For the complete documentation index, see llms.txt. This page is also available as Markdown.

Frends MCP skills for AI clients

Frends also publishes pre-made skills for Frends MCP.

Connecting an AI client to Frends MCP gives it the Tools. Skills tell it how to use them well. Frends publishes a set of skills as an open plugin, so a client knows the order Frends expects, which Tool to reach for, and what to check before it reports success.

The difference shows up quickly in practice. A client with the Tools alone will often add shapes to a Process draft and never connect them, then report the Process as built. A client that has read the authoring skill validates first and fixes what validation returns.

Frends MCP is available from Frends version 6.3.2 onwards.

Skills and guides

Both carry know-how, and they reach the client by different routes.

Guides are served by the MCP server over the protocol. The client asks for them with list_guides and get_guide at the moment it needs one, so they are always in step with the Tenant. See Frends MCP Tools for more details on guides.

Skills are installed into the client. They load when the client recognises the kind of request, before any call is made, so they can shape how the client approaches the task from the start.

You do not choose between them. Guides keep a client correct about your Tenant, and skills keep it sensible about the request.

Available skills

The following skills are available through the Frends plugin.

Skill
Use it for

getting-connected

Connecting for the first time, or fixing a connection that returns 401, 404, or empty lists

find-and-inspect

Finding which integrations exist, where a Process is deployed, and how one is built

diagnose-failures

Investigating a Process run that failed, stalled, or produced nothing

integration-planning

Interviewing you about an integration and turning the answers into a written plan

build-a-process

Building or editing a Process draft up to a passing validation

integration-planning and build-a-process are meant to be used in that order. Planning asks about systems, triggering, volumes and error handling one topic at a time, and produces a plan the build step works from. Starting at the build step tends to produce a Process that runs but solves the wrong problem.

Installing the plugin in Claude Code

Add the marketplace, then install the plugin with these commands:

The plugin reads your endpoint and token from two system environment variables, so no credential is stored in the plugin itself.

Setting the variables on Unix systems:

On Windows, set the same two variables in PowerShell, using setx if you want them to persist between sessions.

Getting the token is covered in How to connect an AI client to Frends MCP.

Set the token as an environment variable rather than writing it into a configuration file. Configuration files get committed to version control, and a Frends MCP token acts on your Tenant.

Using the skills with Codex CLI

The skills are plain Markdown files in the same repository, so they can be used with Codex CLI as well.

Add the Frends MCP server pointing at the same endpoint, reading the bearer token from FRENDS_MCP_TOKEN, either through the CLI or in the Codex configuration file.

Claude web and Desktop

Claude on the web and Claude Desktop add the connector through the client rather than with a marketplace command, and their custom-connector flow authenticates with OAuth by default, which Frends MCP does not offer.

The direct path therefore works only for organizations with access to Anthropic's request-header authentication beta, where an administrator enters the bearer token once for the whole organization.

Without that access the two clients differ: Claude Desktop supports local MCP servers, so it can still connect through a locally running connector that forwards to the remote endpoint and adds the header.

Claude on the web cannot run a local connector, so without the beta it has no route. The endpoint and token are the same in every case.

The plugin is open

The plugin repository is public and MIT licensed, so you can read exactly what each skill instructs the AI client to do before you install it. You can also copy a skill and adapt it to your own conventions.

Last updated

Was this helpful?