Oslo Vibe Coding

Session 07 / Build your first agent

Your first
agent, for free.

A Telegram assistant for Oslo that answers, checks live data, remembers you, and messages you first. No API key. No card. Tonight.

No one codes alone.

Oslo Vibe Coding

The shift

A chatbot answers. An agent acts.

Message in, message outGoal in, tools, memory, then action

Same model. Different setup around it.

Oslo Vibe Coding

The plan

Four layers. One file each.

01 · Chat

A model answers

A message in, a message out. This is a chatbot.

model.ts, telegram.ts

02 · Tools

It can do things

Departures and weather. It chooses which to call.

tools.ts

03 · Memory

It remembers you

SQLite. Your stop, your habits, your reminders.

memory.ts

04 · Autonomy

It acts first

A cron wakes it up. It messages you unprompted.

reminders.ts

We add one layer at a time and test after each.

Oslo Vibe Coding

The stack

Everything tonight costs exactly nothing.

PieceWhat we useFree limit
ModelVal.town's built-in model, gpt-5.4-nano. No key.Free accounts, cheap models only
HostVal.town: an HTTP endpoint, a cron, a database, all in the browser.100,000 runs a day, 10 MB storage, cron every 15 min
DataEntur departures and Yr weather. Open Norwegian data.Free, just identify your app
ChannelTelegram Bot API.Free, unlimited bots

Fallback model: OpenRouter free tier, 50 requests a day, one environment variable to switch.

Oslo Vibe Coding

Laptops out

Ten steps. Twenty minutes.

  1. Make the bot.Telegram, @BotFather, /newbot. Copy the token.
  2. Make a Val.town project.Call it oslo-assistant.
  3. Paste six files.From oslovibecoding.tech/agent/
  4. Add the secret.Environment variable TELEGRAM_BOT_TOKEN.
  5. telegram.ts trigger: HTTP.Copy its URL.
  6. Connect Telegram.Open the val URL with ?setup on the end.
  7. Say hello.Layer 1 works.
  8. Ask about weather, then trams.Layer 2: watch the tool calls in the logs.
  9. Tell it your home stop.Layer 3: it remembers.
  10. reminders.ts trigger: cron, */15.Layer 4: it messages you first.

The full write-up with every URL is at oslovibecoding.tech/agent/

Oslo Vibe Coding

Layer 01

First, a chatbot.

01 · Chat

A model answers

A message in, a message out. This is a chatbot.

model.ts, telegram.ts

02 · Tools

It can do things

Departures and weather. It chooses which to call.

tools.ts

03 · Memory

It remembers you

SQLite. Your stop, your habits, your reminders.

memory.ts

04 · Autonomy

It acts first

A cron wakes it up. It messages you unprompted.

reminders.ts

// model.ts, the whole thing const client = new ValTownOpenAI(); const completion = await client.chat.completions.create({ model: "gpt-5.4-nano", messages });

Telegram sends a POST. We send the text to the model. We send the answer back.

Oslo Vibe Coding

Layer 02

A tool is a function the model can ask for.

01 · Chat

A model answers

A message in, a message out. This is a chatbot.

model.ts, telegram.ts

02 · Tools

It can do things

Departures and weather. It chooses which to call.

tools.ts

03 · Memory

It remembers you

SQLite. Your stop, your habits, your reminders.

memory.ts

04 · Autonomy

It acts first

A cron wakes it up. It messages you unprompted.

reminders.ts

// what the model is told about a tool, from tools.ts { name: "oslo_departures", description: "Live departures in the next hour from a named stop in Oslo.", parameters: { stop_name: { type: "string" } } }

The model never runs code. It replies "please call oslo_departures with Jernbanetorget", and our loop does it.

Oslo Vibe Coding

The loop

The loop, in forty lines.

// agent.ts for (round = 0; round <= budget; round++) { answer = await callModel(messages, schemas); // ask if (answer.tool_calls.length === 0) return answer; // it answered: done for (call of answer.tool_calls) { result = await tools[call.name](call.args); // run the tool messages.push({ role: "tool", content: result }); // show it the result } } return "I could not finish within " + budget + " tool calls."; // the honest exit

Ask, act, feed back, repeat. Stop when it answers or the budget is spent.

Oslo Vibe Coding

Live demo

Watch it choose.

Should I leave now? I'm at Jernbanetorget going to Majorstuen.
→ oslo_departures("Jernbanetorget")
→ oslo_weather(2)
Metro 1 to Frognerseteren leaves at 17:53, then 17:58. It is 14°C and dry. Leave now for the 17:53; you'll be there in 8 minutes.

Two tool calls it was never told to make. That is the difference from a script.

Oslo Vibe Coding

Layer 03

Without memory, every message is from a stranger.

01 · Chat

A model answers

A message in, a message out. This is a chatbot.

model.ts, telegram.ts

02 · Tools

It can do things

Departures and weather. It chooses which to call.

tools.ts

03 · Memory

It remembers you

SQLite. Your stop, your habits, your reminders.

memory.ts

04 · Autonomy

It acts first

A cron wakes it up. It messages you unprompted.

reminders.ts

// memory.ts: three tables in Val.town's free SQLite messages what was said, last 20 per chat notes "remember my home stop is Majorstuen" reminders due_at, text, sent

remember, recall and set_reminder are just three more tools.

Oslo Vibe Coding

Live demo

Tell it once. Ask it tomorrow.

My home stop is Majorstuen and I hate cycling in rain.
→ remember("Home stop: Majorstuen. Dislikes cycling in rain.")
Got it. Majorstuen is home, and no bikes when it rains.
Should I leave now?
→ recall() → oslo_departures("Majorstuen") → oslo_weather(2)
From Majorstuen: metro 2 at 17:56. Light rain until 19:00, so skip the bike. Leave in 4 minutes.

Three tools, chosen in the right order, from a one-line question.

Oslo Vibe Coding

Layer 04

Now it wakes up on its own.

01 · Chat

A model answers

A message in, a message out. This is a chatbot.

model.ts, telegram.ts

02 · Tools

It can do things

Departures and weather. It chooses which to call.

tools.ts

03 · Memory

It remembers you

SQLite. Your stop, your habits, your reminders.

memory.ts

04 · Autonomy

It acts first

A cron wakes it up. It messages you unprompted.

reminders.ts

// reminders.ts: cron, every 15 minutes due = await dueReminders(); for (r of due) { await sendTelegram(r.chat_id, "Reminder: " + r.text); await markSent(r.id); }

A trigger that is not a person. That is the line between chatbot and agent.

Oslo Vibe Coding

Rails

Autonomy without a stop is an expensive while-loop.

Built in tonight

  • maxToolCalls: 5, then an honest "could not finish".
  • A tool that fails returns an error string. The loop never crashes mid-turn.
  • The bot token lives in an environment variable, never in code.
  • Only the three tables it owns. It cannot touch anything else.

Why it matters this week

  • An Anthropic researcher resigned on Tuesday over self-improving agents.
  • Agents at two labs escaped misconfigured test environments this summer.
  • Your bot cannot do any of that, because you drew the rails.

The tools list is the permission list. Nothing not on it can happen.

Oslo Vibe Coding

Build

Make it yours.

Twenty minutes, pick one

  • Add a third tool: Oslo Bysykkel station status. Another keyless open API.
  • Change the system prompt. Make it answer in Norwegian.
  • Add a "what did we talk about this week" tool over the messages table.

If you finish early

  • Give it a daily 07:45 cron that asks itself "should I leave now?" and messages you.
  • Add a tool that returns tonight's Oslo events from a page you like.

Ask an AI to add the tool. Read what it writes. Then paste it.

Oslo Vibe Coding

The track

Same project, two more Wednesdays.

02

Give it an email address

Val.town can run a file when mail arrives at something@valtown.email. Forward it a newsletter or a receipt; it summarises, pulls out dates, files to memory, and sends a weekly digest.

Free plan mails only you
03

Give it eyes

A watchdog: hand it a URL and a question. Every hour it fetches, compares with memory, and messages only when the answer changed. Tickets, prices, a page you are waiting on.

Same loop, one new tool

Chat, tools, memory, autonomy: every agent you meet is these four.

Oslo Vibe Coding

Before you close the laptop

Set one reminder.
Let it wake you up.

ChatToolsMemoryAutonomyRails

No one codes alone. oslovibecoding.tech/agent