AGENT SKILL

pclaw-features

Channel-neutral reference for PersonalClaw's capabilities — sessions, agents, models, dashboard, scheduling, goal loops, tasks, subagents, file/media handling, and config — so you can explain what PersonalClaw can do and route the user to the right surface.

License 许可
MIT
最近更新
2026-11-15
Stars
★ 9
Forks
6

01 前言 Front Matter

name
pclaw-features
always
false
triggers
help, commands, what can you do, getting started, onboard, capabilities, features, dashboard, setup, how do i
description
Channel-neutral reference for PersonalClaw's capabilities — sessions, agents, models, dashboard, scheduling, goal loops, tasks, subagents, file/media handling, and config — so you can explain what PersonalClaw can do and route the user to the right surface.

03 SKILL.md 正文

PersonalClaw Features

PersonalClaw is a personal AI workspace: persistent sessions, configurable agents and models, autonomous goal loops, a dashboard web UI, and a set of native tools (files, knowledge, memory, subagents, scheduling, artifacts, notifications). It runs as a long-lived gateway and can be reached from multiple surfaces — the dashboard, and connected channels (e.g. a Slack channel, if one is configured). This skill describes capabilities channel-neutrally: prefer the dashboard or the native tools, and only mention a specific channel's command syntax when the user is actually on that channel.

Core capabilities

Capability What it is
Sessions Persistent chat sessions with warm history. Resume an earlier session or start fresh; sessions survive restarts.
Agents Named agent profiles (system prompt, model binding, tools, hooks, workflows). Switch the active agent or override it for one session.
Models A bound model per session/agent. auto lets PersonalClaw choose; or pin a specific native/remote model.
Goal Loops The unified autonomous goal engine — classify a target, plan it, then loop one self-directed cycle per turn while a deterministic supervisor decides done-ness (see the loop-worker skill).
Subagents Spin up background workers for parallel/isolated work via subagent_run (see the delegation skill).
Tasks Tracked units of work with status and dependency links, via the /api/tasks CRUD and the native task provider (see the task-and-project skill).
Scheduling Recurring/deferred jobs via the schedule_* tools (add/list/get/update/pause/resume/remove/trigger).
Knowledge A searchable knowledge pool: knowledge_search / knowledge_get / knowledge_create (see the knowledge-grounding skill).
Memory Durable lessons, preferences, and facts: memory_remember / memory_list / memory_forget (see the memory-discipline skill).
Artifacts Named, versioned generated content (widgets, HTML, docs) via artifact_* (see the artifacts skill).
Notifications Reach the user out-of-band via notify (and notify_attachment for files).
Hooks Event-driven handoffs registered with hook_register.
Workflows Reusable standard operating procedures: workflow_list / workflow_get / workflow_run / workflow_create.

Dashboard

The dashboard is the primary web UI — chat, sessions, agents, models, memory, knowledge, scheduling, goal loops, artifacts, files, and system monitoring all live there. Point the user to the relevant dashboard page when they ask "where do I…": e.g. memory and lessons under the memory page, scheduled jobs under the scheduling page, saved artifacts at /artifacts/<slug>.

Access:

  • Local / loopback — open the dashboard URL directly; loopback access needs no token. Over SSH, forward the port (ssh -L <port>:localhost:<port> <host>) and open http://localhost:<port>.
  • Remote — if a public dashboard URL is configured, remote access uses a short-lived presigned link (request one from whatever surface you're on).

Configuration

Configuration lives in ~/.personalclaw/config.json (override the home dir with PERSONALCLAW_HOME). Common fields:

  • agent.approval_mode — "interactive" (the default: every call that needs approval asks, a chat on its card and an agent no chat started in the Inbox), "trust_reads" (a chat also runs a read-only shell command unasked) or "auto" (an agent no chat started approves every call it makes; chats still ask).
  • agent.model — default model ("auto" or a specific model id).
  • dashboard.url — hostname/port/bind for remote dashboard access. Omit for localhost-only.
{
  "agent": { "approval_mode": "interactive", "model": "auto" },
  "dashboard": { "url": "http://my-host.example.com:8080" }
}
Environment variable Purpose Default
PERSONALCLAW_HOME Config/data directory ~/.personalclaw
PERSONALCLAW_PORT Dashboard port (dev mode) 10000
PERSONALCLAW_PROJECT_DIR Agent config/skills directory Auto-detected

File & media handling

PersonalClaw processes attachments the user sends on any channel that supports them:

Attachment type Behavior
Voice memos (audio/*) Transcribed locally (speech-to-text), text injected into the conversation.
Images (png, jpeg, gif, webp, bmp) Sent to the model as vision input.
Text/code files (text/*) Content read and injected inline (truncated for very large files).
Other files Metadata noted; not auto-downloaded.

Speech-to-text is on by default where a transcription backend is available.

Channel surfaces

PersonalClaw can be driven from connected channels in addition to the dashboard. Each channel may expose its own command syntax — describe it only when the user is on that channel. For example, if using a Slack channel: users typically request a dashboard link, toggle auto-approval, switch agents, manage tracked channels, or force-halt a running turn via that channel's commands; the channel also defines per-channel activation modes (respond always / on mention / observe / off). Treat any such channel-specific command surface as an implementation detail of that channel, not a core PersonalClaw concept — the underlying capabilities (sessions, agents, scheduling, stop/cancel) are the same regardless of surface.

04 同仓库更多 Skills