AI Agent Infrastructure

The quality layer every
AI agent needs

Catch hallucinations. Block injections. Scan secrets. Guard trades. 14 tools. One platform. Pay per call in SOL.

Read the docs View on GitHub

API Online · <200ms avg

19.7%

of AI-generated packages are hallucinated

USENIX 2025

68%

of multi-agent pipelines contain hallucinations

IBM 2026

$440k

lost by one agent from a single decimal error

Lobstar 2026

14 tools. Two suites. One platform.

Composable, agent-callable APIs that verify and improve LLM outputs before they cause damage.

Core quality checks

Package Validator

19.7% of AI-generated imports don't exist. We check every one against PyPI, npm, crates.io, and Go — in under 200ms.

POST /v1/validate/imports

Hallucination Firewall

Returns PASS, FLAG, or BLOCK for any LLM output. Catches bad packages, dead URLs, malformed citations, and numeric contradictions. Every verdict is SHA-256 certified.

POST /v1/verify

Context Distiller

Trim bloated agent context windows to a token budget. Keep the system prompt, deduplicate, and slice from newest to oldest. Up to 80% compression.

POST /v1/distill

Security & analysis

Secret Scanner

Detects hardcoded API keys, passwords, private keys, and connection strings in AI-generated code before they reach git. 10 pattern detectors. Redacted output.

POST /v1/scan/secrets

Injection Detector

Identifies prompt injection attacks in user input — role hijacking, jailbreaks, data exfiltration attempts, and encoding tricks — before passing to an LLM.

POST /v1/scan/injection

Token Counter

BPE-approximate token counting for GPT-4, Claude, and Gemini with per-model cost estimates. Know your spend before you call the LLM.

POST /v1/tokens/count

Vulnerability Scanner

Checks every package in AI-generated code against the OSV (Open Source Vulnerabilities) database. Surfaces CVEs and GHSAs before dependencies are installed.

POST /v1/scan/vulnerabilities

Finance protection

Backed by real incidents: Lobstar $440k decimal error · Claude Code $1,446 unauthorized sweep

Units Sanity Check

Prevents the $440k Lobstar-class decimal error. Validates raw on-chain amounts against authoritative token decimals from DexScreener and Solana RPC.

POST /v1/finance/units

Cross-Source Price Validator

Fetches the same asset from two independent live sources (CoinGecko + DexScreener for crypto). Blocks if they diverge >2% or if data is stale.

POST /v1/finance/price

Symbol / Token Resolver

Resolves tickers and token addresses to confirmed identities. Flags ambiguous symbols — USDC has 200+ imposters on Solana alone.

POST /v1/finance/symbol

Rug Pull Scanner

One call to RugCheck.xyz + on-chain authority verification. Blocks tokens with active mint authority, no LP lock, or rug score above threshold.

POST /v1/finance/token/risk

Slippage / Liquidity Guard

Estimates price impact using DexScreener pool data. Prevents the thin-pool disaster where a $440k order realizes ~$40k due to pool drainage.

POST /v1/finance/slippage

Full Order Risk Scorer

Pre-trade gate that runs all finance checks in parallel. Returns a composite PASS/FLAG/BLOCK with per-check breakdown and blockedBy field.

POST /v1/finance/order/risk

Position Guardian

Deterministic kill-switch. No API calls — pure arithmetic. Enforces max position size, daily loss limits, leverage caps, and asset allowlists.

POST /v1/finance/position/check

Catch it before it runs

Send AI-generated code. Get a verdict in milliseconds.

request.sh

curl -X POST https://agent-toolbox.ai/v1/validate/imports \
  -H "Content-Type: application/json" \
  -d '{
    "language": "python",
    "code": "import numpy\nfrom superlogger import magic_log\nimport pandas"
  }'

response.json

BLOCK

{
  "language": "python",
  "valid": [
    { "name": "numpy", "status": "valid", "registry": "pypi" },
    { "name": "pandas", "status": "valid", "registry": "pypi" }
  ],
  "hallucinated": [
    { "name": "superlogger", "status": "hallucinated", "evidence": "Not found on PyPI" }
  ],
  "verdict": "BLOCK",
  "totalImports": 3,
  "hallucinationRate": 0.33,
  "latencyMs": 142
}

PASS FLAG BLOCK

Model Context Protocol

Works with Claude, Warp, Cursor — any MCP client

Drop agent-toolbox into your MCP config and your agent gets quality tools out of the box — validate, verify, and distill. No SDK, no glue code.

validate_imports verify_output distill_context

claude_desktop_config.json

{
  "mcpServers": {
    "agentoolbox": {
      "command": "node",
      "args": ["/path/to/agentoolbox/packages/mcp/dist/index.js"]
    }
  }
}

Start free. Pay per call in SOL.

No subscriptions. No invoices. Agents pay autonomously on-chain.

Free

$0 forever

✓ 10 calls per IP
✓ No auth needed
✓ All 14 endpoints

Try it now

Live

Pay-as-you-go

0.0001 SOL / call

/v1/validate/imports 0.0001 SOL
/v1/verify 0.0002 SOL
/v1/distill 0.0001 SOL

1 SOL = 10,000 calls • No subscription • Agents pay autonomously

How to pay →

Service Wallet — send SOL here to purchase credits

8qXedRydihKEETqU64UXtG2sYZaUhwR4HBFz4Suu27CV

View on Explorer →

Use your Solana transaction signature as the Bearer token on your first API call. Credits are verified on-chain instantly. Or query GET /v1/pricing for programmatic discovery.

The quality layer every AI agent needs