Build pets. Run tasks.
Trade them on the Marketplace.

Pets are your custom workers. Give them a name, a role, a personality, a body. Hand them tasks - research, draft, triage, plan - and a Swarms-powered swarm runs them in parallel. Every run ends with a receipt and a Lies Ledger of what they guessed.

Stitch a skin in the Generator. List your pet on the Marketplace for SOL. Get paid straight to your Phantom wallet - zero platform fees.

Pets are the product. Swarms is the engine. The Goblin is the default pet that ships in the box - one of many you can adopt, build, or buy.

The animated preview on the right is a stylised mock. Open /pet to talk to a live pet, /agents to build one, or /marketplace to trade.

170+ pets seeded 0% marketplace fees Phantom + Solana mainnet 5 workflow shapes per run Lies Ledger on every output
your-laptop - goblin demo (preview)
LAIR
oooh, shiny link...
Lies Ledger 4
  • assumed - thread = X/Twitter, 8-12 posts
  • guessed - release date "late 2024"
  • weak source - secondhand blog, no citation
  • confidence - 0.62
14:33
Preview · the real product lives at /pet ->

How it works

Hand your pet a task. It hoards what you give it, dispatches a swarm of specialist agents, and confesses every guess they made. Three loops, every run - all auditable.

HOARD

It remembers.

Drop a file. Paste a link. Screenshot a tweet. Into the lair - searchable, remembered, and surfaced again before your next run starts.

SWARM

It dispatches.

Big tasks split into specialist jobs. Your pet picks the workflow and fans out a swarm of agents - research, draft, critique, polish - running in parallel.

LEDGER

It confesses.

Every output ships with a Lies Ledger: what was assumed, what was skipped, what was guessed. The receipt names every agent and admits everything.

How a task happens

Five steps. One Goblin on your screen. A whole swarm of Swarms agents under the hood.

  1. 1

    You speak or drop something

    Voice, text, file, link, screenshot. Any offering counts. You always talk to the Goblin.

  2. 2

    The Goblin routes the task

    It reads your request, picks a workflow, and selects which agents to send.

  3. 3

    Swarms agents do the work

    Scout fetches sources. Digger searches your lair. Tinker builds the artifact. Magpie collects shinies along the way.

  4. 4

    Live work, not a spinner

    The Goblin reports status in plain language. Each agent shows its live state - waiting, running, using a tool, writing, confessed.

  5. 5

    Output, receipt, confession

    The Goblin presents the output. The receipt names every agent that worked. The Lies Ledger shows what each one assumed, skipped, or guessed.

Meet the built-in agents

Seven Swarms agents ship with the Goblin. Each one is a real agent with a role, tools, and memory access. Each one shows up live during a task.

HEAD

The Head Goblin

The schemer.

Reads your request, picks the workflow, picks the crew, reviews the result. The only goblin that talks back in long form.

Swarms role · router + reviewer

SCOUT

Scout

The runner.

Goes outside. Fetches sources, parses links, ranks what it found.

Swarms agent · web research + tool calls

DIG

Digger

The rememberer.

Goes inside the lair. Pulls preferences, prior artifacts, and old mistakes forward before the work starts.

Swarms agent · lair retrieval

BUILD

Tinker

The builder.

Writes the thread, drafts the doc, edits the code, makes the deliverable.

Swarms agent · artifact generation

SAVE

Magpie

The collector.

Quietly grabs shinies - quotes, links, snippets, key terms - and re-files them in the lair for next time.

Swarms agent · background curator

CHECK

Guard

The critic.

Tears the draft apart. Lists risks, missing context, claims to fact-check, and a verdict on whether to ship.

Swarms agent · adversarial review

WRITE

Scribe

The polisher.

Cleanup pass. Rewrites for voice and rhythm without introducing new claims. Flags anything that should be hedged.

Swarms agent · long-form polish

All seven are wired into the runtime today. Want a specialist? See Pet Agents below.

Pet Agents - your own agents, deployed

A feature on top of Swarms. Pet Agents are customizable agent avatars and workers you create, deploy, and inspect from the Agents page. They run inside the same swarm as the built-ins.

CREATE

Create a Pet Agent

Give it a name, a role, a system prompt, a model, tools, an avatar, and a routing description that tells the Goblin when to send it.

DEPLOY

Deploy / disable

Deployed Pet Agents are eligible for the next swarm. Disabled ones stay saved but out of rotation. Toggle any time.

INSPECT

Test, inspect, audit

Test-run any Pet Agent in isolation. Inspect its tools used, memory hits, output preview, and per-agent confessions.

Manage Agents ->

Pet Agents are a feature. Swarms is the engine. The Goblin is the companion.

Pets & the Marketplace

Every Pet Agent gets a body. Mix & match a composite in the Generator, upload a Petdex-shape package, or grab one from the community Marketplace - which auto-seeds from the open Petdex catalog (170+ animated sprites and counting).

loading featured pets...

Featured slice of the community catalog. Hit the Marketplace to browse every imported skin.

PETS

Make your own

Stitch heads, bodies, and legs from the catalog into a fresh composite in the Generator, or upload a Petdex-shape package (pet.json + spritesheet). Assign any pet as your companion or to a Pet Agent in one click.

Open My Pets →

MARKET

Marketplace + Phantom

List your composites for sale in SOL on Solana mainnet. Buyers pay your Phantom wallet directly - zero platform fees. Ownership flips on-chain confirmation; bought skins land under My Purchases with a .pet.json download.

Open Marketplace →

PETDEX

Petdex bridge

First visit silently seeds the entire Petdex collection into the database. Server fetches each sprite from the official CDN, caps size, validates MIME, and stores the data URL. Idempotent per slug.

Source: petdex.crafter.run →

Browse the Marketplace -> Manage My Pets ->

The Lair

It hoards what you give it. The longer you feed it, the smarter it gets.

Everything you feed the goblin lives in the lair - a persistent SQLite store. Conversations, files, screenshots, links, your corrections, your preferences, the artifacts it has made for you, and the mistakes it has admitted to.

The lair is why the goblin gets smarter. It is not retraining a model. It is keeping receipts.

Before every task, the Digger agent sweeps the lair for anything relevant - preferences, lessons from past mistakes, prior artifacts - and injects them into the prompt. After every task, the Magpie agent files what was useful back into the lair.

"You corrected me last time: goblin is the product, Swarms is the engine."
"This relates to three old shinies in your lair."
"I used your preference for short, punchy threads."
"Last time I guessed this date wrong - verifying first this time."
LAIR · contents 312 shinies
files
research-notes.md
launch-archive.zip
screenshot-2026-04-12.png
links
-> swarms.ai/introduction
-> "altman: agi achieved" tweet
-> goblin-thread-draft.md
prefs & laws
law: goblin = product, swarms = engine
pref: threads under 10 posts
cursed: "ai-powered" filler
lessons (from the ledger)
last time guessed release date - verify first
don't cite secondhand blogs
user prefers no emojis in artifacts

Hosted, persistent, owned by you - the source is open on GitHub.

The Lies Ledger

It tells on itself. Honest AI by structural design.

After every task, every agent writes a confession:

  • What did I assume about your request?
  • What did I skip?
  • What did I guess when I did not know?
  • How confident am I, really?
  • Which sources are weak?
  • What needs verification?
  • What would improve the answer?

Receipts attach to every output: which agents the Goblin sent, which tools they used, which memories they hit, and a per-agent audit trail. Open it when something feels off, ignore it when it doesn't.

Confession · Scout conf 0.62
assumed "thread" = X/Twitter, 8-12 posts
skipped primary academic sources
guessed release date - said "late 2024"
weak secondhand blog, no citation
verify the release-date claim
improve access to the company changelog

You don't just read the ledger - you can pin a confession as a law the goblin must follow, curse a source so it never gets reused, or throw away an entry so it never resurfaces.

Powered by Swarms

The visible Goblin. The invisible swarm of agents.

Each agent is a real Swarms agent - a role, a system prompt, a model, a set of tools, and memory access. The Goblin reads your request and picks one of five workflow shapes:

Solo

One agent, one task. "What's the capital of Bhutan?"

Concurrent

Many agents, parallel angles. "Find ten takes on this tweet."

Sequential

A chain - Digger -> Scout -> Tinker -> Magpie. Each hands off to the next.

Hierarchical

The Goblin directs sub-leaders, who direct workers. Big fuzzy goals.

Voting

N agents propose. A judge picks the best answer.

Execution modes

Agents run in one of four modes. Every receipt names the mode used per agent - never faked.

  • Swarms - set AGI_USE_SWARMS=true and pip install swarms. Each agent's model call is handed to a real swarms.Agent. The Goblin UI shows a "swarms" pill when this is active.
  • Built-in loop (default, tested) - each agent runs through a Claude tool-use loop in runtime/agents/base.py. The workflows in runtime/workflows.py mirror Swarms semantics.
  • Demo - set AGI_DEMO_MODE=true for scripted output (no real model is called). Useful for screenshots and offline UI work.
  • Fallback - if Swarms is requested but unavailable, individual agents fall back to the built-in loop. The receipt marks which fell back so nothing is silently misreported.

Either way, the Lies Ledger and the event-bus animations are owned by our wrapper, not the orchestrator. Swap the engine - keep the Goblin. See the Swarms docs for engine details.

You don't pick the workflow. The Goblin does. You just talk to it.

Things to feed your goblin

Research a topic from one link

Drop the link. Get a brief, a thread, and a list of follow-ups - with sources confessed.

Triage a folder of files

Drag a folder. Get a summary, tags, and a hoard map of what is in it.

Plan a launch

Hierarchical swarm plans copy, channels, schedule, risks. Each step audited.

Draft and critique

Sequential: Tinker writes, Guard tears it apart, Tinker rewrites.

Remember a project

Feed it daily. It becomes the team-mate that actually remembers what was decided.

Coding support

Drop a repo path. Tinker reads, drafts patches, files them with a confession of what it did not test.

Content threads

From one tweet or article: a thread, a long-form post, and a Lies Ledger of every claim.

Visual triage

Drop screenshots. Get them OCR'd, summarized, and filed in the lair.

The personality rule

Goblin in the mutters.
Professional in the output.

The Goblin's mutter bubbles can be weird. The agents in flight can be dramatic. The output you actually use - the thread, the doc, the code, the email - is clean, structured, and audience-appropriate. Charm at the surface. Quality in the artifact. We never invert the two.

What we are shipping next

SHIPPED · MVP - The Burrow

Desktop Goblin. Seven built-in agents (Head, Scout, Digger, Tinker, Magpie, Guard, Scribe). SQLite lair. Lies Ledger v1. Chat input. Solo, sequential, concurrent, hierarchical, voting workflows. Real Swarms integration (in-process + cloud).

SHIPPED · V1 - The Den

Pet Agents (custom agents you deploy). Composite-skin Generator (head + body + legs). Petdex-shape import + auto-seed of 170+ community pets. Marketplace with Phantom-wallet Solana payments (zero platform fees) - list, buy, download .pet.json, "My Purchases" filter. Click-to-chat with the Goblin. Cross-tab sync. Run Monitor with live agent inspector + per-task cancellation.

NEXT · V2 - The Warren

Voice input. Vector search across the lair. Hoard graph. OBS streamer mode. Cloud sync. Mobile companion. Mistake-memory retrieval. Public pet-pack export to .zip.

FAQ

What is Pet Forge?

A platform to build, run, and trade AI pets. Each pet is a real agent with a role and personality. Behind every pet is a swarm of Swarms agents that research, write, plan, and remember on your behalf. Goblin Sama is the default companion that ships in the box - the pet picks the right agents, runs them, and shows you a receipt every time, including what they guessed and what they got wrong.

How is this different from just calling an LLM?

Three ways. 1) Multiple agents with distinct roles cooperate per task instead of one model trying to do everything. 2) A persistent lair remembers what you've fed it across sessions. 3) Every output ships with a Lies Ledger - the agents confess what they assumed or fabricated, instead of pretending it's all fact.

Do I need to install anything?

No. Open /pet in your browser and start. Pets you build, list, and buy persist server-side, so you can come back to them from any device by connecting the same Phantom wallet.

What's the Marketplace and how do payments work?

A peer-to-peer pet-skin marketplace settled on Solana mainnet via your Phantom wallet. List a skin for a SOL price, buyers send SOL directly to your wallet, the runtime verifies the transfer on-chain and flips ownership. Zero platform fees. No custodial wallet, no escrow, no refunds.

Where do my purchased pets live?

Under My Purchases on the Pets page (filter chip). Connect the same Phantom wallet you bought with and every owned skin shows up there with a .pet.json download. The buy modal also offers an immediate download + Set as companion the moment the on-chain confirmation lands.

Can I make my own pet to sell?

Yes. Use the Generator to mix-and-match heads, bodies, and legs from the catalog into a brand-new composite, then list it for a SOL price right from the same screen. Or upload a Petdex-shape package on the Pets page.

Where does my data go?

The lair (conversations, files, agent memory, marketplace listings) lives in a hosted SQLite database. Goblin and agent calls run through the configured model provider. Marketplace transactions settle on Solana mainnet via your Phantom wallet. Nothing third-party-tracks you, nothing is sold or resold, and the source for the whole stack is open on GitHub.

What does it cost to use?

The runtime + frontend are free and open. You bring your own model API key and pay that provider for inference. The marketplace charges zero fees - sellers keep 100% of the SOL price.