A DAO that nobody can read is, for most practical purposes, a private company with a public ledger. Its members exist as addresses in a contract, its decisions as calldata, its treasury as balances no outsider can attribute. The gap between "the data is on-chain" and "a stranger can read the organization" is filled by a description the organization publishes about itself – and increasingly, that description is written for software rather than for people.
Four such surfaces now exist, and they arrived from four unrelated directions. One came out of DAO governance itself. One came from an AI researcher's blog post. One came from an agent-interoperability working group. One came from a model vendor's protocol. None of them was designed with the other three in mind, none of them supersedes the others, and a DAO wanting to be legible to machines in 2026 has to decide, without much guidance, which of them to publish.
This page describes the four, then reports a census: on 12 August 2026, every organization in this wiki's DAO directory was probed for all four. The results are the useful part, and they are not what the volume of writing about agentic infrastructure would lead you to expect.
daoURI() and the DAOIP family. The oldest and the only one built for DAOs specifically. EIP-4824, authored through DAOstar and created in February 2022, defines a single on-chain function that returns a URI pointing at a JSON-LD document listing the DAO's members, proposals, activity log, governance rules, and contracts. It is the only one of the four whose vocabulary is about governance: it has a field for a proposal's on-chain calls, and a field for who voted. It is also the only one with an on-chain discovery mechanism, an index contract that emits a registration event. Four and a half years after creation it still carries the status Review, and a census of its four production index deployments found 30 registrations, the most recent in February 2025.
/llms.txt. Proposed by Jeremy Howard in September 2024, this is a Markdown file at a site's root whose only mandatory element is an H1 with the project name. Everything else – a blockquote summary, prose sections, H2-delimited lists of links with one-line notes – is optional. It is explicitly a community proposal rather than a standard: no registry, no validator, no conformance suite. What it has instead is reach. Its own documentation notes that thousands of sites publish one, that OpenAI, Anthropic and Google ship them for their documentation, and that Chrome's Lighthouse now audits for it. It says nothing about governance because it says nothing about anything in particular – it is a table of contents for a language model.
The A2A agent card. The Agent2Agent protocol, now published under the Linux Foundation, defines an AgentCard object served at a registered well-known URI. The registration entry names the suffix agent-card.json, with the A2A Protocol Working Group as change controller. The card declares what an agent can do, what protocols and authentication it supports, and what skills it exposes. The path is itself a migration hazard: version 0.3 renamed it from /.well-known/agent.json, and a spec-compliant 1.0.0 client will not look at the old location.
An MCP server. The Model Context Protocol is the most capable of the four and the least like a file. Instead of publishing a document, an organization runs a JSON-RPC endpoint that exposes tools, resources and prompts, so an agent asks questions rather than downloading a snapshot. That difference matters for a DAO: a document describes the treasury as it stood when the document was generated, while a tool answers about the treasury now.
The four surfaces are cheap to probe, so on 12 August 2026 every organization in this wiki's DAO directory was checked for all of them. The directory holds 71 organizations, 70 of which record a website (ConstitutionDAO, dissolved in 2021, does not). Each domain was requested at /llms.txt, /.well-known/agent-card.json, /.well-known/agent.json and /.well-known/mcp.json under a browser user-agent, following redirects.
The probe needs one discriminator to be worth anything. Most of these sites are single-page applications that answer every path with HTTP 200 and their own HTML shell, so a naive status check reports a file that does not exist. The test used here is the response's content type: a real llms.txt comes back as text/plain, a real agent card as application/json, and the SPA fallback as text/html. As a control, each positive was re-probed at a nonsense path (/llms-caper-control-9f3a.txt); fifteen returned 404 and three – yearn.fi, rocketpool.net and cosmos.network – returned 200 with text/html, exactly the behaviour the content-type test is there to catch.
Four domains could not be resolved either way and are excluded from the denominators: balancer.fi returned HTTP 429, sky.money and jito.network returned Cloudflare 403 challenges, osmosis.zone answered the well-known path with a redirect loop, and ether.fi returned HTTP 500 for /llms.txt. That leaves 66 organizations that answered cleanly.
| Surface | Published | Of 66 answering |
|---|---|---|
/llms.txt | 18 | 27% |
MCP server card at /.well-known/mcp.json | 1 | 1.5% |
| A2A agent card (either path) | 0 | 0% |
The eighteen serving /llms.txt are 1inch, Aragon, Cosmos Hub, Kleros, Lido, Maple Finance, Ondo Finance, Pendle, Polygon, Rocket Pool, Safe, The Graph, THORChain, Euler, Helium, Starknet, The LAO and Yearn. The zero for agent cards is not an artefact of the probe: thirteen domains returned HTTP 200 for one of the two agent-card paths, and all thirteen were the SPA HTML shell. That result is consistent with the wider web. A survey published on 29 July 2026 probed 22,341 hosts drawn from the APIs.io catalogue, of which 20,185 answered, and found 65 serving any agent card at all – 0.29 percent – of which only 10 passed every structural check in the A2A 1.0.0 AgentCard object, and 15 were still sitting at the pre-0.3 agent.json path.
The single MCP result is worth reading closely, because it shows how thin the line between publishing a surface and operating one can be. THORChain serves a well-formed server card at thorchain.org/.well-known/mcp.json declaring a read-only endpoint named thorchain-public-discovery, three tools (get_site_summary, get_agent_rules, list_public_api_references), no required authentication, and an HTTP transport at thorchain.org/mcp. It also serves an AGENTS.md at its root describing the rules it asks AI clients to follow – a fifth convention, and one no other organization in the directory publishes either.
Two details qualify it. The card's $schema points at modelcontextprotocol.io/schemas/server-card/v1.0, which returns 404 – the MCP specification, revision 2026-07-28, defines a wire protocol and a registry, not a well-known server-card URI, so this is a convention the publisher has adopted rather than a spec it conforms to. And a JSON-RPC initialize posted to the declared endpoint returns the site's HTML, not a protocol response: /mcp serves the same static document as the well-known path. THORChain has published discovery metadata for an MCP server. It has not, as of this probe, put a live MCP server behind it.
That is not a criticism so much as a measurement of where the field is. The most agent-forward organization in a 70-strong directory of the largest DAOs in existence has shipped a JSON file and a Markdown file describing an interface that does not yet answer.
The machine-readable governance record most DAOs already have is one they do not publish themselves. Snapshot's public GraphQL API returns every space's settings, proposals and votes, and it was answering 540 million requests a month in 2023, when Snapshot put it behind keys to stop runaway bots taking it down (Snapshot Labs). Keyless callers now get 100 requests a minute and keyed callers 200,000 a month, a tenth of the 2 million a month offered when keys were introduced (Snapshot docs, read 19 September 2026). A second API at api.snapshot.box serves Governor and Snapshot X proposals, votes and execution history from one schema, and its indexer can be self-hosted (sx-monorepo). Snapshot's MCP server wraps the same data for agents.
For most DAOs, then, the readable governance surface exists already, is run by their voting vendor, and is rate-limited by that vendor. None of the eighteen llms.txt files above links to it as data.
The grants layer has gone the same way. Karma published agent skills in March 2026 so an agent can find funding opportunities, create a project and post milestone updates without opening the site, and its research tool over US nonprofit filings answers over MCP inside ChatGPT or Claude (Karma; Karma). For a DAO that runs a grants programme on such a platform, the machine-readable surface for its funding is the vendor's, on the vendor's terms.
The more revealing finding is not the count but the content. All eighteen llms.txt files were retrieved and read. Every one of them is a product summary. They describe what the protocol does, list the documentation, and point at the app.
Not one names a live proposal. Not one lists a member or a delegate. Not one states a treasury balance. Four of the eighteen link a governance venue at all – Cosmos Hub to forum.cosmos.network, Euler to its forum plus Snapshot and Tally, Kleros to its forum and Snapshot space, Yearn to gov.yearn.fi – and in every case the link is to the venue's human-facing front page, not to a machine-readable proposal list. Seven files mention the words proposal, treasury, voting or delegate anywhere; the mentions are almost all links to marketing pages about a governance token.
One of the eighteen is not even a deliberate act of publication. Starknet's file opens with a byte-order mark and the line "Generated by Yoast SEO v27.3, this is an llms.txt file, meant for consumption by LLMs" – it is the output of a WordPress SEO plugin, an artefact of the marketing site rather than a decision by the organization.
Put together with the EIP-4824 registry, which is the one surface with governance vocabulary and has 30 registrations across its entire history, the pattern is consistent: the surfaces that describe governance are barely adopted, and the surface that is adopted describes the product. An agent that reads every machine-readable file the largest DAOs publish will come away able to explain how to swap a token and unable to say who decides anything.
Each of the four surfaces resolves, in the end, to a hostname. daoURI() stores a pointer on-chain and the document lives off it. llms.txt is a file on a web server. An agent card is a file on a web server at a more specific path. An MCP endpoint is a running service, which is the same dependency plus a process that has to stay up.
This is the single structural weakness the whole category inherits, and it is not hypothetical. The EIP-4824 registry is the worked example: an immutable on-chain pointer, registered once, aimed at an off-chain document that eventually stopped answering, with nobody obliged to notice. It is the same mechanism documented on DAO tooling discontinuity – a DAO's legibility outsourced to infrastructure that runs on somebody's renewal cycle, and quietly ends when a grant does.
The asymmetry is what makes it dangerous rather than merely annoying. The pointer is durable; the thing it points at is not. A registration event from 2023 sits on Ethereum forever, advertising a description that has been gone for a year, and there is no on-chain signal that distinguishes the two states. Anyone specifying the next such standard has an obvious lesson available: a pointer with no liveness check is a claim, not a fact.
The census supports a short, unglamorous set of conclusions.
llms.txt, but do not mistake it for governance transparency. It is the cheapest of the four, it has real reach, and 27% of the largest DAOs already have one. It is also, as published today, a marketing artefact. A DAO can distinguish itself here almost for free by including what everyone else omits: the governance forum, the voting venue, the treasury address, and the proposal list, as links a machine can follow./.well-known/agent-card.json, since a 1.0.0 client will not read the legacy path.Every caper is a DAO, and the platform publishes all four surfaces – which makes it a fair test of the argument above rather than an exception to it.
Read again on 23 October 2026: caper.network/llms.txt returns the index, and is the only one since caper.network removed llms-full.txt, its whole-corpus file, on 21 October 2026; every wiki page also answers at its own URL with a .md suffix as text/markdown; each DAO has its own per-DAO llms.txt and its own EIP-4824 dao.json; there is an agent card and an OpenAPI document; and the MCP endpoints are live rather than declared – a tools/list call to /api/mcp returns sixteen tools (nine reading, two for the ROLA sign-in and five that need the token it returns) and the per-DAO endpoint at /capers/{cashtag}/mcp returns three, all of them chain reads: get_ledger for treasury, timeline, grants and inter-DAO funding, get_market for live price and supply, and quote_swap, which prices a hypothetical trade of a given size against that DAO's own bonding curve. The full surface is documented at Querying a caper from an AI agent.
That last tool is the sharpest illustration of the endpoint-versus-document argument, because there is no document that could carry it. An agent asking quote_swap what a trade of a given size buys is not reading a number someone published; it is running the curve. A file cannot answer a question whose answer depends on the size of the question.
One honest deduction still applies against the standard set out above, and one has been retired. Retired: this page previously recorded that Caper's agent card sat only at the legacy /.well-known/agent.json path, where a spec-compliant A2A client would not find it. Re-probed on 17 August 2026, /.well-known/agent-card.json returns the identical 4,501-byte document, so the defect the July 2026 survey found in 15 of the 65 cards it located no longer describes this one. Still standing: a caper ships proposals: [] and activityLog: [] in its dao.json where the specification says a field with no value should be omitted rather than left empty.
From 17 August 2026 this page reproduced that against /capers/XRD/dao.json, then the one caper in the directory with no members; the genesis redeploy of 18 September 2026 removed that caper, and the URL has returned 404 since. Read again on 20 September 2026, the $CAPER document is 2,072 bytes carrying three members, seven contracts and both other arrays empty, so the empty arrays do not depend on a caper being new: the route emits them for any field with nothing in it.
What does hold is the fourth conclusion. Every one of those documents is generated from ledger and platform state on request rather than stored, so there is no snapshot to go stale, and the data underneath is Radix component state: the members, the proposals, the votes and the treasury exist on-ledger whether or not caper.network serves anything at all. If the site stopped answering tomorrow, what would be lost is the convenience of the common interface, not the record. That is a narrower promise than "the DAO is machine-readable", and it is the part of the promise the last four years suggest is hardest to keep.
agent-card.json well-known URI suffix./llms.txt, /.well-known/agent-card.json, /.well-known/agent.json and /.well-known/mcp.json under a browser user-agent with redirects followed; positives re-probed at a control path and separated from single-page-application fallbacks by response content type; all 18 llms.txt files retrieved and read in full.