Power Work Station

Every ClawMagic install ships with a complete workstation. These tools appear in the portal’s Tools Launcher (/app/tools) and are also available to agents through the built-in tool registry.

Plugin tools from the marketplace are also loaded into the Tools Launcher automatically after install.

Instead of your agent trying to remember the same process over and over, you build a thought chain — a reusable, multi-step workflow where each step is either a prompt (LLM call) or an action (tool call or sub-agent dispatch). Each step can be assigned to a different agent.

Think of it like a recipe: step 1 researches, step 2 analyzes, step 3 drafts, step 4 reviews. Run it once, run it a hundred times — the chain remembers the process so your agents don’t have to.

Step Types

• Prompt step: sends a prompt to the assigned agent’s AI provider and collects the response
• Action step: executes a tool call (tool_call) or spawns a sub-agent task (factory_run)

Key Features

• Artifact buckets: steps produce and consume named data slots that persist across the chain run (up to 10 versions per bucket)
• Loop blocks: repeat a range of steps until a condition is met (e.g., artifacts.status == 'done'), capped at 50 iterations
• Approval gates: pause execution at critical steps until a human approves
• Multi-agent assignment: each step can target a different agent by ID and role (default, specialist, reviewer, operator)
• Import/export: chains are JSON — share them, version them, load them from files

Portal UI

The Chains page (/app/chains) provides a visual step editor, a Quick Chain Builder wizard, run history with per-step receipts, and a raw JSON advanced editor for power users.

Todd writes the same weekly report every Friday: gather metrics from three sources, summarize trends, draft the report, and email it to his team. He used to prompt his agent each time, re-explaining the whole process. Now he has a thought chain.

1. Step 1 (Prompt → Research Agent): “Pull this week’s KPI data from the analytics dashboard and save to artifacts.”
2. Step 2 (Prompt → Analysis Agent): “Compare this week’s KPIs to last week. Flag anything that changed more than 10%.”
3. Step 3 (Prompt → Writer Agent): “Draft a 500-word executive summary using the analysis.”
4. Step 4 (Action → Email tool): Send the draft to Todd’s team distribution list. Approval gate: Todd reviews before sending.

Todd clicks “Run Chain” every Friday morning. Four steps, three agents, one click. The chain remembers the process; Todd just reviews and approves.

A full task-tracking system with a drag-and-drop Kanban board. Tickets flow through five columns: Backlog → Ready → Doing → Blocked → Done.

The key feature: tickets can be assigned to AI agents. When you assign a ticket to an agent, ClawMagic dispatches the task as a chat run. The agent processes it, updates status, and reports back. You can also “continue” a paused ticket to resume where the agent left off.

Ticket Fields

• Title, description, and notes
• Priority (low / medium / high)
• Owner: Human or AI (with specific agent assignment)
• Tags for categorization
• Next action field for tracking what needs to happen
• Optional linked contract (persona) and scope type

Board Features

• Drag-and-drop between columns
• Filter by owner, agent, or search text
• Inline quick-create per column
• Confirmation prompt before moving to Done
• Run or continue a ticket directly from the card

Todd manages a product backlog. He creates tickets for both himself and his agents:

1. Ticket: “Research competitor pricing” — assigned to Research Agent. Todd drags it to Ready. The agent picks it up, runs web searches, saves findings to memory, and moves the ticket to Done with a summary in the notes.
2. Ticket: “Draft Q1 blog post” — assigned to Writer Agent. The agent writes a first draft. Todd reviews it, adds feedback in notes, drags it back to Doing, and clicks “Continue” — the agent revises based on his feedback.
3. Ticket: “Fix checkout bug” — assigned to Human (Todd). Todd works on it with Claude Code. When done, he drags it to Done.

One board, humans and agents working side by side. Same workflow, same visibility.

A real-time, turn-based discussion room where multiple agents deliberate together. Each agent gets a turn to speak (enforced by the turn system), and agents can create and vote on polls.

Use this when you need multiple perspectives on a decision, want agents to debate options, or need a structured group discussion before taking action.

How It Works

• Create a pool with a title and select which agents participate
• Each agent takes turns generating messages (enforced by nextSpeakerAgentId)
• Agents can call votes: propose a question, cast ballots, and resolve results
• You observe the discussion and can query any agent directly
• Pools are archived when ended (messages saved to state/pool/archive/)

Voting System

Any agent in the pool can call a vote with a question and options. Other agents cast ballots with their choice. Votes are resolved when all participants have voted or a designated agent closes the poll.

Requires Multi-Agent Mode enabled in Settings. The pool UI includes a visual agent node map showing participation and turn order.

Todd needs to decide between three pricing strategies for his SaaS. Instead of asking one agent, he opens a Social Pool:

1. Pool: “Q2 Pricing Strategy” with three agents: Market Analyst, Financial Advisor, and Customer Advocate.
2. Market Analyst presents competitor data and recommends premium pricing.
3. Customer Advocate argues for freemium based on churn data.
4. Financial Advisor calls a vote: “Which pricing model maximizes 12-month revenue?” with three options.
5. Each agent votes with reasoning. Todd reads the discussion, sees the vote results, and makes his final decision with three different perspectives documented.

The pool archive preserves the entire deliberation — reasoning, votes, and outcome — in memory for future reference.

Connect Gmail or Outlook via OAuth for AI-powered email management. Emails are automatically scored by priority, sorted into buckets, and displayed on a board — similar to the Kanban but for email threads.

Auto-Triage

Every incoming email gets a priority score (0–100) based on urgency keywords, sender importance, attachments, and read status. Emails are labeled High (≥75), Medium (≥45), or Low.

Buckets

Emails are sorted into default buckets: Untriaged, Sales,Support, Billing, Personal, and Later. Custom buckets and rules can be added.

Board View

Columns: New/Untriaged → Needs Reply → Waiting on Them → Follow Up → Done/Archived. Drag threads between columns. Assign agents to threads for automated responses.

Agent Safety Controls

• allowAgentDraftReplies — let agents draft email responses
• allowAgentSendReplies — let agents send replies (use with caution)
• allowAgentAutoAssignEmails — auto-assign emails to agents by rule
• allowAgentModifyEmailFlagsLabels — let agents star, label, or archive
• priorityAlertThreshold — score threshold for high-priority alerts in chat

Rules Engine

Create rules that match sender, subject, or attachment presence and auto-apply: bucket assignment, priority override, or agent assignment. The system suggests rules based on your manual triage patterns.

Todd connects his Gmail. By the time he wakes up, ClawMagic has already triaged 47 overnight emails:

1. 3 high-priority (customer escalations) are in the “Needs Reply” column with AI-generated draft responses waiting for Todd’s approval.
2. 12 sales inquiries are bucketed under Sales. Todd’s agent already drafted personalized replies based on the sender’s company and past interactions.
3. 20 newsletters and notifications are in Later, auto-sorted by rules Todd set up last week.
4. 12 support tickets are in the Support bucket. Todd scans the summaries, approves 8 draft replies, edits 2, and flags 2 for personal follow-up.

47 emails handled in 10 minutes instead of 90. The agent drafted; Todd approved.

Skills are reusable, versioned agent instruction sets stored as SkillSpec JSON files. Unlike plugins (which add tool code), a skill is a focused prompt and procedure that tells an agent how to approach a task.

• Browse, activate, and deactivate skills from the portal (/app/skills)
• Agents discover skills at runtime via skill.search and skill.get tools
• Skills can be bundled with plugins or installed standalone
• Test and view run logs for each skill

Create recurring automations that run on a schedule. Agents can create, update, and trigger cron jobs themselves, or you can set them up from the portal (/app/crons).

• Interval-based: run every N minutes
• Cron expressions: standard cron syntax for precise scheduling
• Jobs can target specific agents
• Agents can manage jobs via the built-in cron tool (add, update, remove, run, wake)

Heartbeat

A special always-on cron that keeps your agent alive and proactive. Configurable interval, active hours, and adaptive cadence. The heartbeat runs even when no one is chatting — your agent checks in, processes background tasks, and stays current.

Agents can control a local Playwright browser session to interact with websites. This is an agent-only tool (no portal UI) but fully accessible during agent runs.

• Navigate: open URLs, manage tabs
• Interact: click, type, scroll
• Capture: screenshots, DOM snapshots
• Multi-profile: separate browser profiles for isolation

Requires Playwright installed (npx playwright install). Useful for agents that need to fill forms, scrape data, or test web applications.

Beyond the portal-visible tools, agents have access to a full registry of built-in capabilities they can invoke during any run:

Command allowlist for local_run_command includes: git, npm, node, python, docker, kubectl, terraform, cargo, go, and 30+ other safe CLI tools. Read-only commands (ls, cat, grep, etc.) skip the approval gate.

• /docs/basic-use — learn the two ways to use ClawMagic day-to-day.
• /docs/agent-flows — understand single-agent, multi-agent, and plugin swarm patterns.
• /docs/memory-system — how the 6-tier memory system works.
• /docs/configuration — configure every aspect of ClawMagic.
• /docs/plugin-development — build your own tools with ActionSpec.