Tools

Tool layers

Layer	Tools	When included
Built-in	Read, Write, Glob, Bash, Fetch	Always (when tools enabled)
MCP	Auto-discovered from mcporter	Full tool profile on Anthropic and Codex provider paths
Agent	code_with_agent, check_code_agent	When chatId + config present

For coding worker details, see Coding Agents. For configured agents and reusable Discord @alias profiles, see Agents.

Agent coding tools

code_with_agent runs an external coding CLI (not an in-process Read/Write/Bash tool call).

Exact command builders live in src/code-agents/utils.ts (buildCodeAgentArgs):

• Claude worker: claude -p --verbose --output-format stream-json --dangerously-skip-permissions ... <task>
• Codex worker: codex exec --full-auto --json --color never ... <task>

Discord interactive coding sessions currently use Claude only.

Related files:

• Tool schema: src/tools/definitions.ts (code_with_agent)
• Runtime: src/code-agents/index.ts
• Command construction: src/code-agents/utils.ts

Built-in tools

• Read — read a file (restricted to allowedPaths)
• Write — write a file (restricted to allowedPaths, uses file locking for concurrent writes)
• Glob — find files by pattern (restricted to allowedPaths)
• Bash — run shell commands (blocked list via isBashCommandSafe())
• Fetch — make HTTP requests and return the response. HTML is auto-converted to plain text. Use for APIs, web search (e.g. https://duckduckgo.com/html/?q=your+query), or fetching page content

Tool config

Tools are configured per-channel or per-cron-job:

"tools": {
  "enabled": true,
  "allowedPaths": ["${HOME}/.skimpyclaw"],
  "maxIterations": 20,
  "bashTimeout": 30000
}

maxIterations is kept for older configs as a soft finalization checkpoint. A positive value asks the model for a final answer every N tool-use rounds, but it no longer caps the loop or returns a max-iteration error. Automatic context compaction keeps long tool runs within the configured context threshold.

MCP tools (mcporter)

SkimpyClaw auto-discovers MCP tools at runtime via mcporter.

How it works:

1. On first tool request, SkimpyClaw reads ~/.mcporter/mcporter.json
2. Calls listServers() to find all configured MCP servers
3. For each server, discovers available tools with their schemas
4. Maps each tool to mcp__{server}__{tool} format (e.g. mcp__my-server__search)
5. Results are cached for the lifetime of the process

mcporter config (~/.mcporter/mcporter.json):

{
  "mcpServers": {
    "my-tools": {
      "command": "npx",
      "args": ["@example/mcp-server"]
    },
    "my-server": {
      "url": "http://localhost:3001/sse"
    }
  }
}

Two transport types:

• Stdio: command + args — mcporter spawns the process
• SSE: url — mcporter connects to an HTTP SSE endpoint

CLI commands:

skimpyclaw tools list                                         # list all tools (built-in + MCP)
skimpyclaw tools install my-server --command npx --args '["@some/mcp-server"]'
skimpyclaw tools install my-server --url http://localhost:3001/sse
skimpyclaw tools remove my-server

tools install and tools remove modify ~/.mcporter/mcporter.json. Restart SkimpyClaw after changes.

Exec Approval

The Bash tool has a built-in safety gate that pauses and asks for human approval before running dangerous commands (risk tiers 2-3). See the dedicated Exec Approval page for full details on risk tiers, configuration, and channel behavior.