Browserbase MCP server

Use Browserbase from Claude Code, Codex, Grok Build and MCP clients through PipMCP. One PipMCP API key connects this package to private execution.

Activate with PipMCP · Integration directory

Build something useful

Inspect existing sessions, approve creation only if needed, then request release of the returned session ID and inspect its final state. Actual browser driving requires a separate CDP client.

The package is an authenticated client: discovery, schemas, policy decisions and execution come from PipMCP. Provider implementation remains on the server.

Related MCP server: Azure DevOps MCP server

Quick start

Requires Node.js 22+. Clone this repository, then:

npm ci --ignore-scripts
npm test
npm start
# After configuring your provider in PipMCP and setting PIPMCP_API_KEY:
npm start -- --live
# Launch the stdio MCP bridge:
npm run server

Create a PipMCP API key with explicit connectors:read and connectors:call scopes. Set PIPMCP_API_KEY in the calling process environment. Provider credentials are configured in PipMCP. Multiple accounts require explicit PIPMCP_INTEGRATION_ID and PIPMCP_CONNECTION_ID from discovery. Existing keys do not gain scopes automatically.

Built for coding agents

Install this repository's skill into your project:

node lib/cli.mjs agents install --client codex --scope project --project /path/to/project
# Use claude or grok for those clients.

Connect AI agents to Browserbase using one PipMCP API key and actual hosted tool discovery. Native AGENTS.md, CLAUDE.md and project skills support the current coding session. No additional AI model key is required. capabilities.json describes the package; llms.txt indexes its instructions.

Inspect actual tools before calling

import {PipMCPGatewayClient} from './lib/index.mjs';
const client = new PipMCPGatewayClient(); // PIPMCP_API_KEY
const inventory = await client.getTools("browserbase");
console.log(inventory.data.tools);
// Select a real tool and validate your arguments against its returned schema.
// client.callTool(provider, actualToolName, reviewedArguments, {idempotencyKey});

PipMCP enforces workspace, role, connection, policy/approval and usage controls. Preserve the original idempotency key when retrieving/resuming a receipt. Pending and confirmation_required are not completed provider results; uncertain writes are held for reconciliation. The bridge never auto-approves a hosted action.

Version 1 migration

This release uses the authenticated PipMCP gateway for real discovery/execution. Upgrade your MCP launch command and configure PIPMCP_API_KEY before reconnecting. Historical local runtimes remain available under their original MIT license in Git history.

Evidence and support

Transport, package and agent fixtures are tested. Authentication/provider access must be checked in your own configured workspace; synthetic fixtures do not establish live provider execution. Do not include credentials or customer records in public issues. MIT covers this public client.

Provider setup reference

Official provider documentation. Credentials are entered in PipMCP; this package does not accept or store provider secrets.