Home About

Composable Claws

For hosting a personal agent, or claw, there is the seminal OpenClaw and its monolithic cousins, such as Hermes and IronClaw. At the opposite end of the architectural spectrum are the avant-garde, build-from-nothing projects such as NanoClaw and SeedBot.

Both approaches have advantages, but I wanted the best of both: well-tested, well-engineered components with the flexibility to choose or build the features I need.

Composable claws are built from modular components connected by established protocols. A client connects to a harness using ACP, while the harness connects to tools via MCP servers and to CLI tools directly. You choose the components that make up your stack:

  1. Model
  2. Harness (ACP server, MCP client)
  3. Instructions
    1. AGENTS.md
    2. Skills
  4. Tools
    1. CLIs
    2. MCP servers
  5. Frontend (ACP client)
  6. (Optional) Sandboxing
  7. (Optional) Network isolation
  8. (Optional) Credential proxy
Composable claw architecture: a frontend connects to a harness over ACP; the harness connects to a model, MCP tools through MCP servers, and CLI tools directly.

The harness can be wrapped in a container for sandboxing and network isolation.

Here is my stack:

  1. Model: GPT-5.6 (Sol and Luna), Qwen3.8-27B
  2. Harness: Pi, with pi-acp, pi-mcp-adapter
  3. Instructions: AGENTS.md
  4. Tools:
    1. CLIs: gog, leda, matrix-commander
    2. MCPs: mcp-cron, chrome-devtools-mcp, mnemosyne
  5. Clients: matrix-acp-bridge, mcp-cron
  6. Network isolation: squid
  7. Sandboxing: Docker

To run multiple profiles, copy the same stack and update the instructions and tools for each one.

Think it is too much to configure? We have a tool for that!

Example Multi-profile Docker Compose Setup