Skip to main content
Glama

Typed live-app actions for MCP-compatible AI agents, over WebSocket.

Built by Eigenwise


NEW: Join our community on Discord at discord.gg/J3W9b5AZJR — protocol questions, feedback, and SDK-contribution chat all welcome.

Tesseron is an accessibility layer for AI agents. You instrument your app once — the way you'd add ARIA to a web page — and any MCP-compatible agent can drive it through typed actions you define. Think of it as an API for agents, written by the people who built the app.

The premise: the agent doesn't need to click a button, it needs to do the thing the button does. Say an agent wants to add five items to a list. Through the UI that's click add, type, submit, click add again — five round-trips through a brittle, re-render-happy interface. With Tesseron it calls one action:

addTodos(['buy milk', 'finish report', 'call mom', 'book flight', 'water plants'])

Your handler runs against your real state. No clicking, no scraping, no brittleness.

Live applications (browser tabs, Electron/Tauri desktop apps, Node daemons, CLIs) declare actions with a Zod-style builder; agents (Claude Code, Claude Desktop, Cursor, Copilot, Codex, Cline, ...) call them as MCP tools. Your real handler runs in your real process, against your real state. No browser automation, no scraping, no Playwright.

It's a protocol, not just a TypeScript library — and not just for the web. The JS/TS SDKs are the reference implementation and already cover the browser, Node, and desktop (Electron/Tauri). But the wire protocol is CC BY 4.0 and language-agnostic: anything that can open a WebSocket and speak JSON-RPC 2.0 can host actions, so a Python daemon, a Rust desktop app, or a .NET line-of-business tool can speak it too. Web is simply where the first SDKs landed — not the boundary of what Tesseron is.

WebMCP is a useful browser-native path for public sites that want an assistant to fill forms. The Tesseron and WebMCP section explains the difference, including Tauri and Python processes, cross-app flows, and agents like Claude Code.

Why Tesseron

  • Typed actions, not scraped DOMs. Declare with Zod or any Standard Schema validator; the handler is a plain function against your real state.

  • Framework-agnostic. Same API for vanilla TS, React, Svelte, Vue, and Node. Pick your stack.

  • MCP-native. Every action, resource, and capability maps to a standard MCP primitive. Users pick their agent.

  • Click-to-connect. Six-character claim code handshake. No API keys, no OAuth dance, no per-client configuration.

  • First-class capabilities. ctx.confirm for yes/no, ctx.elicit for schema-validated prompts, ctx.sample for agent LLM calls, ctx.progress for streaming updates, subscribable resources for live reads.

  • Cross-client delivery. First-class install paths for Claude Code, Codex, OpenCode, and Pi — each one a single command or short config snippet that wires the MCP gateway in. No bundled binary; the gateway is npx -y @tesseron/mcp@<version> on demand.

Related MCP server: Touch Flow Base (MCP Counter POC)

Install

Tesseron has first-class install paths for four agent clients. Pick the one you use:

Claude Code

/plugin marketplace add eigenwise/tesseron
/plugin install tesseron@tesseron

Installs the tesseron Claude Code plugin. The MCP gateway and docs server are launched on demand via npx.

Codex CLI

codex plugin marketplace add eigenwise/tesseron

Codex consumes the same plugin manifest as Claude Code, so the gateway, docs server, and skills come along automatically.

Pi

Pi has no built-in MCP support, but the community-maintained pi-mcp-adapter is the canonical bridge. Install it once:

pi install npm:pi-mcp-adapter

Then add Tesseron to .mcp.json in your project root (or ~/.config/mcp/mcp.json for a global install):

{
  "mcpServers": {
    "tesseron": { "command": "npx", "args": ["-y", "@tesseron/mcp@2.10.4"] },
    "tesseron-docs": { "command": "npx", "args": ["-y", "@tesseron/docs-mcp@2.10.6"] }
  }
}

pi-mcp-adapter discovers and exposes the Tesseron tools to Pi automatically. Use "directTools": true per server entry to surface each Tesseron action as a top-level Pi tool instead of going through the mcp proxy.

To pick up the skill bundle as well, point Pi's settings skills array at a clone of plugin/skills/ — the same folder Claude Code / Codex use.

OpenCode

OpenCode reads MCP servers from opencode.json rather than a plugin manifest. Save this as .opencode/opencode.json in your project root (or ~/.config/opencode/opencode.json for global use):

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "tesseron": { "type": "local", "command": ["npx", "-y", "@tesseron/mcp@2.10.4"], "enabled": true },
    "tesseron-docs": { "type": "local", "command": ["npx", "-y", "@tesseron/docs-mcp@2.10.6"], "enabled": true }
  }
}

To also pick up the skill bundle, point OpenCode's skills.paths at a clone of plugin/skills/ — see plugin/README.md for the snippet.

Other MCP clients

Claude Desktop, Cursor, VS Code Copilot, Cline, and any other MCP-compatible client work too — the gateway is plain stdio MCP. See the one-time setup in examples/README.md.

Then in your app

Drop @tesseron/web, @tesseron/server, @tesseron/react, @tesseron/svelte, or @tesseron/vue into your project, declare actions, and let the agent drive your real UI:

import { tesseron } from '@tesseron/web';
import { z } from 'zod';

tesseron.app({ id: 'todo_app', name: 'Todo App' });

tesseron
  .action('addTodo')
  .input(z.object({ text: z.string().min(1) }))
  .handler(({ text }) => {
    state.todos.push({ id: newId(), text, done: false });
    render();
    return { ok: true };
  });

await tesseron.connect();

See TypeScript examples/ for working apps in vanilla TS, React, Svelte, Vue, Express, and plain Node.

Packages

Package

Purpose

@tesseron/core

Protocol types, action builder. Zero runtime deps beyond Standard Schema.

@tesseron/web

Browser SDK.

@tesseron/server

Node SDK.

@tesseron/react

React hooks adapter.

@tesseron/svelte

Svelte 5 adapter.

@tesseron/vue

Vue 3 adapter.

@tesseron/vite

Vite plugin: dev-server bridge for browser tabs to dial the gateway over the same origin as your app.

@tesseron/mcp

MCP gateway server (tesseron-mcp CLI; launched by each client's install path via npx).

@tesseron/docs-mcp

MCP server that serves the Tesseron docs (search_docs, read_doc, list_docs) for chapter-and-verse spec lookups inside agent sessions.

The Claude Code / Codex plugin lives at plugin/, exposed via the marketplace manifests at .claude-plugin/marketplace.json (Claude) and .agents/plugins/marketplace.json (Codex).

Client capability support

Tesseron's action context gives handlers four capabilities beyond plain tool invocation, each backed by an MCP primitive. Whether a given call actually fires depends on what the user's MCP client advertises:

SDK surface

MCP primitive

tool(...) (action invocation)

tools

resource(...) (live reads, subscriptions)

resources (+ resources.subscribe)

ctx.sample(...)

sampling

ctx.confirm(...) / ctx.elicit(...)

elicitation

ctx.progress(...)

notifications/progress (client must pass _meta.progressToken on tools/call)

For the authoritative, continuously-updated list of which client supports which primitive, see the official MCP client compatibility matrix — filter by Sampling or Elicitation to see how narrow the field still is. A few points worth knowing before you pick a capability:

  • Tools are universal — every MCP client can invoke your actions.

  • Sampling is the rarest. Claude Code, Claude Desktop, and Claude.ai do not expose it; today's support is concentrated in VS Code + GitHub Copilot, goose, and fast-agent.

  • Elicitation (MCP 2025-06) landed in Claude Code (2.1.76, March 2026), Cursor, Codex, VS Code Copilot, goose, and fast-agent, but not Claude Desktop, Claude.ai, ChatGPT, Windsurf, or Zed.

  • When a capability is missing, Tesseron raises a typed error (SamplingNotAvailableError, ElicitationNotAvailableError) or collapses to the safe default (ctx.confirm returns false), so handlers can branch explicitly rather than silently misbehaving.

Status

v1.0 shipped April 2026; the SDK is at v2.10 as of writing. The protocol is at 1.2.0 and intentionally kept small: bidirectional JSON-RPC 2.0 over WebSocket, dynamic MCP tool registration, click-to-connect handshake, streaming progress, cancellation, sampling, confirmation, schema-validated elicitation, subscribable resources, session resume.

The seven TypeScript SDK packages, @tesseron/{core,web,server,react,svelte,vue,vite}, share one version and should be installed together. The hub packages @tesseron/mcp, @tesseron/docs-mcp, and @tesseron/conformance release independently. See the compatibility contract for cross-language rules.

The JS/TS SDKs are the reference implementation; the protocol spec is CC BY 4.0 so anyone can write a compatible client or server in any language. The conformance fixtures are the executable half of that.

Language SDK repositories

Language

Source

Install

TypeScript

tesseron-typescript

npm scope @tesseron: core, web, server, react, svelte, vue, vite

Rust

tesseron-rust

crate tesseron: cargo add tesseron

Python

tesseron-python

PyPI tesseron: uv add tesseron

C++

tesseron-cpp

CMake FetchContent from that repo, link tesseron::tesseron

All four speak protocol 1.2.0 and pass the shared conformance suite apart from fixtures their transports or claim flow do not support. SDK code and examples live in the language repositories; docs, protocol fixtures, and issues stay here.

On the roadmap: a Streamable HTTP transport and more desktop-native integrations. The Rust SDK already includes a Tauri example.

Development

pnpm install --frozen-lockfile
pnpm gate
pnpm build
pnpm docs:build

Contributing

Bug reports, protocol refinements, new framework adapters, and improvements to the reference implementation are welcome.

Star history

License

Reference implementationBusiness Source License 1.1 (source-available). You may embed Tesseron in your own applications, use it internally, fork it, and redistribute it freely. You may not offer Tesseron or a substantial portion of it as a hosted or managed service to third parties. Each release auto-converts to Apache-2.0 four years after publication.

Protocol specificationCC BY 4.0. A compatible implementation in any language, for any purpose including commercial, is explicitly encouraged.

Contributions are welcome under the Developer Certificate of Origin — every commit must be Signed-off-by.


Related MCP Connectors

Related MCP Servers

  • F
    license
    Not graded
    quality
    D
    maintenance
    Enables deployment of autonomous AI agents with memory and tool execution capabilities through a WebSocket-based MCP protocol. Provides production-ready infrastructure with REST API access, persistent state management, and extensible function registry for building self-hosted AI systems.
    -
  • F
    license
    Not graded
    quality
    D
    maintenance
    An MCP server that enables AI agents to execute sandboxed JavaScript code to control external web applications in real-time. It utilizes WebSockets to synchronize script execution with frontend animations, allowing complex UI interactions to be handled in a single agent turn.
    -
  • F
    license
    Not graded
    quality
    B
    maintenance
    MCP server that exposes WebSocket session management as tools, allowing AI agents to open persistent connections, send typed frames, drain buffered messages, and monitor per-frame Shannon entropy to detect schema divergence.
    -
  • -
    license
    Not graded
    quality
    C
    maintenance
    Enables AI agents to discover and call business actions from browser-only legacy web applications as typed MCP tools, executing them through the original GUI via Playwright.
    -