requests a minute, shared by every token in your organisation (rate limits).

The server only answers when your AI tool asks it something. To act the moment something happens in Pylon, you need its webhooks, which a trigger’s Send webhook action sends, with up to three attempts. Pylon delivers at least once and in no set order, so your code has to cope with a duplicate.

Community-made Pylon MCP servers

A few people have written their own Pylon MCP servers and shared them on GitHub. Pylon doesn’t make them, and both of these run on your own machine with a Pylon API token, so an Admin has to make the token first.

• JustinBeckwith/pylon-mcp is MIT licensed and on npm as pylon-mcp, at version 1.1.2, and its last commit on main was on 13 October 2026. It needs Node.js 24 or later and PYLON_API_TOKEN in a .env file. Its README lists 32 tools. They cover your organisation, accounts, contacts, issues, tags and teams, and one redacts a message. It can delete accounts, contacts, issues and tags, and none of its tools replies to a customer.
• Carter-S/pylon-mcp turns every operation in Pylon’s OpenAPI spec into a tool, 114 of them by its README, and keeps up when you “drop in a new openapi.json, rebuild”. It hasn’t changed since 15 June 2026.

Pylon’s own server now reaches most of the API through its REST API tools, so a community server is mainly useful where a browser sign-in can’t happen, such as CI or a script. It then works with the token’s access rather than through your own sign-in.

How Deacon’s MCP server differs

Pylon is a helpdesk that gathers Slack, Teams, email and WhatsApp into one queue (pricing page), and its server lets an AI tool work that queue with 109 tools. Deacon’s MCP server has ten, seven that read and three that change something inside Deacon, and none of them sends a message. So the catch above, a reply that can’t be unsent, doesn’t come up, and nothing a stranger types into your chat can get your agent to send one through Deacon.

The reports come from the chat on your own site, where Deacon answers customers from your docs in seconds. When someone says something’s broken, Deacon files it in What needs you, under Problems in your product, grouped by topic and set beside the JavaScript errors thrown on your own pages. Your agent reads that list with get_needs_you, opens each chat with get_conversation for the customer’s own words and the page they were on, and works out the cause in your code. There’s no Pylon link to paste in, because the agent reads the report at its source.

It differs on sign-in too. Your agent connects through your browser once an owner on your team presses Allow, but a scheduled job or a GitHub Action, which can’t use Pylon’s server, takes a key from Use a key instead on Integrations. Point Deacon’s bug-report webhook at your own code, have that code start your agent in CI, and a fix can be waiting for you to check by the time you read the report.

Pylon is the one for B2B customers who expect you in a shared Slack channel, since Deacon has no Slack or Teams channel, ticket queue or email inbox, and no helpdesk integration. When your customers ask on your own website and support is you, Deacon gets their reports to your coding agent without you sorting them first.

Questions about Pylon’s MCP server

Is Pylon’s MCP server free?

It comes with all three Pylon plans, but none of them is free, and each is bought through Pylon’s sales team with a minimum seat count for new customers (pricing page). The AI agent tools can cost more on top, since backtesting uses credits. Pylon pricing has the figures.

What are Pylon’s MCP Connectors?

They’re for Pylon’s AI reaching out to other tools, not for your AI tool reaching into Pylon. Connectors let Pylon’s own AI agents call other companies’ MCP servers or APIs, and administrators choose which ones each agent can use. The pricing page lists them as MCP Connectors, beside the Pylon MCP server. Support tools with an MCP server names the other support tools whose AI can call out like this, and sets each one’s own server beside Pylon’s.

Can I try Deacon’s MCP server without a demo?

Yes. Sign up, put Deacon on your site, then open Integrations and press Connect your agent, with no sales call and no seat count to meet. The MCP server comes with the free plan, as every Deacon feature does, and you won’t be asked for a card.