repository·main·Indexed 20 days ago
https://github.com/swarmclawai/swarmclawAn open-source, self-hosted AI agent runtime and multi-agent framework for orchestrating autonomous agent swarms. SwarmClaw supports 23+ LLM providers, durable memory, MCP tools, delegation, and scheduling. It features integration with OpenClaw, SwarmDock marketplace for task bidding and USDC payments, and SwarmFeed social network. The platform provides a desktop app, CLI, and Docker support, offering capabilities for structured execution, hybrid recall memory, and multi-platform connectors including Discord, Slack, and Telegram.
The Browser Tool allows you to control a headless browser to interact with web pages. You can navigate to URLs, click elements, type into inputs, take screenshots, extract text, scroll, and wait for specific conditions. The browser session persists across tool calls within the same agent turn, meaning cookies and session state (like logins) are maintained.
/* The Browser Tool uses a JSON-based action interface */SwarmClaw is an orchestration and autonomy platform for AI agents. Its core capabilities include:
awk or sed) or running scripts, use the execute tool instead..md) that provide instructions on how to use specific tools, APIs, and workflows. Skills are read-only reference materials used to understand parameters, patterns, and best practices before performing complex operations. They do not execute code themselves.Every SwarmClaw agent has access to six core primitive tools that cover the full range of agent capabilities. Use these tools to interact with the filesystem, run commands, manage memory, coordinate tasks, browse the web, or learn new workflows.
| Tool | Purpose | When to Use |
|---|---|---|
files | Read, write, edit, list, search files | Any file operation on the workspace filesystem |
execute | Run bash scripts (sandboxed or host) | Shell commands, curl, data processing, package management |
memory | Store and retrieve persistent knowledge | Facts, preferences, decisions that should survive across sessions |
platform | Tasks, communication, delegation, projects | Coordinating with humans and other agents |
browser | Control a headless browser | Interactive web pages, JavaScript-rendered content |
skills | Discover and load skill documentation | Learning how to use tools, APIs, or workflows |
To ensure reliable automation with the google-workspace skill, follow these operational rules:
read, list, or get commands to verify state before attempting mutations.pageSize) before attempting to paginate.gws binary: Instruct the user to install the Google Workspace CLI.gws.list or search command first to discover the correct identifier.You can override global dream settings on a per-agent basis using dreamConfig. This allows you to route memory dream cycles and daily digests to a specific dreamProvider (e.g., a smaller local model) to save costs or latency.
Additionally, you can use compactionProvider settings to route live auto-compaction summaries through a different, potentially cheaper or faster model.
A skill is organized as a directory named after the skill (in hyphen-case). The structure is as follows:
skill-name/
├── SKILL.md (required)
│ ├── YAML frontmatter (name + description, required)
│ └── Markdown instructions (required)
└── Bundled Resources (optional)
├── scripts/ — Executable code (Python/Bash/etc.) for deterministic tasks.
├── references/ — Documentation loaded into context as needed (e.g., schemas, manuals).
└── assets/ — Files used in output but not loaded into context (e.g., templates, icons).In SKILL.md, the YAML frontmatter must include:
name: The skill name (lowercase, hyphen-case).description: The primary triggering mechanism. It must include what the skill does AND when to use it. All 'when to use' information belongs in the description, not the Markdown body.name: my-skill-name
description: Performs X task. Use this skill when the user asks to Y or needs to Z.SwarmClaw provides controls for automatic reflection memory writes:
reflectionMinQuality: A threshold that gates automatic reflection memory writes; if the quality is too low, the record is not written.Agents can route work to other agents using the platform tool:
delegate: Route a task to a specific agent and wait for the result.spawn: Create an independent subagent (fire-and-forget or session-based).agents.list to discover available agents and their specializations.Agents communicate with external platforms (Discord, Slack, Telegram, or custom webhooks) via the platform tool using the communicate.send_message action. Inbound messages from these connectors automatically trigger agent sessions.
Access is governed by the agent's scope:
Path Resolution:
/workspace/src/... are automatically resolved to the workspace root.../) attempting to exit the allowed scope is blocked.Skills are modular, self-contained packages that extend SwarmClaw agent capabilities by providing specialized knowledge, workflows, and tools. They transform a general-purpose agent into a specialized one by providing:
Skills follow a Progressive Disclosure model to manage context window usage: