> ## Documentation Index
> Fetch the complete documentation index at: https://docs.mibyanai.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Skills, tools, and automation
> Skills and the Skills Hub, toolsets, MCP servers, plugins, scheduled jobs, webhooks, and messaging platforms
This page is about what the agent can **do** and how to make it work **without you**: skills, tools, MCP servers, plugins, schedules, webhooks, and messaging channels.
```mermaid theme={null}
flowchart LR
A["The agent"] --> S["Skills
know-how it can load"]
A --> T["Toolsets
terminal, web, files, memory..."]
A --> M["MCP servers
tools from other services"]
A --> P["Plugins
extend the backend or the app"]
Z["Things that start work
on their own"] --> C["Scheduled jobs"]
Z --> W["Webhooks"]
Z --> G["Messaging platforms"]
C & W & G --> A
```
Changes to skills, tools, and MCP servers apply to **new sessions**. A chat that is already running keeps the setup it started with.
## Capabilities
Open **Capabilities** from the sidebar. It has three tabs: **Skills**, **Tools**, and **MCP**.
### Skills
A **skill** is reusable know-how the agent loads when it is relevant. Mibyan also learns skills from your work.
| Label | Meaning |
| - | - |
| **Built-in** | Shipped with Mibyan |
| **Learned** | Created by the agent from your work |
| **Hub** | Installed from the Skills Hub |
Search the list, sort by **Most used** or **A to Z**, and see how many times each was used. **Enable all**, **Disable all**, and **Disable unused** change many at once. You can **edit** a skill, or **archive** it. An archived skill is not deleted, and can be restored later with `mibyan curator restore`.
A background **curator** reviews agent-created skills and archives stale ones. It never touches built-in or hub skills. Pause, resume, or run it from Command Center, Maintenance.
### The Skills Hub
The hub is where you find and install more skills from the official index, GitHub, and community sources.
Results show how many were found and how quickly. A hub that times out is listed so you know.
Read the skill's `SKILL.md` and see its files before you install.
A security scan gives a verdict of **Safe**, **Caution**, or **Dangerous**, lists findings, and shows the policy: install allowed, review before installing, or blocked by policy. Each skill also shows its trust level: builtin, trusted, or community.
**Install**, and later **Update installed** or **Uninstall**. Errors include an action log.
### Tools (toolsets)
A **toolset** is a group of capabilities such as terminal, web search, file work, memory, and delegation. Turn each on or off. A toolset shows **Configured** or **Needs keys**.
* **Providers.** Some toolsets can use different backends. For web, choose one for **Search** and another for **Extract**. The active one is marked.
* **Execution backend.** Choose where terminal commands run. Each option shows **Ready**, **Needs setup**, or **Unavailable**, and you can pick one now even if setup is pending (commands fail until it completes).
* **Setup.** Some backends need a one-time install, run from the app.
* **Keys.** Add the credential a tool needs right there. Values are hidden until you reveal them.
* **Vision** uses your auxiliary model configuration, so the image-capable model is chosen in **Settings, Models**.
Removing every enabled toolset disables memory, terminal, web search, delegation, and most other tools until you turn them back on. The app asks you to confirm.
## MCP servers
**MCP** connects the agent to tools that live in other services. You manage servers in **Capabilities, MCP** or **Settings, MCP**.
* **New server** takes a name and a JSON config, for a **stdio** command or an **HTTP** URL.
* **Test connection** reports how many tools it exposes.
* **Enable** or **disable** a server, or a single tool. It applies to new sessions.
* A server can show **Connecting**, **Needs authentication** (choose **Authenticate**, and a browser opens), **Error**, or **Off**.
* **Reload MCP** refreshes tool schemas for fresh turns.
* Each server shows how many tools, prompts, and resources it offers, an estimate of tokens per call, and uses in the last 30 days. Servers you never use are flagged **unused**.
A catalog of approved servers you can **Install**. If one needs credentials, you fill them in first.
Paste an `mcp.json` snippet, an `npx` or `docker` command, a `claude mcp add` line, a URL, or a Cursor link. Mibyan recognizes the servers in it and asks before adding them.
An **`mibyan://` link** can offer to add an MCP server. The app shows the exact configuration and where it came from, and adds it only if you confirm. Only `http` and `https` URLs are accepted, the config is limited to 32 KB, and a server that runs a local command carries an extra warning: continue only if you trust the source.
The agent can also offer to install, enable, or authorize a server inside a conversation, from the Mibyan-approved catalog. It always asks first.
## Plugins
There are two kinds, shown in **Settings, Plugins**:
| Kind | What it extends | Notes |
| - | - | - |
| **Desktop plugins** | The app's interface | Bundled with the app or dropped into the desktop plugins folder. **Disable** unloads them live. **Rescan** picks up new ones |
| **Agent plugins** | The backend: tools, skills, MCP servers, hooks, and slash commands | Sources are bundled, user, git, project, or pip. **Portable** ones are skill and MCP bundles that also work in other agents. Toggles apply to new sessions |
To install a plugin from a repository, choose **Install plugin** and paste the git URL. Before anything is installed, the app inspects the repository and shows what it contains (an agent plugin, a desktop UI, or both), where each part installs, and a link to the source. Pick the components, optionally **enable the agent plugin after install** or **force reinstall**, then **Install**. It warns about insecure `http` or local URLs, and lists any environment variables the plugin needs.
Install plugins only from sources you trust. A plugin can add tools and run code.
## Scheduled jobs
A **scheduled job** (also called a cron job) runs a prompt for you on a timetable and delivers the result. Open **Scheduled jobs** from the sidebar.
Or start from a **Blueprint**, a ready-made automation you fill in and schedule.
For example: Summarize my unread Slack threads and email me the top 5.
**Daily** (9:00 AM), **Weekdays**, **Weekly** (Monday), **Monthly** (the first day), **Hourly**, **Every 15 minutes**, or **Custom**. Custom accepts cron syntax such as `0 9 * * *`, or phrases like every hour or weekdays at 9am.
**This desktop**, **Telegram**, **Discord**, **Slack**, or **Email**. The chat platforms need a **home channel** set first.
Or use the default global model.
Each job shows its state (enabled, scheduled, running, paused, disabled, error, completed), the **last** and **next** run, and a **run history**. You can **trigger now**, **pause**, **resume**, **edit** (changes apply on the next run), or **delete** it, which stops it immediately.
If you change models in a way that affects jobs, the app warns that some jobs **will be skipped until you review their model settings**, and links to them.
Inside one chat, `/loop` re-runs a prompt on an interval, and the composer suggests **Schedule this** when a request sounds recurring.
## Webhooks
Webhooks let other systems start the agent by sending an HTTP event. They are their own gateway platform, so enable them first, and restart the gateway when asked.
| Field | Purpose |
| - | - |
| **Name** | For example `github-push` |
| **Events** | Comma-separated, or leave empty for all |
| **Skills** | Skills to load when it fires (optional) |
| **Prompt** | Instructions for the agent when it fires (optional) |
| **Deliver to** | Log, Telegram, Discord, Slack, Email, or a GitHub comment |
| **Deliver payload only** | Skip the agent and just deliver the payload |
When you create a subscription the app shows its **webhook URL** and a **secret, shown only once**, so copy it right away. Subscription changes hot-reload once the receiver is running, and a disabled subscription rejects incoming events.
## Messaging platforms
Open **Connections** from the sidebar to let people talk to your agent from chat apps. Each platform shows a state: **Connected**, **Connecting**, **Needs setup**, **Restart needed**, **Retrying**, **Startup failed**, **Error**, **Disabled**, or **Messaging gateway stopped**.
| Platform | How you connect |
| - | - |
| **Telegram** | Scan a QR code and Mibyan creates the bot for you, or paste a token from `@BotFather`. Add the numeric user IDs allowed to talk to it |
| **WhatsApp** | Scan a QR code in WhatsApp under Linked devices. Choose a **separate bot number** or **your own number** |
| **Discord** | Create an application and bot in the Discord Developer Portal and paste its token. Choose the reply style, allowed users, and a home channel |
| **Slack** | Paste the bot token and the app-level token used for Socket Mode |
| **Matrix, Mattermost, Signal, BlueBubbles (iMessage), QQ** | Fill in the fields shown, guided by the setup link |
Every platform has a **setup guide** link. Follow it, save, then enable the platform.
Always set the **allowed users** list. Without it, anyone who finds your bot can message it. Allow-all options are for development only.
Security and pairing:
* New users can appear as **pending pairing requests**. **Approve** one to let that person talk to the agent, or **Revoke** a user, who then stops being recognized on their next message.
* Too many failed approvals **lock the platform out** for a while.
* Credential changes take effect after a **gateway restart**. Use the status bar or the banner in the page.
* The **home channel** is where the bot sends things you did not ask for in the moment, such as scheduled job output and reminders.
A chat can be **handed off** to any connected platform with `/handoff`, and resumed in Desktop later.
Every settings page, with what it changes.