AI Agents Native to the Browser.

Stop dealing with the security nightmares of server-side API keys and brittle Chrome extensions. Aeion WebMCP extends the Model Context Protocol (MCP) directly into the user's browser via the `navigator.modelContext` API. It enables AI agents to perform complex, session-aware CRUD operations securely under the user's exact permissions.

The Browser as the AI Workspace

In-Browser Execution

Tools run purely client-side. The AI agent executes JavaScript within the browser context, piggybacking on the user's established authentication (JWT/Cookies) to call your APIs. No server-side secrets or API keys are ever exposed to the LLM.

Dynamic Tool Registration

React components can inject context-specific tools dynamically using the `useAeionTool` hook. When the user navigates to the "Invoices" page, the agent instantly gains the ability to execute `finance_get_client_invoices` with closure access to the current active client ID.

Enterprise Registry & Governance

Every tool call is intercepted by a registry that enforces strict permission checks (an agent can't run a `finance:write` action without that entitlement), applies rate limits, and logs the call to an immutable audit trail — so giving an AI agent real capabilities doesn't mean losing visibility into what it actually did.

Universal CRUD Generator

Every collection in the OS gets a matching set of five tools — list, read, create, update, delete — generated automatically with strictly typed schemas. Your team never hand-writes AI tooling boilerplate for a new collection; it's simply there the moment the collection exists.

Module-Specific Actions

Go beyond basic CRUD. The platform ships with 180+ semantic workflow tools out of the box. Agents can intelligently invoke tools like `commerce_bulk_publish_products` or `contact_merge` with explicit intent.

The Universal Backend Bridge

Some operations are too heavy to run in the browser. The bridge lets a browser-side agent discover and call the exact same server-side MCP tools the backend already exposes — one consistent tool surface, whether the action runs in the tab or on the server.

Secure AI Collaboration

Granular Permission Modeling

Every tool declares the exact entitlement it needs to run. An agent without `mymodule:write` on file gets blocked instantly with a clear `PERMISSION_DENIED` — the same rule a human without that role would hit. There's no separate, weaker rulebook for AI.

Automated AI Auditing

Every action an agent takes — which tool, which user, how long it took, whether it succeeded — is written to an immutable audit log automatically. When something goes wrong, "what did the AI actually do" is a query against `/api/webmcp/audit`, not a guess.

navigator.modelContext Polyfill
Client-Side JWT/Cookie Auth
Immutable AI Audit Log
Granular Scoped Permissions

Frequently Asked Questions

An MCP (Model Context Protocol) implementation that runs *inside the browser* — letting AI assistants integrated into your admin UI take actions on your page directly. Customer asks Claude (via your admin's AI panel) "update this order to shipped"; Claude calls the WebMCP tool which manipulates the UI directly without an HTTP round-trip.

Server MCP (the standard) routes the AI's tool calls through the server's MCP endpoint. WebMCP routes them through the browser-running AI tools — useful for UI manipulation, browser-side data validation, and capabilities only the browser can do (clipboard, native file picker, geolocation, getUserMedia for camera).

Three reasons. (1) Many UI workflows aren't exposed as APIs (form-fill, multi-step wizards, modal navigation). (2) Browser-only capabilities (clipboard, camera, native pickers) aren't reachable from the server. (3) Lower latency — the AI interacts with the page directly without HTTP overhead.

Yes — register your tool from any React component with `useAeionTool({ name, description, handler })`. The tool appears in the AI's available toolset while that component is mounted; unmounts when the user navigates away. Per-component scoping prevents tool-namespace bloat.

RBAC enforces what each user can do; AI inherits the calling user's role. WebMCP tools can't bypass auth checks — they invoke the same React component handlers the user would invoke by clicking. AI prompt-injection attacks that try to trick the AI into doing harmful actions get blocked at the same auth layer that blocks the user.

Chromium-based (Chrome, Edge, Brave, Opera, Vivaldi), Firefox, Safari. Older browsers without modern JS features fall back gracefully — WebMCP tools simply aren't registered if the runtime can't support them.

Ready to embed AI properly?