For: everyone. Every role may hold their own MCP connection, members included — this is the one workspace connection a member can create.
MCP is the way your playbooks and personas become available inside Claude, ChatGPT, Claude Code, Codex, Cursor, VS Code and any other client that supports it. There is one endpoint, and each person connects their own client to it.
Go to Settings → MCP Server and pick your tool's tab. Each tab shows the exact steps and, where the tool needs one, the file or command to copy — with your endpoint already filled in.
Claude, ChatGPT, Claude Code and Codex sign you in with your sqemes account; no key needed.
Cursor and VS Code use a key, which you create on the same page. The snippet reads it from an environment variable or asks for it once, so the key never sits in a file you might commit.
A tool that is not listed: add the endpoint as a remote MCP server, sign in if it offers to, and otherwise send a key as a bearer token.
Nothing about this is configured inside sqemes beyond creating the connection — the rest happens in the other tool.
It gets tools for searching and fetching your playbooks and personas — search_templates, get_template, read_context_file, download_skill, list_personas, get_persona and others, depending on what your key is allowed to do.
This is the most capable of the three channels, because the client can find the right playbook mid-task without anybody choosing it from a list. It is also the only one besides sqemes Chat where personas work.
When Claude looks at your playbooks. Claude uses them when you ask it to, or when a request clearly names one, such as a skill or a persona. To have it check on its own whenever a task fits, copy the instruction from Settings → MCP Server into your Claude instructions — your own, a project's, or, for the whole team, your organisation's. sqemes does not send that instruction by itself: it is yours to set. The instruction was rewritten on 19 October 2026: it now checks your personas first and asks when two of them fit. If you pasted it before then, paste it again.
A skill's references reach the client as text, when the task needs them. Files that are meant to be used rather than read — scripts, templates, fonts, anything that is not text — come as a download link instead. The link is valid for 15 minutes, and the client saves the file next to the skill.
A client that installs skills, such as Claude Code or Codex, can fetch a whole skill with download_skill: the SKILL.md plus a link for every file.
Reading one file. read_context_file takes either the file's uri from get_template, or the playbook (its name or id) plus the file's path inside the skill, exactly as SKILL.md writes it — for example references/animation.md. Text comes back as text; scripts, assets and other files come back as a download link. Pass as: "text" or as: "file" to get the other one.
Downloads count towards the workspace's daily download limit, which normal use never comes near.
Claude fetches those links from its own sandbox, and on Team and Enterprise plans the organisation decides which domains that sandbox may reach. The defaults do not include sqemes: Team allows package managers only, and new Enterprise organisations start with network access switched off.
To let Claude use the files in your skills:
Open Organization settings → Capabilities.
Under code execution and file creation, choose Allow network egress to package managers and specific domains.
Add api.sqemes.com.
The same domain covers uploads: when Claude saves a file to sqemes, it sends it to api.sqemes.com as well. Without it, Claude still reads every text file of a skill — only downloading and uploading files fail. On a self-hosted instance, add your instance's own domain instead.
This is the part worth understanding properly. A connection is bound to the person who created it, and the server answers with that person's playbooks — not the workspace's.
So sharing a connection string with a colleague does not give them their own view; it gives them yours, under your name. Everyone creates their own. It takes a minute and it keeps access rules meaning what they say.
It went dormant. A connection unused for 90 days is refused rather than deleted. If something that worked months ago now returns nothing, create a new connection.
The person left the workspace. Connections go with their owner. That is the intended behaviour — it is also why a shared connection string is a bad idea, because it stops working for reasons the person using it cannot see.
Check in this order: the key's permissions, whether you personally have access to the playbooks you are expecting, and whether the connection is pointed at the workspace you think it is. A connection that authenticates cleanly and returns an empty list is almost always an access question, not a setup one.
Updated 19 October 2026 · Article version 6
