Overview

@inbrowser/agent

Agent runtime plus an agent-friendly CLI: AgentSession, AgentStrategy, ToolRegistry, ModelClient, and an `agent` binary.

Agent runtime + agent-friendly CLI in one package. Three consumer surfaces:

Library — import { createAgentSession, createToolRegistry, ... }. Pure TypeScript, no React, no DOM, no fetch. The playground UI consumes the same primitives.
CLI — the agent binary. NDJSON output by default, --json stdin, --dry-run, schema introspection, automatic session logs.
MCP server — agent serve exposes agent tools over stdio so external LLM hosts (Claude Desktop, others) can drive them as MCP tools.

The library is browser-safe; the CLI and MCP server are Node-only.

Install

TERMINAL

bun add @inbrowser/agent
# or
npm install @inbrowser/agent

The agent binary lands in ./node_modules/.bin/.

A 30-second example

Library

import {
  createAgentSession,
  createReactLoopStrategy,
  createToolRegistry,
  createDispatch,
  createMetricsCollector,
} from '@inbrowser/agent';

const session = createAgentSession({
  strategy: createReactLoopStrategy(),
  llm: yourModelClient,
  tools: createDispatch(createToolRegistry()),
  toolContext: () => ({ workspace, runtime, sandbox, lint, signal: new AbortController().signal }),
  metrics: createMetricsCollector(),
  history: [],
  id: 'sess-1',
  systemPromptBuilder: (workspace, runtime) => '…',
});

for await (const event of session.submit('hello')) {
  // SessionEvent stream — apply to your store / UI / log
}

CLI

TERMINAL

# Headless single session — text in TTY, NDJSON when piped.
agent run "build a chess rule set"

# JSON payload via stdin.
echo '{"prompt": "reset rules"}' | agent run --json -

# 10 isolated sessions in parallel.
agent fleet --size 10

MCP server

TERMINAL

# Expose agent tools over stdio for an external host.
agent serve --project my-project

CLI commands

Command	Purpose
`run`	Headless single session — text in a TTY, NDJSON when piped
`fleet`	Run N isolated sessions in parallel
`serve`	Expose agent tools over MCP stdio
`events`	Inspect a session’s event log
`undo`	Roll back the last mutating operation via the event log
`describe`	Print tool schemas / session metadata
`schema`	Emit machine-readable schema introspection
`migrate`	Migrate session logs / event-log format
`version`, `help`	Version string and usage

Subpath exports

Entry	Surface
`@inbrowser/agent`	Browser-safe library — `createAgentSession`, `createToolRegistry`, strategies, metrics, storage adapters, event utilities
`@inbrowser/agent/cli`	Node-only CLI internals — `main`, command handlers, arg parsing, the `CLI_SPEC`
`@inbrowser/agent/node`	Node-only event-log writer — `openEventLog`, `buildRollbackEvent`, id generation

Where to go next

Documentation is organised under docs/ following the Diataxis framework:

If you want to	Read
Follow a complete lesson	Tutorials
Accomplish a specific task	How-to guides
Look up a function, event, or CLI flag	Reference
Understand the design choices	Explanation

Starting points by role

Building an agent in code? Start with Drive a session from your code.
Running the CLI? Start with Run the agent CLI.
Exposing tools to an external host? Start with Serve agents over MCP.
Implementing a new LLM provider? See Implement a custom ModelClient.

Position in the stack

@inbrowser/agent is the agent runtime — independent of domain packages. Hosts compose it with their own AgentDefinitions; the bare CLI ships zero built-ins.

See Inference vs inverse architectures for the two distinct consumer modes.

Licence

Same as the parent workspace.