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 openhttp://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
ledger-read
Read a ledger CSV
budget-coach
Weekly spend review
yt-transcript
Fetch the transcript of a YouTube video and summarise it with timestamps. Use when the user shares a youtube.com or youtu.be link and asks what it says.
incident-writeup
Turn my incident notes, Slack excerpts and graphs into a blameless postmortem in Cartwheel's format. Use after an on-call incident once the mitigation is in.
feedsmith-release
Cut a feedsmith release - bump the version, move the Unreleased changelog section, tag, and draft the GitHub release notes. Use when I say "release feedsmith" or "cut 0.x".
feedsmith-bench
Benchmark feedsmith fetch against a real OPML, cold and warm, and compare with main. Use before merging anything that touches fetch.py or store.py.