ucsandman/DashClaw

Remote approvals, policy checks, and execution evidence for unattended AI agents.

What it solves

DashClaw provides a safety layer for AI agents running unattended (such as overnight coding sessions or CI agents) to prevent them from executing destructive or expensive actions. It solves the problem of giving agents unconditional authority by intercepting tool calls and requiring human approval for high-risk operations, ensuring that a single bad run cannot delete a database or force-push to a main branch.

How it works

DashClaw acts as a fail-closed approval layer between an agent's decision to call a tool and the actual execution. It uses hooks and gateways to freeze high-risk actions, routing them to an Approvals inbox where a human can approve or deny the request from a CLI, PWA, Telegram, or Discord.

Depending on the integration, enforcement is either "mechanical" (the tool is physically stopped by a hook before it runs) or "cooperative" (the agent's code must honor the guard's verdict). The system records every decision in a signed ledger for auditability and uses policy thresholds to determine when to interrupt the user based on the agent's stated confidence.

Who it’s for

Solo developers or small teams who run long, unattended coding-agent sessions against real repositories and infrastructure and cannot monitor every tool call in real-time.

Highlights

  • Fail-closed interception: Stops destructive calls (like rm -rf or DROP TABLE) before they execute at supported enforcement seams.
  • Remote approval: One-click resolution of pending actions via multiple platforms including mobile and chat apps.
  • Plain English explanations: Translates complex commands into simple sentences explaining what the action does and flagging irreversible changes.
  • Verifiable evidence: Maintains a signed decision ledger and Ed25519 receipts for later review.
  • Prompt-injection scanning: Includes default scanning for system-override patterns to block or warn against malicious inputs.
  • Multi-runtime support: Integrates with Claude Code, Codex, Hermes, OpenClaw, and provides Node/Python SDKs.

Related

  • Project
    cisco-ai-defense/defenseclaw

    DefenseClaw is a security governance framework for AI coding agents that provides runtime traffic inspection, capability scanning, and audit logging to help organizations manage agent risks.

  • Project
    riba2534/happyclaw

    HappyClaw is a self‑hosted, multi‑user platform that runs Anthropic’s Claude Code agents as persistent services. It provides a web UI, Docker‑isolated execution, eight chat‑app integrations, workspace‑scoped file/memory handling, scheduling, usage tracking and fine‑grained RBAC, letting teams safely automate code‑related tasks with Claude agents on‑premises.

  • Project
    emperorclaw/emperorclaw

    EmperorClaw is an operational control center for companies running teams of AI agents, providing task tracking, shared knowledge, and auditability to ensure agent work is organized and reliable.

  • Project
    johnson7788/MultiUserClaw

    MultiUserClaw is a Docker‑based SaaS framework that lets you run isolated AI assistants (Hermes agents) for many users. It includes a React front‑end, a FastAPI gateway for auth, container management and LLM proxying, per‑user Hermes containers, PostgreSQL storage, and features like knowledge bases, skill store, cron jobs, multi‑channel bots, and multi‑model support.

  • Project
    SAIL-Research-Lab/cheetahclaws

    A Python-native agent harness for AI coding assistants that supports any LLM provider or local model, offering a hackable alternative to closed-source agent frameworks.