Core Concepts
In a nutshell: MyClaws is split into Admin (configuration & coordination) / Claw (execution on real machines) / Client (operator interface), layered on top of Agent / Skill / MCP / Trigger / Memory as composable building blocks.
One table for every role
| Term | What it is | Where it runs |
|---|---|---|
| Admin | Account, config, coordination hub, docs site | Your self-hosted server |
| Claw | Execution engine (Bash / files / toolchain) | Your real machines |
| Client | Operator interface (desktop / mobile) | Your computer or phone |
| Agent | Role / system prompt / tool whitelist | Loaded by Claw |
| Skill | Host-local capability package (with SKILL.md) | Claw host filesystem |
| MCP | External tool protocol | Launched by Claw |
| Trigger | Cron / Webhook automation | Admin β Claw |
| Memory | Cross-conversation memory (clue / dream) | Centralized in Admin |
The Admin / Claw / Client trio
- Admin: Fastify backend responsible for authentication, Agent/MCP config CRUD, download page, push proxy, and Memory API. You deploy it; all Claws and Clients connect to it.
- Claw: Another Fastify backend + WebSocket responsible for launching Agents, executing tools, reading/writing files, and running Bash. It runs on whatever machine you want the Agent to work on.
- Client: Desktop (Vue + Electron) and mobile (Expo + RN) frontends. Handles conversation UI, Claw management, and settings.
The three communicate via REST + WebSocket.
Agent is role plus tool whitelist
Each Agent is a combination of system prompt + enabled tool set + model preferences. A new Agent has all built-in tools enabled by default (write / read / bash / web_search, etc.); you can restrict them as needed. Public Agents are visible to all users; private Agents only to their creator. Built-in Agents (e.g., conversation title summarizer, Bernard) are maintained by admins; regular users cannot edit their system prompts.
Skill is a host-local capability package
A Skill is a directory containing a SKILL.md with frontmatter name / description.
Installed Skills live at ~/.myclaws/claw/data/skills/<name>/ on the Claw host filesystem.
For day-to-day use, ask the General Assistant β see Skill Overview.
MCP is the external-tool protocol
MCP (Model Context Protocol) lets Agents call external tools (databases, automation, custom services) through a unified protocol. You manage MCP Server configurations on the Claw side; mount the ones you need when starting a session.
Trigger is the scheduled / event entry
- Cron: starts a session on a cron schedule.
- Webhook: HTTP endpoint (with SDK) for GitHub / CI / custom system calls.
Memory is long-term cross-conversation context
user: user profilefeedback: behavioral corrections / confirmationsproject: project background factsreference: pointers to external information
Memory is stored centrally in Admin; Claw accesses it during sessions via the memory_recall / memory_write tools.
Common confusions
Claw is neither "your computer" nor "the AI backend"
Claw is an independent runtime that can be installed on any machine (including your own computer). It is not the model, nor a browser extension.
Skills are host-local β do not store them in Admin centrally
An earlier version stored Skills in Admin; that approach is deprecated. The new flow: install on the target Claw host via the General Assistant.