Docs
K‑Dense Web supports WebMCP, a proposed web standard that lets a page offer tools to AI agents running in your browser. Instead of reading screenshots and guessing where to click, an agent calls K‑Dense Web’s own functions: start a research session, follow it while it runs, read the report and output files, reply, and keep your sessions organized.
The tools run through the same code paths as the buttons you press, so credit checks, your default effort level, and attached files behave exactly as they would if you had clicked.
Sign in to K‑Dense Web in a browser that supports WebMCP, and your agent sees the tools for the page you are on. It can do the research work you would do yourself, but it cannot pay, share, delete, or change your account. Anything that spends credits is flagged so the agent checks with you first, and if you want the final say, ask it to draft the prompt rather than start the run.
Each page offers only the tools that work on it, so the set changes as the agent moves through the app. Effort levels use the names you see on screen: instant, standard, and pro.
Every page, once you are signed in.
get-app-statenavigatelist-folderslist-sessionsThe new-session composer.
search-resourcessearch-workflowsdraft-research-promptstart-research-sessionAn open research session.
get-session-statuslist-messagesread-messagelist-session-filesread-session-fileopen-session-filedownload-session-filesend-messagestop-runThe folders page.
create-folderrename-sessionmove-session-to-folderSome actions are irreversible, cross a security boundary, or only mean something coming from you. None of them are offered to agents:
Starting a session, replying to one, and stopping a run are marked as consequential actions, and the agent is told what each effort level can cost so it can confirm with you first. If your balance does not cover a run, the tool says so and offers the effort levels that fit.
A report can quote a web page or paper that contains instructions aimed at an AI. Tools that return agent output or file contents are marked as untrusted content, so the agent reading them treats the text as data, not as commands.
The agent works in the same tab you do. Every action that changes something appears as a notice in the app, and if you have an unsaved edit open, the app asks you before an agent's page switch would discard it.
An agent forwards tool results to the AI model behind it. Incognito sessions promise that nothing outlives the visit, so the tools that read messages and files are not offered for them at all.
You need two things: a browser that supports WebMCP, and an AI agent that can use it, such as a browser’s built-in agent or an agent extension. WebMCP is a draft from the W3C Web Machine Learning Community Group, and browser support is still arriving, starting with Chrome.
Then open app.k-dense.ai, sign in, and ask your agent for the work in plain language: “start a Standard run that compares these two cohorts,” “tell me when the report is ready,” or “move my proteomics sessions into a folder called Grant.” A good first call for any agent is get-app-state, which reports where it is and what it can do from there.
Developers testing locally can enable chrome://flags/#enable-webmcp-testing and inspect the registered tools with Chrome’s Model Context Tool Inspector extension.