skillit

Overview

skillit is a generic dispatcher agent. Operators upload markdown files via the admin UI; each filename becomes the name of a "skill" whose content is treated as the system instructions for handling matching user requests. The agent loads skills from Redis at boot, refreshes its in-process cache via pub/sub on admin changes, and either routes a user message to the best-matching skill via the run_skill tool or executes an explicitly-named skill when the user prefixes their message with /<skill-name>.

Each skill can optionally be scheduled to run on cron, with the agent invoked in-process via agent.generate() and the result persisted to Redis (and posted to a Slack channel if one is configured on the schedule).

Usage

Adding and editing skills

Sign into /admin with ADMIN_PASSWORD. Drag-and-drop one or more .md, .markdown, or .txt files into the upload area; the filename (minus extension) becomes the skill name. Existing skills can be deleted from the same view. Changes propagate to the agent's in-process cache immediately via Redis pub/sub — no restart required.

A bundled example skill is seeded on first boot if Redis is empty and is auto-filtered from routing once any user-uploaded skill exists.

Chatting with the agent

In any messaging adapter (web playground, Slack channel where the bot is installed, etc.):

• Ask a question. The LLM picks the best matching skill from the dynamic system prompt, calls run_skill to load it, then follows the loaded instructions.
• Prefix a message with /<skill-name> to dispatch directly, skipping LLM routing. The skill content is inlined into the prompt and applied as-is.

Scheduling skills

For each skill, click Schedule in the admin UI and provide:

• Cron expression — a standard 5-field expression (e.g. 0 9 * * 1-5).
• Prompt (optional) — sent as the user input to the skill on each fire. Leave blank to fire the skill with no additional input.
• Timezone (optional) — IANA TZ name; defaults to the container's TZ.
• Slack channel (optional) — a Slack channel ID (C0123456789) to post the run's output to. Requires the platform's Slack adapter to be configured separately; the bot must be a member of the target channel.

The most recent run for each schedule is shown inline (timestamp, duration, output or error). Use the Run now button to fire a schedule on demand.

Web search

Skills can use the web_search tool to fetch live results from the web via Tavily. The tool is only available when TAVILY_API_KEY is set — if it is not configured, the tool is silently omitted and skills that reference it will fall back to the model's training data.

To enable it, set TAVILY_API_KEY while deploying the agent blueprint.

Inputs

Input	Required	Description
ADMIN_PASSWORD	yes	Password gating the /admin upload UI
SESSION_SECRET	yes	HMAC secret used to sign admin session cookies
TAVILY_API_KEY	no	Tavily API key for web search; if omitted, web_search is disabled

Limitations

• Single replica only. The scheduler runs in-process; with multiple replicas the same cron job fires on each.
• Mastra memory is in-process. Each cron fire creates a fresh thread in LibSQLStore({ url: ':memory:' }). An hourly sweep deletes threads older than 24h, but they don't survive a process restart.
• No history of cron runs. Only the most recent run per schedule is persisted in Redis (skillit:last-run). Older runs are overwritten.
• Slack posting requires the bot to be in the target channel. Posts to channels the bot isn't a member of are silently dropped with not_in_channel. Use the channel ID (e.g. C0123456789), not the channel name.
• Session storage is ephemeral. Admin sessions are HMAC-signed cookies — they survive process restarts only if SESSION_SECRET is set to a stable value (otherwise an ephemeral secret is generated and sessions reset on restart).
• Output truncation. Cron run output is truncated at 8 KiB for Redis storage and ~35 KB for Slack posts.