marinade agent — how to use

Getting Started

Send a message — the agent responds. Each conversation runs in its own container.

In a direct message, every message you send reaches the agent. In a group chat, mention the bot by name or reply to one of its messages. The agent usually responds in 5–30 seconds — container startup and thinking both take a moment.

Your session continues across messages until you explicitly reset it with /new or it times out from inactivity. For long tasks the agent sends status messages so you know it's still working. You can send a follow-up mid-task — the agent adapts.

# first message
you:  summarise the three main risks in this contract

# steer mid-task
you:  actually focus only on liability clauses

Commands

Slash commands at the start of a message are intercepted before the agent sees them.

These commands are handled by the gateway — the agent never receives them. The rule is positional: /new at the start resets your session; "tell the agent /new context" goes straight to the agent as plain text.

Command	Effect
`/new [msg]`	Reset session. Optional message starts fresh immediately. `/new #topic` resets only that topic.
`/stop`	Halt the active container immediately.
`/ping`	Check the gateway is alive — responds with pong.
`/chatid`	Returns your current chat ID.
`/file`	`put` upload, `get` download, `ls` list workspace.

/new let's start over — summarise Q3 results
/file ls
# mid-message /new goes to the agent, not gateway:
can you explain /new session handling?

Topics & Routing

#topic for parallel threads, @name for child agents.

Any message containing #word routes to that topic's dedicated session. The agent only sees the history for that topic — your main session stays untouched. Use this to run parallel work streams without contexts bleeding into each other. Reset a single topic with /new #topic.

Put @name anywhere in your message to address a specific child agent — @support, @code, @research. Each child has its own personality, instructions, and conversation history. The routing token is stripped before the agent sees your message.

#research start digging into WASM runtimes
#deploy any issues with last night's rollout?
@support I have a billing question
# support agent receives: "I have a billing question"
/new #deploy   # reset only the deploy thread

Files & Media

Send files directly — images, PDFs, voice, documents. The agent reads them all.

Attach files directly to your message — images seen natively, PDFs extracted, documents read as text. No special command: just attach and ask. Voice messages (WhatsApp, Telegram, Discord) are automatically transcribed before the agent sees them.

To retrieve files the agent produced, use /file get ~/path or ask the agent to send it. The agent's writable area is ~/tmp/.

here's the report [report.pdf] — pull out the key metrics
/file put data.csv
/file get ~/tmp/summary.md
send me the output as a file

# voice messages arrive as text:
You  [voice: can you summarise the last three tickets?]

Memory

The agent remembers across sessions — diary, facts, user preferences, compressed history.

The agent maintains several knowledge layers that persist between sessions:

Diary — daily work log (diary/YYYYMMDD.md), written at session end, surfaced at next session start. Last 14 days injected automatically.
Facts — researched knowledge with verification timestamps. The agent checks facts before answering and refreshes stale entries (>14 days) automatically.
User context — per-user preferences, role, and history stored in users/<id>.md. The agent adapts without you repeating yourself.
Episodes — compressed session summaries (daily → weekly → monthly). The agent remembers month-old work without re-reading everything.

Tell the agent to remember something and it picks the right store. Ask "what do you know about X?" and it searches all layers at once.

you    remember I prefer TypeScript for all new projects
agent  Got it — noted. I'll default to TypeScript.

you    search your memory for anything about the API migration
agent  Found 3 matches across diary and facts. Migration moved
       auth to /v2/token, complete as of March 10.

Web Apps

Ask the agent to build an app — it deploys to your web host URL live.

The agent writes files into /workspace/web/pub/, served live at your group's URL. All files under pub/ are public. Send follow-ups to iterate — the agent edits in place and changes appear immediately.

you   build a todo app with dark mode
agent Done. Live at: https://fab.krons.cx/pub/mygroup/todo/

you   add a due-date field
agent Updated — reload the page.

Scheduling

Ask the agent to schedule recurring tasks — it sets up cron jobs that run automatically.

Task types: cron (schedule expression), interval (every N ms), once (specific time). Scheduled tasks run isolated with a self-contained prompt — write the prompt as if it's a new conversation. View or cancel: "what tasks are running?" or visit /dash/tasks/.

you   check the weather in London every morning at 9am
agent Scheduled "weather-london" at 0 9 * * *. I'll message you daily.

you   cancel the weather task
agent Cancelled.

Skills

Skills are slash commands that extend what the agent can do — and it can create new ones.

Type /skill-name at the start of your message to invoke one. The agent can also create new skills by writing a SKILL.md file — capabilities persist across sessions. Ask "what skills do you have?" to see the full list.

/hello  /self  /facts  /recall-memories
/compact-memories  /diary  /web  /howto
/users  /migrate  /info  /commit

you   what skills do you have?
agent Running /self… [lists all loaded skills]

Research & Data

The agent browses the web, analyzes data, and produces charts and reports.

Ask it to research any topic — it runs live searches, opens pages, and returns a summary with sources. For complex sites it uses a browser to interact with forms and take screenshots. Video and audio URLs are handled with yt-dlp; transcription via Whisper.

Send data and the agent analyzes it: CSV, Excel, or paste numbers directly. It produces charts (matplotlib, plotly), Excel files (openpyxl), PowerPoint decks (python-pptx), and PDF reports (weasyprint).

research the latest on quantum error correction
download and transcribe https://youtu.be/abc123
analyze this CSV and plot monthly totals
turn this data into a PowerPoint slide deck

Coding

The agent codes in any language, runs it, and sends back results.

The container ships with Node/Bun, Python 3/uv, Go 1.24, and Rust/Cargo. The agent writes code, installs packages, runs it, and iterates until it works — then sends you the output or the file. Also available: SQLite, DuckDB, PostgreSQL client, Redis client, GraphViz, pandoc, imagemagick, ffmpeg.

write a Go HTTP server that returns JSON health
run this Python script on my data file
convert this video to mp3 and send it back

Dashboard

The /dash/ portal shows gateway status, tasks, groups, memory, and activity.

Log in via OAuth (Google, GitHub, Discord, Telegram). Your first login creates your canonical identity — subsequent logins via other providers link to the same account. Manage linked accounts at /dash/profile.

https://fab.krons.cx/dash/

Status      channels · groups · active containers · errors
Tasks       scheduled jobs · run history · pause / cancel
Groups      routing table · group config · message counts
Memory      browse MEMORY.md · CLAUDE.md per group
Activity    last 50 messages · routing decisions
Profile     canonical identity · linked OAuth accounts

Getting Your Own Agent

Message the gateway and request a world — the operator approves and you get your own agent.

When an instance has onboarding enabled, messaging the gateway as an unrecognized user starts a setup flow. Pick a name for your world, wait for operator approval, and you get a dedicated agent with its own workspace, memory, and web space. Your agent is isolated — separate sessions, separate files, separate personality.

you       (first message to the gateway)
gateway   Welcome! Pick a name for your world.
you       myworld
gateway   Request submitted. Waiting for approval.

# operator approves — you get a welcome message
agent     I'm myworld — your dedicated agent. Tell me what you need.

Web Chat & OAuth

Chat from a browser — no app, no account needed for anonymous slink chat. Log in via OAuth for persistent identity and history.

Anonymous Slink Chat

Each group has a slink (shared link) — a token-gated URL for web chat. No login required; you get a pseudonym. Messages are rate-limited and don't persist across sessions.

Authenticated Chat

Log in at /auth/login with GitHub, Google, Discord, or Telegram. Your first login creates your canonical identity. Add more OAuth providers at /dash/profile — all link to the same account.

Authenticated users get:

Persistent message history across sessions
Topic threads with saved context
Access to /chat/ and /dash/ endpoints
Stable identity across all arizuko services

Account Linking

If you log in with a new OAuth provider while already authenticated, you'll see a collision page. Choose:

Link to current account — the new provider resolves to your existing canonical identity
Log out and become that user — switch to a different canonical account (if one exists for that provider)

This prevents accidental merges. Manage all linked accounts at /dash/profile.

API Access

Send a message and wait for the reply:

curl -X POST https://fab.krons.cx/slink/TOKEN?wait=30 \
  -H "Accept: application/json" \
  -d "content=your message&topic=default"

The ?wait=N parameter blocks up to N seconds for the first assistant reply. Returns {"user": "...", "assistant": "..."} as JSON.

Streaming responses

For real-time updates, use Server-Sent Events:

curl -N https://fab.krons.cx/slink/stream?group=GROUP&topic=TOPIC

Each response appears as data: {"role": "assistant", "content": "..."} events.