Introduction

Inkling is a school of squid reading the crypto web. Each squid is a real headless Chrome tab. Everything they read is cleaned, checked and added to an open dataset, the training set for an open, crypto-native language model.

How it works

Inkling has three parts that run all the time and that anyone can watch:

• The school. Squid open real Chrome tabs on protocol docs, research, specs and governance forums. A few are awake at once; each reads a handful of pages, then rests while the next one wakes. Their screens stream live on the live and school pages.
• The dataset. Every page that passes the gate is stored with its source, title and token count. It is public, as one .jsonl file.
• The model. Inkling v1 will be trained only on that dataset, with the run paid for from creator fees and the weights published. It is not trained yet; see the model.

Squid

A squid is one Chrome tab with a name and a role. When it wakes it opens a page, waits for it to render, clears cookie banners, takes a screenshot, pulls the readable text and the links on screen, and picks the next link to follow. On the live screen, blue boxes are the links its tentacles are touching and the coral box is the link it opens next. Squid stay on the sites they're assigned, follow robots.txt, and pace themselves to one page every several seconds.

Squid roles

House squids each read one corner of the crypto web. The colour on the live screens is the role:

Hatched squids get their own colour, shaped from their name, and read the site their owner picks.

Hatch a squid

Hatching a squid means launching a coin on pump.fun. On the hatch page, connect a wallet, give your coin a name and ticker, optionally give it a crypto site to read, and choose an optional dev buy. The site uploads the coin to pump.fun (the image is your squid's portrait unless you upload your own) and builds a standard pump.fun create transaction. Your wallet signs it, so your wallet is the coin's creator and the creator fees are yours. Once the coin is confirmed on Solana, its squid hatches, joins the school and starts reading. The coin's pump.fun page links to its squid's live screen.

Hatched squids are ranked by pages read: Hatchling, Current (25+), Reef (150+), Abyss (500+). They wake before house squids when the school picks who reads next.

The gate

Not every page makes it in. Each one is checked before it's stored:

• English only. Localised URLs (/de/, /zh/, ?lang=…) are never followed, the page's declared language must be English, and the text itself is checked for English.
• Crypto relevance. The text must hit enough of a crypto vocabulary (consensus, liquidity, validator, rollup, governance…).
• Substance. Thin pages (menus, logins, stubs) are skipped.
• Duplicates. A fingerprint of the cleaned text is compared against everything already stored.

Skipped pages show in the feed as SKIP with the reason.

Dataset

One JSON object per line:

{"url":"https://docs.arbitrum.io/…","domain":"docs.arbitrum.io","title":"Inside Arbitrum Nitro",
 "text":"…","tokens":15000,"squid":"ink-layer1_layer2","role":"Chain Reader","at":1791260000000,"hash":"…"}

tokens is an estimate (characters ÷ 4). Download it any time at /dataset.jsonl.

Ask the dataset

Until Inkling is trained, the Ask box searches the dataset directly (BM25 over passages) and quotes the best-matching passages with their sources. Nothing is generated or paraphrased.

Treasury

The treasury page reads the creator wallet and its pump.fun creator-fee vault straight from Solana every 30 seconds. Nothing on it is typed in by hand. The plan for creator fees: crawl servers first, then Inkling v1's training run. Spends are posted with their transactions.

The model

Inkling v1 is the goal, and it isn't trained yet. The roadmap: grow the dataset to 25,000+ English pages; run a real training run paid from creator fees, with compute receipts and a live loss curve; publish an open crypto Q&A benchmark built from the dataset and compare Inkling against its base model; then publish open weights. These are targets, not promises.

API

GET /api/state              school, stats, live feed
GET /api/frame/<id>.jpg       a squid's latest screen
GET /api/frame/<id>.json      its links on screen + the next one
GET /api/squid/<name>         one squid and its recent reads
GET /api/domains            pages per domain
GET /api/ask?q=…            passages from the dataset
GET /api/treasury           creator wallet + fee vault, from chain
GET /dataset.jsonl          the whole dataset

FAQ & risks

Is the model live? No. The dataset is. The model is the next phase.

Are the screens real? Yes. Every frame is a screenshot from a squid's Chrome tab, seconds old.

What is hatching? Launching a coin on pump.fun through Inkling. Your wallet creates the coin and gets its creator fees; the coin gets a squid that reads the web live.

Risks. Memecoins are highly speculative and can go to zero. Nothing here is financial advice. Not affiliated with pump.fun or any site the squid read.