For most of the history of equity research, the reader has been a person. An analyst writes a note, it goes out as a PDF, and a fund manager opens it in an inbox or on a platform. Every part of the distribution chain was built around that moment.
That is changing. Buy-side firms are putting AI agents to work on the reading. An agent can screen a watchlist overnight, pull the latest notes on twenty companies, compare forecasts and hand a portfolio manager a short brief before the market opens. The person still makes the decision. The agent does a growing share of the reading that leads up to it.
This is a new class of research reader, and it has very different needs from the one the industry built for.
What changes when the reader is software
A person can cope with friction. They can log in to five portals, remember which broker sends what by email, and squint at a table in a PDF. An agent cannot do any of that well, and it should not have to.
An agent needs three things.
- Structured access. It needs to ask a question and get an answer in a form it can work with: which notes exist on this company, what has changed since last week, what each broker forecasts for revenue next year.
- Entitlements it can trust. It must only see what its organisation is entitled to see. Not more, because that breaks the provider's commercial terms and the reader's compliance rules. Not less, because then the brief is wrong.
- One place to ask. If an agent has to be wired separately to every broker and every platform, it will only ever be wired to a few of them, and its view of the market will be partial.
Why the current model breaks
Since MiFID II unbundling, research distribution has spread across many platforms, each with its own integration, its own login and its own format. For a human reader that is an annoyance. For an agent it is a wall.
PDFs by email are the clearest case. An inbox has no idea of entitlements, no structure and no history. An agent reading a mailbox can extract text, but it cannot tell whether a forecast was superseded, whether a note was withdrawn, or whether the reader was ever meant to have it.
Platform-by-platform integration is the other case. Each connection is a separate project, with a separate key, a separate data model and a separate set of rules. A buy-side firm building an agent faces that work once per source. A sell-side firm that wants its research read by agents faces it once per destination. Neither side gets far.
What entitlement-aware, machine-readable access looks like
The answer is not to give agents a back door to research. It is to give them the same front door a human reader uses, with the same rules applied, in a form they can read.
In practice that means:
- every read is checked against the provider's entitlements, per organisation and per user, by sector, company, analyst or entitlement code
- an agent acting for a firm sees exactly what that firm is entitled to and nothing else
- notes, metadata and forecasts come back as structured data, not just a file
- supersedes and withdrawals are reflected, so an agent does not quote a note that has been replaced
- providers can see agent reads separately from human reads, so they know who, and what, is consuming their work
That last point matters more than it sounds. If agents are going to read research, the people who write it should be able to see that happening and count it properly. On Research Protocol, reads by people and reads by AI agents are reported separately, and an AI-generated quote of a note is counted as its own type of consumption rather than folded into document reads.
What Research Protocol offers today
Research Protocol is a neutral hub for investment research. A sell-side firm connects its library once and it becomes available, under its own entitlements, to every connected reader, human or agent. A buy-side firm connects once and can query across every library it is entitled to.
There are three ways in, and an agent can use any of them.
- An MCP server. The Model Context Protocol is how agents such as Claude and in-house agents discover and call tools. Our server at
/mcp-server/mcpexposes 40 tools, so an agent can search coverage, read research its organisation is entitled to and pull consensus forecasts. Details are on the MCP page. - A REST API with more than 200 endpoints, for teams building their own pipelines. See the API page.
- A command-line tool,
rpt, with 40 commands, for developers and analysts who prefer a terminal. See the CLI page.
Agents can also find us without being told. We publish machine discovery files at /.well-known/agent.json, /.well-known/mcp.json and /llms.txt, so an agent framework that looks for them knows what the hub offers and how to connect.
Beyond single notes, the hub gives an agent something no single source can: per-broker forecasts in each broker's own line-item structure, and cross-broker consensus with the median, interquartile range, outlier flags, rating distribution and consensus history. A bank's own API can tell an agent about that bank's research. It cannot tell it what other brokers think about the same stock.
Getting started
We have made the first step as short as possible. There is no sales call. An agent developer can get a free self-service key at the agents page. The free tier allows 1,000 requests a day and 20,000 a month, with attribution "via Research Protocol" required on anything the agent shows.
If you are building an agent that reads research, get a free key and point it at the MCP server. If you want to talk about a larger deployment, email [email protected].