Skills and MCP: extending the agent from the composer
What the agent can do is not fixed. You can give it skills and MCP servers, and you never have to leave the chat to do it — paste a link and say "install this", or just describe the skill you want and it gets written, installed and used in the same turn. This page covers the difference between the two, what you can ask for in a chat, and how scope and timing work.
Two words, kept apart
- A skill is a document describing how to handle a class of task, plus optional scripts and reference files. Once installed, the agent follows it whenever a matching task comes up.
- An MCP server is an external tool service the agent can call, so it can reach your own API or a third-party service.
Prefer a skill when you have the choice. A skill is read only when it is relevant; an MCP server's tool definitions cost context on every single turn. Unless the capability is a sentence or two and needed everywhere, package it as a skill.
Two routes: the profile pages, or just asking
| Profile pages | Asking in a chat | |
|---|---|---|
| Where | My skills, My MCP servers | Any chat composer |
| Best for | You already have a SKILL.md or a packaged zip | Building one from nothing, changing one, trying one out, or not wanting to hand-write files at all |
| How | Import from GitHub, upload a SKILL.md, upload a zip; for MCP, paste an mcpServers config | See below |
The pages are covered in 05-misc.md; this page is about the composer.
Five things you can ask for in a chat
| You say | What happens |
|---|---|
| "What does this skill do, is it any good?" (with a link) | It gets fetched and read, and explained back to you. Nothing is written to your account |
| "Install this" (with a GitHub URL or a file) | It gets installed, and you are told how to use it |
| "I want a skill that does X" | The SKILL.md is written for you and installed — you write nothing |
| "Change that skill so it also does X" | The installed one is updated |
| "What do I have installed?" / "Remove X" | Listed, or removed |
If you drop a link with no instruction, the agent reads it first and then asks whether you want to try it or keep it — it will not decide on its own to write to your account.
Two scopes — be clear which you want
| Trial | Installed | |
|---|---|---|
| Where it applies | This chat only | Saved to your profile, loaded automatically in every later chat |
| Shows up under My skills | No | Yes |
| Good for | Seeing whether it is any use | You know you want it around |
Say "keep this for later" or "add it to my skills" when you want it to persist, and "just try it" when you don't. A trial install is gone the moment you start a new chat — that is the easy mistake to make.
They do not take effect at the same time
- A skill works in the same turn. Once installed, the agent follows it immediately; you do not need a fresh chat.
- An MCP server takes effect on your next message. MCP servers are started when a session starts, so writing the config is not enough for the turn you are already in. Send one more message and it is there — nothing has gone wrong.
Everything gets read before it gets installed
The agent reads the SKILL.md before deciding what to do:
- Harmless behaviour — read-only requests to a known domain — is installed directly
- Anything that wants credentials, writes local files or runs external commands gets its risks explained to you first, and you decide
- If the package states usage terms (free tier, commercial restrictions), those are passed on to you
Third-party packages often ship an install.sh. It should not be run — those scripts install to paths this platform does not load from. Quodara has its own way of installing them.
Common questions
Q: I can't write a SKILL.md. Can I still use this? A: Yes. Describe when it should kick in, what it should do and what to avoid; it gets written in the right format and installed for you. Changing it later is another single sentence.
Q: Does installing something affect other people? A: No, it applies to your account only.
Q: I installed it but the agent doesn't seem to use it. A: Skills are read when a matching task shows up, not executed every turn. If it genuinely should have triggered and didn't, the trigger conditions in the SKILL.md are probably too narrow — ask the agent to widen them.
Q: My MCP server is installed but the agent says it can't call the tools. A: Expected. Send one more message. See the timing section above.
Q: Can I install a skill from a private repository? A: A public URL is the easiest route. For private content, upload the files, or paste the content into the chat and have the agent turn it into a skill.
Q: Does installing a lot of them slow things down? A: Skills, no — unused ones are never read. MCP servers, yes, because their tool definitions cost context every turn. That is why a skill is the better default.
Where to go from here
That covers the quick start, the core concepts, apps, chats, publishing, your profile, Explore, content marketing, scheduled tasks, Android delivery, media generation, documents, and skills and MCP. When something is unclear in practice, ask the official default assistant in Chat; when you need a human, open a support ticket.
Happy building.