Vibe Coding
Back to FastClaw

FastClaw / Ability level

Tools and Sandbox

FastClaw has built-in tools such as files, commands, networks, memories, messages, subagents, Cron, preferences, project previews, etc., and uses sandbox policy to control execution boundaries.

Tools and Sandbox key concepts infographic
Tools and Sandbox key concepts infographic

Built-in tools

Agent.NewAgent will register memory_search, web_fetch, load_skill, exec, message, delegate_task, MCP tools, etc.; bash_output, kill_shell, apply_patch, file tools, image_gen, tts, cron, timezone, preference, goal and other tools can also be seen in the source code.

tool setRepresent toolPurpose
Fileread_file、write_file、list_dir、edit_file、apply_patchRead and write workspace and project files
commandexec、bash_output、kill_shellRun shells, trace long commands, terminate processes
networkweb_fetch、web_searchRead web pages and search data
Agentload_skill、install_skill、delegate_taskLoad skills, install skills, and dispatch subtasks
runtimestart_app_preview、app_preview_logsStart application preview and read dev server logs

sandbox strategy

policy/defaults.go provides three types of policies: permissive, standard, and restricted. Restricted denies network and exec, restricts system directory reading and writing, and sets CPU, memory, and timeout upper limits; standard denies system writing and shadow reading, but the network is more relaxed.

  • permissive: Allows all operations, suitable for trusted environments or development experiments.
  • standard: The default security is better, and system writes and sensitive reads are prohibited.
  • restricted: Suitable for public chat, low-trust users or scenarios where plug-ins are uncertain.

Execution backend

sandbox.Executor abstraction includes exec/read/write/list/backend/close; ExecutorPool can reuse backends by agent/project/session and supports Docker, E2B, Boxlite and other paths.

Coding Agent Key Points

When the Project Runtime is started, the file tool will be redirected to the app subdirectory to prevent the Agent from mixing the working directory with the runtime root directory.