Agent Toolbox MCP server

Read from the package source without running it. The installed server may list more.

• check_citation

  Validates US case-law citations (volume / reporter / page / year) against a bundled table of reporter abbreviations, flags malformed or implausible citations, and \u2014 when source text is supplied \u2014 checks quote fidelity to catch fabricated quotes. Deterministic and offline. Returns PASS/FLAG/BLOCK. Call before an agent presents a legal citation or quotation.

• check_infra_plan

  Static blast-radius / risk analysis of infrastructure-as-code against a bundled CIS/OPA-style ruleset. Accepts a Terraform plan JSON (terraform show -json), an AWS IAM policy JSON, or a Kubernetes manifest JSON. Flags public exposure, IAM wildcards, destroy/replace of stateful resources, privileged pods, and more. Deterministic and offline (no cloud credentials). Returns PASS/FLAG/BLOCK with findi

• check_tool_args

  Validates a proposed tool/function call's arguments against a caller-supplied schema and policy: types, required fields, enums, numeric ranges, string length/pattern, null-safety, unknown args, unit-coercion (dollars-vs-cents), and cross-field rules. Deterministic and offline. Returns PASS/FLAG/BLOCK with violations. Call before executing a tool call involving money, quantities, or destructive eff

• compute_deadline

  Computes a legal deadline by counting court days (skipping weekends and US federal holidays) or calendar days from a start date, forwards or backwards. Deterministic and offline. Returns the resolved date and what was skipped. Use for filing deadlines and statute-of-limitations math.

• count_tokens

  Approximate token count and cost estimate for a string or a chat-formatted messages array. Use before an LLM call to budget cost and check context-window fit. Supports gpt-4, gpt-3.5, claude, gemini, and generic.

• distill_context

  Compresses a conversation context to fit within a target token budget using TF-IDF importance scoring. Use this when your context window is getting large to reduce token costs while preserving the most important content. Keeps system prompts and the most recent messages.

• finance_order_riskwrite action

  Full pre-trade gate. Runs token risk, slippage, price validation, and position limits in parallel and returns a composite PASS/FLAG/BLOCK with the check that blocked it. One call replaces the individual finance checks.

• finance_position_check

  Deterministic position-limit and kill-switch gate \u2014 no external calls, pure arithmetic. Enforces max position size %, daily-loss limits, leverage caps, open-position count, and an asset allowlist. The final non-overridable gate before executing a trade.

• finance_price

  Cross-validates an asset price against two independent live sources (CoinGecko + DexScreener for crypto, yahoo-finance2 for stocks). Blocks stale or divergent data and flags a proposed price that deviates from consensus. Provide symbol or tokenAddress.

• finance_slippage

  Estimates price impact for a trade using DexScreener pool liquidity (constant-product AMM approximation). Prevents draining a thin pool. Flags implausible volume/liquidity ratios (wash trading).

• finance_symbol

  Resolves a ticker symbol or token to a confirmed identity and flags ambiguity. For crypto, prefer resolving by address \u2014 symbols collide (USDC has 200+ imposters on Solana). Returns matches ranked by liquidity plus a verdict.

• finance_token_risk

  Rug-pull scanner for Solana tokens: RugCheck.xyz score plus on-chain mint/freeze authority verification. Blocks tokens with active mint authority, unlocked LP, or a risk score above threshold.

• finance_units

  Validates that a raw on-chain token amount matches the intended human (UI) amount given the token's authoritative decimals. Catches the catastrophic decimal-scaling error (e.g. sending 1000x too many tokens). Call before building any transfer transaction.

• rx_check

  Medication safety gate: deterministic unit-confusion, overdose, and drug-drug interaction checks for a list of medications. Returns PASS/FLAG/BLOCK findings with a signed certificate. Informational only \u2014 not medical advice. Call before an agent finalizes any medication list, prescription, or dosing instruction.

• scan_command

  Scans a shell command for dangerous / destructive patterns before execution \u2014 rm -rf /, curl|sh remote-exec, dd/mkfs raw disk writes, fork bombs, chmod 777, privilege escalation, force-push to protected branches, kubectl/docker destroys, firewall/security disables, and data exfiltration. Quote- and substitution-aware (content inside quotes never triggers). Deterministic and offline \u2014 nev

• scan_injection

  Detects prompt-injection attacks (instruction overrides, role hijacking, jailbreaks, data exfiltration, encoding tricks) in user-supplied input. Call this before passing any untrusted input to an LLM. Returns a risk band (safe/suspicious/injection), a score, and matched patterns.

• scan_pii

  Detects and redacts PII/PHI/PCI (SSNs, credit cards, IBANs, UK NHS numbers, emails, phone numbers, IPs) in text before it is logged, sent to a third party, or persisted. Uses deterministic checksum validation (Luhn, ISO-7064, mod-11). Returns PASS/FLAG/BLOCK, a redacted copy of the text, and a signed certificate. Call this before any egress of user or record data.

• scan_secrets

  Detects hardcoded secrets and credentials (AWS/GitHub/OpenAI/Anthropic keys, private keys, DB connection strings, passwords) in code before it is committed or executed. Matches are redacted \u2014 the raw secret is never returned. Returns findings with type, severity, and line number.

• scan_sql

  Scans SQL for dangerous patterns before execution \u2014 DELETE/UPDATE without WHERE, DROP/TRUNCATE, tautologies (WHERE 1=1), stacked statements, UNION-based injection, and privilege changes. Comment- and string-literal-aware; no database connection. Returns PASS/FLAG/BLOCK with findings.

• scan_url

  Scans a URL / host for SSRF and egress-policy violations before an outbound request or browser navigation \u2014 cloud instance-metadata endpoints (169.254.169.254), private/loopback/link-local targets, decimal/octal/hex IP obfuscation, denied schemes (file:, gopher:), credentials-in-URL, punycode/homograph hosts, and allow/deny-list + port policy. Deterministic and offline by default; optional DN

• scan_vulnerabilities

  Checks package names against the OSV (Open Source Vulnerabilities) database and returns CVEs/GHSAs affecting them. Use after validating imports to catch known-vulnerable dependencies before installing.

• screen_sanctions

  Screens one or more party names (people, companies, vessels) against bundled OFAC sanctions lists (SDN + Consolidated) using deterministic exact, alias, and fuzzy matching. Returns PASS/FLAG/BLOCK with matched records and a signed certificate. Call before onboarding, paying, shipping to, or contracting with any counterparty.

• validate_identifier

  Validates structured identifiers (IBAN, ABA routing, SWIFT/BIC, credit card, EIN, EU VAT, VIN, NPI, US SSN, Ethereum/Solana address) via deterministic checksums and format rules. Auto-detects the type when not given. Card and SSN values are masked. Returns PASS/FLAG/BLOCK. Call before an agent stores, pays to, or transacts against an identifier.

• validate_imports

  Validates all imports/packages in AI-generated code against live registries (PyPI, npm, crates.io, Go). Returns lists of valid, hallucinated, and unknown packages. Use this before executing any AI-generated code to catch hallucinated package names ('slopsquatting').

• validate_schema

  Validates a JSON value against a caller-supplied JSON Schema (Draft-07 subset), deterministically and offline. Use to gate an LLM's or a tool's structured output before acting on it. Returns PASS/FLAG/BLOCK with per-error JSON paths.

• verify_outputwrite action

  Runs the hallucination firewall on an LLM output. Checks for hallucinated packages (code), invalid URLs, malformed citations (DOI/arXiv), and numeric contradictions. Returns PASS, FLAG, or BLOCK verdict with a tamper-evident certificate. Use this before accepting any LLM response in a critical pipeline.

• Code scan2 source files scanned25/25
• –Live reliabilityno gateway calls yet and no remote to proben/a
• –Tool poisoningtools not inspected (local package is not executed); not countedn/a
• Auth qualitylocal package, no credentials required12/15
• Maintenancelast push 57 days ago12/15
• Maintainer identityregistry namespace matches repository owner6/10

Safety score

85/100

higher than 73% of 4,460 AI and agents servers

Tools

26

the typical AI and agents server has 6

Tools that change things

8%

typical: 0%. Fewer write tools means less an agent can break

Downloads a week

33

more than 22% of 1,538 AI and agents servers with a package

Agent Toolbox: common questions

Is Agent Toolbox MCP server safe?

Yes, by our scan: it is graded A (85/100). Read the Agent Toolbox safety report

How do I install Agent Toolbox?

It runs on your machine. Copy the Claude Code, Cursor, VS Code or Claude Desktop config from the install section.

Does Agent Toolbox need an API key?

Not as far as the registry entry and our scan can tell: no credentials are declared or required.

Is Agent Toolbox maintained?

The last commit was 96 days ago (2026-07-31). The latest release is v0.1.4.

What can I use instead of Agent Toolbox?

Servers from other publishers that do the same job: Emilia Protocol MCP server, three.ws Provenance MCP server and Decionis MCP server.

How does Agent Toolbox compare with other AI and agents MCP servers?

safety score 85/100 (higher than 73% of 4,460 AI and agents servers); tools 26 (the typical AI and agents server has 6); tools that change things 8% (typical: 0%. Fewer write tools means less an agent can break). Full comparison.