AgentScope packages Claw and Builder as runnable Java tools for local and team agent development

AgentScope has been packaged into two runnable Java applications that let developers exercise the Java Harness runtime and workspace conventions locally and, when needed, scale to team deployments. The examples — Claw, a personal — assistant-style single — user app, and Builder, a collaborative multi — tenant platform — are provided in the agentscope repository and can be built and launched as standalone JARs to demonstrate the Harness in practice. This matters because teams can iterate locally and then migrate the same agent logic into a multi — user environment without rewriting business code.

The quick — start walkthrough supplies concrete build — and-run steps for both examples. For Claw the guide describes building the example with Maven and then running the produced JAR; for Builder it shows exporting an API key (export DASHSCOPE_API_KEY=sk-xxx), building with Maven and launching the builder JAR. Both services start as web applications you can run locally to validate the Harness runtime and workspace layout before any production deployment.

Claw targets a single user on a single machine and exposes a Web UI together with five instant‑messaging channels. When launched it opens at http://localhost:8080 and uses a default workspace at ~/.agentscope; developers can extend it by editing ~/.agentscope/agentscope.json to add DingTalk, WeCom or Feishu channels. Claw runs with no process isolation (it executes under the user identity) and is intended as the fast local iteration environment for an individual developer.

Builder is aimed at teams and multi — user deployments: it provides a React single — page application plus a JWT REST API, listens on port 8080, and ships with default logins admin/admin, bob/bob and alice/alice. The Builder README documents production topics such as database migration, sandbox image setup and configuring a distributed backend, reflecting its role as the path from local experimentation to production — ready, team-scale operations.

The guide spells out concrete technical differences so teams can choose a path. Isolation and storage differ: Claw offers no isolation and uses LocalFilesystemWithShell, while Builder supplies (userId, agentId) namespaces and an optional Docker sandbox and uses a CompositeFilesystem. Sharing and permissions differ too — Claw has no sharing model, whereas Builder supports three — tiered permissions (run / edit / fork) and can scale horizontally by switching to a BaseStore backend instead of single — process, single — node operation.

A recommended workflow ties the two together: a developer tunes an agent locally in Claw until satisfied, commits the workspace directory (including AGENTS.md, skills/, subagents/) into version control as a template, and an administrator then imports or pushes that workspace into Builder for team use. The authors emphasize that moving from Claw to Builder does not require changing Agent business logic; only the storage medium and the isolation layer differ.

Underpinning both examples is the previously described Java Harness runtime, which includes HarnessAgent, workspace conventions, pluggable file systems and Hook pipelines. The examples are presented primarily as runnable teaching and evaluation tools for those Harness components and workspace patterns, enabling a path from single — developer iteration to multi — tenant deployment.