Skip to main content
The Tightknit MCP server is currently in pilot. Contact Support if you’d like to join the pilot and turn it on for your community.
The Tightknit MCP server lets AI agents work inside your community on your behalf. It implements the Model Context Protocol, an open standard that connects AI apps to external tools, so an agent like Claude, ChatGPT, or Cursor can read your feeds, post and moderate discussions, run polls, manage members, and more, all through natural language. This is a remote, hosted server: there is nothing to install and no API key to manage. You add one URL to your agent and sign in with your Tightknit account.

Let your agent set it up for you

The fastest way to get started is to hand this page to your agent and let it walk you through the setup. Paste the following prompt into Claude, ChatGPT, Cursor, or the agent of your choice:
Copy this into your agent
Your agent will read this guide, add the server to its configuration, and prompt you to sign in. If you’d rather set it up by hand, follow the steps below.

Connection details

Access requires the MCP User role (admins have full access), in a community that’s in the pilot. If you can’t sign in, ask Support to enable the MCP server for your community.
You never paste a token or secret. On first connect, your agent opens a browser window where you sign in, and the communities you can access are scoped to that account automatically.

Connect your agent

Add the server from the terminal:
Then run /mcp inside Claude Code and choose Authenticate to sign in through your browser. Once connected, /mcp lists the Tightknit tools as available.

How access works

The MCP server never holds your password and issues no long-lived key. Your agent completes a standard OAuth 2.1 sign-in, and every request is authorized against your role and permissions in the Studio. A few things follow from this:
  • Role-based — Admins have full access. Everyone else needs the MCP User role, which grants the mcp:access permission to connect. Without it, you can’t reach the server.
  • Tools follow your Studio permissions — Beyond connecting, each tool checks your Studio role and permissions, so an agent only sees the tools your roles allow. For example, pairing MCP User with Events Manager exposes the events tools but not the polls, members, or site tools. Admins can use every tool.
  • Scoped to your communities — The agent only acts on communities where your role has access. If you can reach more than one, ask the agent to run list_tenants first, then reference the community you want by the tenant_id it returns.

Available tools

Each tool is annotated so your agent understands the impact of an action before it runs one:
  • Read — Fetches data. Never changes anything.
  • Write — Creates or updates content in your community.
  • Destructive — Deletes content. Cannot be undone.
Tools for features your plan doesn’t include won’t appear in your agent’s tool list, and a few fields on the tools you do see (such as group flair, award certificates, and uploading a custom font file) are only accepted with the matching feature. Every destructive tool also takes a confirm parameter the agent must set to true. It makes a delete an explicit act rather than an incidental one, but the agent supplies it, so it is not a substitute for your own approval: keep a human in the loop for deletes. Destructive here means the row is gone, not hidden: there is no trash and nothing to restore from. Where a reversible option exists, the tool’s description points at it — an event can go back to draft instead of being deleted, a journey can be disabled, a feed can be archived, and a post or comment can be unpublished.
Agents can create, edit, publish, and delete real content in your community. Review what your agent proposes before you approve write actions, and keep a human in the loop for moderation and member changes.

Communities

Events

Activity

Slack

Polls

Members

Posts and comments

Feeds

Groups

Site

Onboarding

Journeys

Awards

Content library

Tags

Custom fonts

Available on plans that include custom fonts. A font family is created empty, then one font file is attached to it per weight and style you want the site to render in — the same two steps as uploading a font in Studio. Deleting a face or a family deletes the font files behind it. Anything the site rendered in a deleted family reverts to the default typeface.

Branding

A branding set is a named collection of style tokens for your site, including its colors, typefaces, and corner roundness. Your site has one active branding set at a time which is applied. The rest are saved alternatives, so an agent can prepare a look and you apply it when you’re ready.

Files

Several tools can attach an image you hold as a local file, such as an event cover, an award badge, a site logo, a group banner, or a content cover. The agent uploads the file with prepare_file_upload first, then passes the returned file id to the tool that uses it. Custom font files (.woff2 and .woff) go through the same two steps, then attach with add_custom_font_face. On the edit tools, every image parameter takes null to remove the image that is currently set. Omitting the parameter leaves the existing image alone. The same holds for optional text such as a description or a link: pass null to clear it, omit it to leave it as it is. Uploads are capped both by what the file is used for and by your plan’s file size limit, whichever is lower. prepare_file_upload states the cap it applied when it turns a file down.

Let the agent reach the upload domain

The agent sends the file bytes itself, from its own sandbox, so that sandbox needs network access to mcp.tightknit.ai. This is a separate setting from the connector you already approved, and most clients restrict it by default. Claude desktop and claude.ai: unless your Domain allowlist is set to All domains, uploads fail until you add the Tightknit domain. Go to Settings > Capabilities, enter *.tightknit.ai under Additional allowed domains, and select Add. ChatGPT web and desktop: sandbox internet access is all or nothing rather than per domain. Turn on Allow public internet access under Settings > Data controls > Work network access. While it is off, the sandbox reaches only OpenAI’s own managed allowlist and uploads fail. This setting may be controlled by your workspace admin. Every other tool reaches Tightknit through the connector itself, so only file uploads need this.

Rate limits

Tool calls are limited per minute, per person, so one busy agent can’t slow things down for everyone else. Each tool draws on one of three budgets: Budgets are separate, so a run of searches never eats into what’s left for writes. A budget covers all the communities you administer rather than each one separately. There is also an overall cap of 300 requests a minute per connected agent. Limits are counted per region, so an agent spread across regions may get more than the figures above, never less. If your agent goes over a tool budget, the tool returns a message saying which one was exhausted. Going over the overall 300 cap instead fails at the connection level with an HTTP 429, which most clients report as a request error rather than a tool result. Either way, ask the agent to wait a minute and continue. Tools that take a list, such as give_award and add_group_members, cost one call no matter how many members you pass. Batching is cheaper against these limits than looping, but it does more at once, so keep the batches to what you actually intend to happen.

Troubleshooting

Access requires the MCP User role (or Admin) in a community that is in the pilot. Confirm you’re signing in with the right account, and ask Support to enable the MCP server for your community if it isn’t already.If sign-in completes and the client still reports a generic error with no tools, remove the connector and add it again. Some clients cache a failed handshake and keep replaying it.
If you can access more than one community, tools need to know which one to target. Ask the agent to run list_tenants and then reference the community by the tenant_id it returns.
Each budget refills every minute. Ask the agent to wait and retry, or to batch its work: passing a list of members to give_award or add_group_members costs one call rather than one per member. See Rate limits for the budgets.
The agent uploads file bytes from its own sandbox, which needs network access to mcp.tightknit.ai. In Claude, add *.tightknit.ai under Settings > Capabilities > Additional allowed domains. In ChatGPT, turn on Allow public internet access under Settings > Data controls > Work network access. See Files for the full steps.
The server uses the Streamable HTTP transport with OAuth. Make sure your client supports remote MCP servers (not only local stdio servers) and that you entered https://mcp.tightknit.ai/mcp exactly, including the /mcp path.
It doesn’t need them. Enter https://mcp.tightknit.ai/mcp and leave the OAuth fields blank: the server publishes its own metadata, and clients register themselves automatically, so there is no client ID or client secret to issue.If your client requires the values manually:

CLI

Manage your community from the terminal with the Tightknit CLI.

API Keys

Generate keys for direct access to the Tightknit Admin API.