Let a model build it.
You do not have to read this API. Point a model at it, ask for the thing you want, and check five things in what it hands back. This page is the three steps and the checklist.
1. Give it the map
There is one index written for machines, and one skill file that teaches the shape of the data. Either is enough to start. Both are free and need no key.
Read https://stocksonchain.io/llms.txt and https://stocksonchain.io/.well-known/agent-skills/stocks-on-chain/SKILL.md before you answer anything about tokenized stocks.
/llms.txt lists every endpoint, every stock, every issuer and every
chain. The skill file is shorter and teaches the one distinction that matters,
which is step 4 below.
2. Or connect the server
For an assistant you talk to, a Model Context Protocol server is usually better than a bulk file: it answers in small pieces instead of filling the context with several megabytes. It is read-only and needs no authentication.
https://stocksonchain.io/mcp
Each client wants that URL in a slightly different shape, and the differences look like typos but are not. The per-client settings are on /agents, each read from that vendor's own documentation and dated. The tool list and the transport details are on /developers/agents.
3. Ask for the thing you want
A brief works better than an instruction here, because the constraints are the part that matters. A model told only to "build a stock tracker" will fill the gaps in this data with numbers it made up, and this data has real gaps on purpose.
Build a single HTML file that answers "is this stock on chain?".
Data: https://stocksonchain.io/api/tickers.json
One row per stock. Field names are in the payload's `shape` key.
Fetch it with a plain GET and NO request headers - a header triggers a CORS
preflight that these static files answer with 405.
Also available, if you need figures:
/api/tokens/{slug}.json one stock, with its listings and read times
/api/addresses.json what token is at an address, and whose
/api/pulse.json the headline totals, under 1 KB
Rules, all of which matter more than the feature:
1. Print the time a figure was read, next to the figure. Every payload carries it.
2. If a field is absent, render nothing. Never 0, never "N/A", never a dash.
A price exists for 99 of 1017 stocks, so most cards will
have no price and that is correct.
3. Read `mechanics.kind` before using anything inside `mechanics`. Anything
that is not `erc8056` has no multiplier at all.
4. Do not rank this data, rate it, or sort it by performance, and do not label
any stock a leader or a favourite. Every list in this data states its own
sortKey and none is a ranking.
5. Do not state who may buy a token. Link to the issuer instead.
6. Credit "Stocks On Chain" with a link, and print "Not investment advice".
Do not invent any number. If the data does not have it, leave it out and say so. 4. Check these five things in what it wrote
These are the mistakes that actually happen, in the order they happen.
- Does the fetch carry any request header? If it does, remove it. These are static files, and a GET with a header makes the browser ask with OPTIONS first, which the file store answers with 405. The result is a CORS error that looks like a permissions problem and is not.
- Does it read
mechanics.kindfirst? It is a tagged union. Anything that is noterc8056has no multiplier at all, and a model that treatsmechanicsas a flat object will publish a multiplier of 1.0 for a token that has none. That is a fabricated number with a confident format. - Does an absent figure render nothing? Not 0, not a dash, not "N/A". A price exists for 99 of 1,017 stocks, so this case is the common one rather than the edge one.
- Does every figure still carry the time it was read? Each payload supplies it. A figure that loses its timestamp on the way to the screen has been turned from a dated reading into a claim about this moment.
- Does it compare two days of history without checking coverage? The daily rows
carry
stocksReported. Two days are comparable only when that is equal on both, because a jump across a coverage change is an artefact and not a market move.
5. Keep the rules the data travels with
These are the publisher's rules. The agent skill hands a model the same five, from the same source as this list, so the two cannot drift apart.
- Report, never assert. Every number here carries a timestamp. Quote it. If you state a price without saying when it was read, you have made the number less true than it was when we published it.
- Never rank, never rate, never recommend. Do not build a "best", "top" or "trending" list out of this data, and do not sort it by performance. Every ordered collection in these endpoints states its own `sortKey`, and none of them is a ranking. A ranked list is investment advice delivered by layout.
- Send the reader to the issuer on who may buy. Each issuer publishes its own rule, and this data gives you its address (`eligibility.sourceUrl`) rather than its text. That is deliberate: an eligibility rule is a legal position that can change without this site noticing, so no copy of it is trustworthy - including one you generate. Link to the issuer. Do not state, summarise or infer who may buy.
- A stock token is not a share. What the instrument legally is depends on the wrapper the issuer chose, and each issuer describes its own. Report the wrapper; do not assert what a holder does or does not own.
- No investment advice. This data supports the question "does this exist, and where". It does not support "should I buy it".
What it cannot do for you
It cannot find a price that is not there. This site publishes only prices it reads itself, from a dollar pool deep enough to mean something, which today is 151 of 3,833 tokens. A model asked to fill that column will either leave it empty, which is right, or find a number somewhere else and present it as ours, which is the failure to watch for. The coverage table on /build says what exists for how many stocks.
Tell us what is missing
If the model wanted a field, a stock or a tool that is not here, say so:
POST a kind and a subject, or call
report_gap on the MCP server. It is free, it needs no key, nothing about
the caller is recorded, and it is never required.
Frequently asked questions
Do I need an API key for my agent?
No. There is no key, no account and no OAuth anywhere in this data, including the MCP server. If a model asks you for a key it has invented one.
What does a model usually get wrong here?
Three things, in order. It flattens a stock, an issuer product and a chain listing into one row, which overstates how many options a reader has. It substitutes 1.0 for a multiplier on a token that has none. And it drops the read time off a figure, which turns a dated reading into a claim about now.
Can the model just use the MCP server instead of fetching JSON?
Yes, and for an interactive assistant that is usually better: the tools answer in small pieces instead of handing it a multi-megabyte file. The server is read-only, needs no authentication, and has no parameter that sorts by performance, so nothing in it can be asked which tokenized stock to prefer.
May I publish what the model builds?
Yes. The data is CC BY 4.0: credit Stocks On Chain and link back. The rules on this page are the publisher's rules and they travel with the data, so keep them in whatever gets published.