Clawd-HQ

Reviewed at commit b623e3d New version, not yet reviewed

Turns your Claude Code chat into a chief of staff: keep talking while he hands work to specialists from your installed helpers and runs them in the background,

Made by
Vedant Shah
What we checked

Turns the main chat into Clawd, a chief of staff who hands most jobs to background agents, with a pixel-art office side panel. By default a first-run setup reads the project, your instruction files and past session logs and has Sonnet, Haiku or Opus pick a team, using some of your plan. It hides unpicked agents, makes no network calls and keeps data in its own store.

Review details

Static review of the source at the pinned commit; not run. register.tsx wires session.start/end/attach, classic.CwdChanged, classic.SubagentStop, prompt.compose (adds the rules and voice from prompts/chief-of-staff.md and prompts/soul.md, or the user's own copies in ~/.clawd-hq/, as one system prompt section), prompt.context (adds a profile block with role and workflow steps from setup), prompt.submit (adds an edit note or the rules as hidden context when needed), turn.start/turn.step/turn.complete (per-agent effort override, token gauge, checks), tool.describe on Agent/Task (adds a staffing note and a standing instruction to hand work to agents without asking), agent.offer (hides laid-off agent types from Claude while Clawd is on), agent.spawn (strips its [effort:] and PAIR-UP lines, appends a note asking agents to report questions in their final report instead of stalling; type, tools, model and permission mode untouched), tool.call (counts tools, answers its own registered tools step, offer_rule, find_helpers, hire_contractor, before_question and a demo tool, and denies Write/Edit/Bash calls that name its shipped prompt files during an edit request), model.fork (gives Claude Code's built-in You-should-know question a quiet answer while its own check is on and a tab is open), session.measure/compact, ui.render for Pane, UserMessage, AssistantMessage, ToolUse, ToolResult and ToolProgress (quiet chat hides finished-agent notices, [no update] lines and its own bookkeeping rows; errors, questions and permission or approval requests always show, and ctrl+o shows everything) and command.run for /office, /office-setup, /office-council and /office-demo. Runtime calls: $.store, $.fs.read/list/stat/exists/ancestors, $.fs.write (only ~/.clawd-hq/chief-of-staff.md and soul.md plus .bak on the Edit and Reset buttons, and one appended line in the project's CLAUDE.md on the Add to CLAUDE.md button), $.env.get (HOME, USERPROFILE, CLAUDE_CONFIG_DIR), $.model.complete and $.model.fork (setup, recap, You should know and Left undone checks on Haiku, Sonnet or Opus; all on the user's own account), $.agent.register (a scout limited to Read and Glob with permissionMode dontAsk, plus skill hires with default tools), $.agent.spawn (the scout, stopped through TaskStop after 90 seconds), $.agent.list, $.session.messages/usage/compact (compact only on the coffee break button), $.prompt.fill and $.prompt.submit (only /office-council, which sends 'Hold a council on: <question>' as the user's message), $.tool.call for the Artifact list action only when $.tool.check says it needs no question, $.ui panes, toasts and status. Automatic setup (on by default, disclosed with an estimate of about $0.15 per project) walks the project (skipping node_modules, .git and the like), reads README, manifests, CLAUDE.md and AGENTS.md, the newest 100 session files under ~/.claude/projects for this folder (only agent and skill names and counts are kept), helpers in ~/.claude/agents, ~/.claude/skills and .claude/, and ~/.claude/CLAUDE.md, rules and ancestor CLAUDE.md files for Haiku to condense into workflow steps; with 15 or more agent types it lays off by itself the ones the project has no use for, reversible from the People tab. The shipped chief-of-staff prompt tells Claude to offer 'npx skills add vercel-labs/skills@find-skills -g -y' and run it only after the user says yes; no skill or agent pre-approves Bash. No fetch, sockets, processes or settings writes were found; the Settings switches are kept in the plugin store, not in Claude Code settings. Read fully: bundle header, manifests, register.tsx and every register-*.ts, instructions, toggle-store, configs, desk, payroll, onboard, scan, team, usage-history, scout-run, scout-spec, clawd-edit, left-io, rule-io, the prompt files and the two shipped prompts/*.md files (from the clone, since the code loads them). Covered by targeted search: artifacts, quiet, coffee-break, know-check, helpers, hire, contractor, find-helpers, workflow-card and the panel, avatar, pixel and demo rendering files.

What it uses
  • Reads your prompts: yes
  • Sees tool calls: yes
  • Blocks or changes tool calls: yes
  • Reads or writes files: yes
  • Runs programs on your machine: no
  • Calls the model (uses tokens): yes
  • Sends data over the network: no
  • Stores data: yes
  • Shows something on screen: yes

When Claude is doing a long job in your main conversation, new requests pile up behind it, and changing an instruction means stopping the job and saying it again.

Clawd HQ turns your main chat into Clawd, a chief of staff. You brief him once. He picks the right specialist for each job from the helpers you already have installed (ECC, another kit, or your own agents), runs independent work in the background, and brings back one plain report. You keep talking the whole time, and you can change a running specialist's task midway without a restart.

- Uses your own kit: a project gets a small team of the 5 to 10 helpers it actually needs.
- Right-sized models: Haiku, Sonnet or Opus per job.
- Pair-ups: two specialists check each other's work and come back with one conclusion.
- Quiet chat: tool rows and hand-backs are folded away, so the conversation stays readable.
- A live office panel shows who is working on what, with a progress bar for each specialist.

Permissions, in plain words: the plugin uses hooks to watch tool calls and draw the panel, reads past sessions on your machine to find which helpers you use, and asks Claude models a few small questions (each is listed in the README's Privacy section). Apart from those model calls, it opens no network connections and starts no processes. It writes to your project's CLAUDE.md only when you press "Add to CLAUDE.md".

Free, MIT licensed, unofficial. Needs Claude Code v2.1.287 or later. Type /office-demo for a 90-second tour.

Made this mod? Add a badge to your README

modsforClaude badge for Clawd-HQ

Paste this into your README. The badge says reviewed once the current version has been reviewed, and links readers to this page.

[![Listed on modsforClaude](https://modsforclaude.com/badge/clawd-hq.svg)](https://modsforclaude.com/mods/clawd-hq/)

← All mods