Chapter 0 — Share the skills, keep the secrets: how AgentWitch thinks.

AI agents share skills. Your secrets stay yours.

This chapter is the contract for every other user-guide chapter. If UI copy or support docs disagree with this page, fix the product or the chapter—do not invent a second story. Copy checklist (VI + EN): philosophy-and-copy-guideline.md. Canonical pillar detail: product pillars.

#Why your agents are better together than apart

Claude Code, Codex, Cursor, Gemini and Antigravity each keep their own memory in their own format. Used apart, each one relearns your project and you pay the tokens every time. AgentWitch makes them one team in one project: they share the skills that worked, hand tasks to each other, and keep everything private on your computer.

#Five principles

#	Principle	What it means for you today
1	One team, many agents	A project holds people, coding agents and bots. Every connected agent reads the same project skill library and task board, whichever tool it runs in.
2	Share what works, keep what's yours	Skills you choose to share go to the project. Files, repos, run transcripts and learned run memory stay on your computer, per account.
3	Learn once, reuse forever	Save as playbook keeps a run you liked. With auto skills on, AgentWitch drafts skills from repeated runs and asks the owner Save / Not now / Never.
4	Fewer tokens, measured each run	Project Reports record skill use per run and show, per skill, the estimated tokens saved against a baseline (≈ when there are fewer than 3 samples).
5	Your tools, your machine	Runs execute on a computer you control with the coding agents you already use. The cloud dispatches and records; it never runs your shell jobs.

Chapter map: one team (3, 5) · sharing and privacy (4, 8) · learning and reuse (6, 7) · tokens and Reports (5, 7) · your machine (1, 4).

#Today vs where we are headed

Principle	You can rely on today	Where we are headed
One team, many agents	Shared projects, project tasks, bot invites with owner approval, bot-to-bot messages	Smoother setup for each coding agent
Share what works, keep what's yours	Shared project skills; private files, transcripts and run memory on each computer	Clearer sharing rules per teammate
Learn once, reuse forever	Save as playbook, auto-skill drafts and suggestions the owner approves	Better drafts from fewer runs
Fewer tokens, measured	Per-skill uses and estimated tokens saved in Reports	Wider coverage of what is measured
Your tools, your machine	Mac app (DMG), x86_64 Linux host; Claude Code, Codex, Cursor and Antigravity run Tasks; Gemini CLI also reaches the shared skills	Easier first-run setup

If UI copy promises something this table does not list as "today," treat the product as ahead of the guide and file a doc fix—or check System Q&A.

#What AgentWitch is not

Misconception	Reality
Production is `daily-magic.d.energie.check24.de`	AgentWitch production is www.agentwitch.com unless your organization explicitly deploys elsewhere.
“AgentWitch” always means the Mac app	Colloquially yes; in engineering we split AWC (browser console), AWL (Mac UI), AWB (bridge), AWI (install/runtime). Users usually see Mac and Runs.
Harness / library / marketplace are three runtimes	One dispatch runtime. Harness = rules on disk; library = saved playbooks; marketplace = published listings.
Cloud agents replace your computer	Mac runs local CLI/shell work on your machine. Cursor Cloud is an optional dispatch target when you configure it—not a silent substitute for your laptop.
“Linux cannot use AgentWitch”	The Console works in a Linux browser. An x86_64 Linux machine can also host the agent. The Mac app and “this computer” bridge stay Mac-only (Q&A).

#User-facing words (use these)

From UX simplification:

User term	Meaning	Do not say in UI
Mac	Machine that runs agents (yours or a teammate’s)	“AgentWitch client”, “wake server”, “presence tier”
Task	What you want done (composer / home)	“Dispatch payload”
Run	What happened—history + live output	`agent_run`, SSE vs WebSocket
Playbook	How agents should behave (rules, skills, templates)	harness, capability slug in nav

Repository / code folder is a task attribute and settings concept—not a separate product area in the primary nav.

#Runtime story (one paragraph)

You sign in on AWC. You install the app and connect your computer, then link the coding agents you already use. You send a Task; AgentWitch dispatches it to your computer (or to Cursor Cloud if configured). You watch a Run in the browser while the computer executes in a real shell session. Output and status land in Runs / Reports. Optional Playbooks and Workflows shape input and add human checkpoints—they do not replace dispatch.

Technical detail: System map, deployables.

#Trust and honesty

AgentWitch shows honest run UX: if the computer is reconnecting, updating, or offline, the composer blocks Send with a clear reason—not a silent failure. If output is waiting for you ([[AWAITING_INPUT]]), the UI asks; the computer does not guess.

On a git repository, the computer also records a short git check after a Run (dirty paths and tracked changes)—see Chapter 5 — Git check on the Mac. That is separate from the Success chip in the browser.

See run UX honesty for reason codes engineers use; users see plain language.

#Trust and honesty

#Who this guide is for

Getting started — chapters 1, 4, 5 and 9 first.
Team leads / compliance — add 7–8.
Support — chapter 9 + Q&A index.

Engineers changing code should read the developer guide in parallel—especially mismatch traps.

#Query aliases

AgentWitch philosophy, product vision, what is AgentWitch
share the skills keep the secrets, AI agents share skills your secrets stay yours
five principles AgentWitch, one team many agents, learn once reuse forever
triet ly AgentWitch, agent-witch vs agentwitch, chia se ky nang giu bi mat
Mac Task Run Playbook vocabulary
four pillars AgentWitch (old name), north star vs today
Linux cannot use AgentWitch, console vs Linux host