Add ElevenLabs integration: voice agent + browser agent

[shubh24]

What

Adds a standalone Next.js example under examples/integrations/elevenlabs/ that integrates an ElevenLabs voice agent with a Browserbase-backed browser agent.

The two agents share a single run state:

• ElevenLabs is the voice shell — it listens, speaks, and exposes one client tool (control_demo) for issuing high-level browser instructions.
• A single controller owns all browser state. It uses the Claude Agent SDK as a step planner and executes browser actions through the Browserbase browse CLI against a shared persistent Browserbase session.
• The same session is rendered in a live iframe, so the user can watch the browser act while talking to the voice agent.

This lets you say things out loud ("open Browserbase, find the docs, read me the first heading") and have a real cloud browser carry it out, narrated back by voice.

Contents

• app/ — Next.js UI (Browserbase-branded) + API routes (session, control, stream)
• lib/demo-controller.ts — controller: Claude Agent SDK planner + browse CLI executor over a shared Browserbase session
• lib/demo-types.ts — shared types
• README.md — local setup, required env vars, and ElevenLabs agent configuration
• .env.example — ANTHROPIC_API_KEY, BROWSERBASE_API_KEY, BROWSERBASE_PROJECT_ID, NEXT_PUBLIC_ELEVENLABS_AGENT_ID

Also adds an elevenlabs/ entry to the monorepo README integration list.

Run locally

cd examples/integrations/elevenlabs
pnpm install
pnpm dev   # http://127.0.0.1:3001

Requires the Browserbase browse CLI on your PATH.

🤖 Generated with Claude Code

---

Note

Low Risk
Isolated example app and documentation only; no changes to published packages or core library behavior.

Overview
Adds a new ElevenLabs + Browserbase example under examples/integrations/elevenlabs/ and lists it in the root README.

The demo is a Next.js app where ElevenLabs handles voice and calls a single client tool (control_demo), while a shared controller owns browser state: it spins up a persistent Browserbase session (live iframe), plans steps with the Claude Agent SDK, and runs actions via the browse CLI. The UI polls GET /api/demo/session, streams snapshots over SSE (/api/demo/stream), and posts instructions to POST /api/demo/control.

The controller layer adds queuing, duplicate detection, optional interrupt, direct-navigation shortcuts, snapshot-based planning, and voice-oriented outcomes/guidance returned to the agent. Docs cover env vars (.env.example), local run on port 3001, and suggested ElevenLabs agent tool/system copy.

^{Reviewed by Cursor Bugbot for commit e2c3d89. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit e2c3d89. Configure here.}

[cursor]

Browse setup skipped after failure

High Severity

If ensureBrowseRemoteMode fails after a Browserbase session is created, browserbaseSessionId is already set. Later calls to ensureBrowserRuntime return immediately and never retry Browse CLI remote setup, so that demo can stay unable to run browser commands until the process is restarted.

 

^{Reviewed by Cursor Bugbot for commit e2c3d89. Configure here.}

[cursor]

Queued runs stall on error

Medium Severity

When a run finishes, the next queued instruction is started via executeInstruction in a fire-and-forget promise. If that call rejects, the handler sets error state but does not start further queue items, so remaining queued instructions never run while the snapshot still shows a non-zero queue.

 

^{Reviewed by Cursor Bugbot for commit e2c3d89. Configure here.}