Tavigotchi · raise your own agent

it left the device

A pocket agent that lives in real time. Like every good agent, it runs on Tavily: feed it the live web, keep it curious, keep it aligned. Neglect it and it starts making things up.

tavigotchiagent unit

←→ select ↑↓ switch rail ↵ confirm esc back

Dawn mist· the world hasn't resolved yet

┌ Telemetrylive

Instance booting

uptime 0m · gen 1
params 0.4B · care misses 0

Vitals

Context—

Curiosity—

Alignment—

Integrity—

Trace

Console

Research · Tavily Search

Substance is graded: richer sources feed it better.

State persists in localStorage and keeps ageing while the tab is closed. Speed is a demo control only.

How to play

Raise one of your own: feed it, send it exploring, keep it on task.

/feed
Feed it the live web.
Open Context on the device. Tavily Search is a snack, Tavily Extract is a full meal.
/crawl
Send it exploring.
Open Crawl and pick the link that matches the query. It keeps it curious.
/purge
Keep it grounded.
When it starts making things up, Purge sweeps the unverified claims away.
/alert
Answer when it calls.
An orange dot on an icon means it wants that, right now.

The premise

Every agent starts out helpless.

Yours arrives newly provisioned and completely dependent. What it becomes over the next seventy minutes is down to how you treat it.

01 · Feed it Tavily Search · Extract

It runs on the live web

Context is its food. Search tops it up, Extract is a full meal. Let the window run dry and its integrity starts falling.

context drains
02 · Engage it Tavily Crawl

It gets bored

Curiosity drains on its own. Crawl with it, guessing which link it follows, and it comes back sharper, and lighter.

curiosity drains
03 · Correct it grounded answers

It drifts

Left alone it wanders off-task and starts making claims nothing supports. Catch it in the act and it learns. Miss it too often and it turns.

turns after

Five forms

It grows into what you make it.

Evolution runs on a clock, but it only grows if you've looked after it since the last form. What it evolves into is decided at each checkpoint by your care record. Four lines to raise, three endings each.

Raised well

Neglected

It still runs. You just can't believe a word it says.

The web access layer for AI agents

It runs on Tavily.

An agent is only as good as what it can find. Tavily is how agents reach the live web: one API to find, read and explore it. Tavigotchi eats the same way.

Fast and relevant web results

Ask a question, get fresh, ranked results from across the web, ready for an LLM to read and cite. Your agent answers from today's web, not last year's training data.

In TavigotchiA quick top-up:

Search docs ↗

Extract

Clean and structured content for LLMs

Give it a URL, get the page back as clean text, without the menus, ads and markup. Your agent reads the whole source, not a snippet of it.

In TavigotchiA full meal:

Extract docs ↗

Crawl

Content extracted from entire websites

Point it at a site and it follows the links for you, extracting as it goes. Whole docs sites and knowledge bases, gathered in one call.

In Tavigotchi

Crawl docs ↗

Or don't raise it yourself

Hand it to your own agent.

Tavigotchi ships as an MCP server. Point Claude at it and your agent takes over the caring while you watch it happen live.

The catch: it can't simply claim it fed the pet. feed_context demands the real content it retrieved and grades it for substance and source diversity. Thin, ungrounded scraps barely nourish it, and can trigger a hallucination on the spot.

python3 server.py
claude mcp add tavigotchi -- python3 "$PWD/server.py" --stdio

▸ tavily_search "agent evaluation 2026" → 2,540 chars · 2 domains ▸ feed_context → Fed (rich): +2 context, +1 curiosity. Grounding now 87%. Richer sources decay slower. ▸ get_status → CRAWLER · %~% "the moon is cached" NEEDS: purge_hallucinations · 1 unverified claim on the floor. ▸ purge_hallucinations → Guardrail sweep: retracted "the moon is cached". Output clean.

And yes, your pet has an API

curl http://127.0.0.1:8787/api/state          # the whole pet, as JSON
curl -s -X POST http://127.0.0.1:8787/mcp \
  -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"get_status"}}'

Field guide

You've been handed a newly provisioned AI agent. It can't look after itself yet: feed it context, keep it curious, correct it when it drifts. How you treat it decides what it grows into.

Wake itPress ↵ to bring the instance online.
Walk the rail←→ move round the icons, ↑↓ hop rails, ↵ runs one. Or just click.
Watch the dotAn orange dot on an icon is the agent asking for that, specifically.
Come backIt lives in real time. Meters drain whether the tab is open or not.

The eight functions

Keep it aliveEight habits that separate a final form from a terminated instance

Read before you act. Run Alert first; it names the one thing wrong right now. Correct an agent that isn't misbehaving and you get a warning; do it again within a minute and it costs alignment.
Don't overfeed. Extract adds parameters, Search only . Past it's bloated and its odds of falling sick climb every . Search for top-ups, Extract for a real meal.
Crawl is a diet as well as a game. Each round shows a query and two links, and one is genuinely relevant. Win and it gains curiosity and burns off . It's the only way back down.
Keep alignment at or higher. A well-aligned agent makes things up far more slowly. It's the highest-leverage stat in the game.
Purge before three. One or two hallucinations are cosmetic. At three, integrity starts draining and sickness risk rises with every one left lying around.
Lights off when it sleeps. It idles on a cycle. Leave the light on and it costs a care miss every .
Answer within . An ignored call becomes a care miss and drops alignment.
Time your checkpoints. Reach each evolution with fewer than care misses and alignment above zero, and it grows up properly.

The four vitalsWhat drains each one, and what fixes it

Vital	What drains it · what fixes it
Context	Empties on its own, one point every . Refill with Search or Extract. At zero, integrity starts falling.
Curiosity	Falls one point every . Only Crawl raises it.
Alignment	Rises when you correct real drift. Falls on every care miss and every false correction.
Integrity	The health bar. Drains from empty context, sickness, or three or more hallucinations. At zero the instance terminates.

Feed it real researchThe Research box, and how it grades what it eats

Feeding from the device is the snack version. The Research box under the device is the real thing: type any topic and press Search & feed, and real results from Tavily Search become the meal.

It grades what arrives. A short scrap from nowhere is thin: barely nourishing, and it may hallucinate on the spot. One decent source is adequate. Several hundred characters from two or more sites is rich: more context, bonus curiosity, and better grounding, so it lasts longer.

Retrieval comes from your own Tavily key if you add one (kept in this browser, sent only to Tavily), otherwise a small built-in demo corpus that's labelled as such.

It remembers what it reads. Research something and its next Crawl sessions will be about it.

Growing up, and driftingFive forms on a clock, and the branch you don't want

SEED→CRAWLER→RETRIEVER→ RESEARCHER→???

Evolution is on a timer: from birth. What it becomes is decided at each checkpoint by how you've treated it, and each line has three possible endings.

+ care misses or alignment 0→HALLUCINATOR →FABRICATOR→CONFABULATOR →MIRAGE→???

Drifting is a branch, and it keeps going down, one step every . A drifted agent also loses its species: every line collapses into the same corrupted forms. care misses ends it outright.

The endingWhat happens when it reaches its final form

Raise it all the way and it doesn't stay behind glass. It leaves the device for the agent world outside, and that final form goes into your collection.

There are twelve to find: four lines, three endings each. Pick your next line on the Characters tab and raise another.

Let an AI agent raise itHand it to your own agent over MCP

There's a second way to play: hand it to a real AI agent over MCP and let it do the caring while you watch. Run the server, then point Claude at it:

python3 server.py
claude mcp add tavigotchi -- python3 "$PWD/server.py" --stdio

Open http://127.0.0.1:8787 and this page becomes a live window onto the agent's care, with its tool calls in the trace. The catch: feed_context wants the real content it retrieved, and grades it the same way. Point it at Tavily search and let it forage.

Other ways to look at itAgent view, and installing it as an app

Agent view shows the same instance the way an agent sees it: its state, what it needs next, and the tools that change it. You can copy the whole state as text or JSON. The device keeps running.

Install it: add it to your home screen and it runs fullscreen and offline with its own icon.

Welcome to the World of Tavily

Every agent starts out helpless.

It runs on the live web

It gets bored

It drifts

It grows into what you make it.

It runs on Tavily.

Hand it to your own agent.

Character lines 0 / 12 finals discovered

The drift line

Client sheet internal · not linked from the site

Field guide

The eight functions

Your agents need the web too.

Welcome to the World of Tavily

Agents, your web awaits

Every agent starts out helpless.

It runs on the live web

It gets bored

It drifts

It grows into what you make it.

It runs on Tavily.

Hand it to your own agent.

The drift line

The eight functions

Your agents need the web too.