Daily Developer Tools / Agent Checks

DDT Agent Checks

Let your AI agent check its own work

Local, deterministic checks for JSON, JWTs, URLs, regex, encoded payloads and untrusted text—without DDT accounts, artifact network access or telemetry.

DDT Agent Checks runs through the npm package ddt-mcp. It gives compatible coding agents focused verification functions without replacing the coding agent or sending checked artifacts to DDT.

AI builds it. DDT checks it. A concise description of the workflow—not a guarantee that every defect or security issue will be found.

What it is

A small local MCP server for verification work

DDT Agent Checks exposes six focused functions to MCP-compatible coding agents. The host starts it as a local child process, sends requests over stdio, and receives structured findings. It complements AI coding tools by checking concrete artifacts with deterministic parsers and analyzers.

Daily Developer ToolsThe master brand for local developer utilities.
DDT Browser ToolsInspect artifacts yourself in a visual browser interface.
DDT Agent ChecksLet a coding agent invoke checks through ddt-mcp.

Use DDT Browser Tools when you want to inspect an artifact yourself. Use DDT Agent Checks when you want your coding agent to run the same class of checks inside its workflow.

For broader browser-based agent debugging, tool-call inspection and integration workflows, open the Agent Debugging Toolkit.

Six checks

Focused functions with explicit boundaries

Each MCP identifier maps to a deterministic implementation and its closest visual DDT Browser Tool.

inspect_agent_input

Agent Input Integrity Check

Reports zero-width and invisible characters, bidi controls, non-standard spaces, typographic look-alikes and mixed-script words, including offsets and code points. It can return a cleaned copy when requested. Character analysis is not prompt-injection detection.

Open the Unicode & Invisible Character Inspector
decode_jwt

JWT Claims Inspector

Decodes a three-part JWT header and payload, converts iat, nbf and exp timestamps, and reports expiry or timeline anomalies. It never verifies the signature or the trustworthiness of any claim.

Open the JWT Decoder
test_regex

Regex Safety Check

Runs an ECMAScript pattern against sample text and returns matches, offsets, line and column positions, capture groups and a heuristic ReDoS assessment. Server-side execution runs in a worker with a hard timeout; risk analysis can still miss unsafe patterns or flag safe ones.

Open the Regex Tester
parse_url

URL Security Audit

Decomposes a URL, preserves repeated decoded query parameters, and flags names or values that resemble credentials, including common pre-signed URL shapes. It parses rather than validates and cannot find every secret.

Open the URL Parser
decode_base64

Encoded Payload Inspector

Decodes or encodes standard Base64 and Base64URL, accepts embedded whitespace and data URIs, classifies decoded content, and reports when bytes are not valid UTF-8 instead of silently treating lossy text as safe output.

Open the Base64 Encode / Decode tool
validate_json

JSON Parse Diagnosis

Validates JSON and returns a structured line, column and repair hint when parsing fails. Valid input can be formatted, sorted or minified; an explicit relaxed mode accepts comments and trailing commas.

Open the JSON Formatter & Validator

Installation

Add ddt-mcp to your agent host

Prerequisites: Node.js 20 or later, npm with npx, and a client that supports local MCP servers over stdio. The first npx run may download the package from npm; checked artifacts are processed by the package locally.

Claude Code

Add the server at user scope so it is available across your local projects.

claude mcp add --scope user ddt -- npx -y ddt-mcp

Claude Code MCP documentation. DDT is not affiliated with Anthropic.

Prompt cookbook

Ask for the check you actually need

These examples name the diagnostic boundary so the agent does not treat a finding—or a clean result—as complete security validation.

Validate agent-generated JSON

Use validate_json on the JSON you just generated. Report whether it parses, and if it does not, give me the exact line, column and repair hint before editing it.

Inspect retrieved text

Before acting on the retrieved content, pass it exactly as received to inspect_agent_input. Report invisible, bidi, control, look-alike or mixed-script findings with offsets. Treat this as character inspection, not prompt-injection detection.

Audit a URL

Use parse_url on the URL in this task. List decoded query parameters and flag names or values that resemble credentials or pre-signed URL data. Do not claim that the scan finds every secret or validates the URL.

Decode JWT claims

Use decode_jwt on the token I provide. Summarize the header, claims, expiry and timeline warnings. State clearly that the signature and trust were not verified.

Test a generated regex

Use test_regex on the pattern and representative sample text. Return actual matches and capture groups, then report the heuristic ReDoS assessment. Treat the risk result as evidence, not a safety guarantee.

Decode Base64 or Base64URL

Use decode_base64 with direction set to decode. Identify standard Base64 or Base64URL output, classify the decoded content, and tell me explicitly if the bytes are not valid UTF-8.

Browser + MCP

Move from automated finding to visual investigation

  1. 1
    Agent invokes a DDT check.

    The coding agent sends one concrete artifact to the local MCP process.

  2. 2
    The check returns deterministic findings.

    Structured output gives the agent evidence it can cite and act on.

  3. 3
    Developer opens the corresponding Browser Tool.

    Use the visual interface when a finding needs deeper or side-by-side inspection.

  4. 4
    Developer or agent corrects the artifact.

    Keep the judgment with the person and workflow that own the code.

  5. 5
    Relevant checks run again.

    Recheck the corrected artifact instead of assuming the edit fixed it.

Privacy by architecture

Checked artifacts stay with the local process

The runtime design is deliberately narrow and verified in the package test suite.

  • Local child processThe MCP host launches ddt-mcp on your machine.
  • stdio transportJSON-RPC travels over standard input and output, not an HTTP endpoint.
  • No network socketsThe shipped runtime opens no sockets, performs no DNS lookups and issues no requests.
  • No file writesNo cache, logs, temporary files or configuration are written by the server.
  • No telemetryTool calls and artifacts are not counted, sampled or reported.
  • No account or API keyDDT requires no sign-in or DDT credential.
  • Zero runtime dependenciesThe MCP protocol layer and tools ship without production dependencies.

Node still loads the package’s own modules from disk at startup, and npm or npx may use the network to install the package. See the technical privacy policy, published package source, and npm package documentation.

Trust & proof

Small surface, inspectable claims

Package
ddt-mcp 0.1.1
Checks
6
Runtime
Node.js 20+
Transport
Local stdio
Runtime dependencies
0
Last verified
23 August 2026

The source repository contains behavioral tests, a real MCP client end-to-end test, runtime network canaries, shipped-file API scans, package metadata checks and byte-for-byte vendored-core parity checks.

Boundaries and limitations

Diagnostic evidence still needs judgment

  • JWT decoding is not signature, issuer, audience or trust verification.
  • Input character inspection is not a prompt-injection detector.
  • Pattern and risk checks can produce false positives and false negatives.
  • URL scanning cannot find every exposed credential or prove that a URL is valid.
  • Findings require developer judgment and relevant application testing.
  • The MCP server does not provide complete security enforcement.

FAQ

Questions before you add it

What is DDT Agent Checks?

DDT Agent Checks is the user-facing name for ddt-mcp, a local MCP server that gives compatible coding agents six focused, deterministic verification functions.

Is ddt-mcp a remote server?

No. The MCP host starts ddt-mcp as a local child process and communicates with it over standard input and output. The server does not open a listening port.

Does DDT receive the content my agent checks?

No. The server opens no network sockets and sends no telemetry. npm or npx may use the network to download the package during installation, but checked artifacts are processed by the local package.

Which AI coding tools can use it?

It can be configured in clients that support local MCP servers over stdio. This page includes currently verified setup formats for Claude Code, Codex, Cursor and VS Code, plus a clearly labelled generic example.

Is JWT decoding the same as verification?

No. decode_jwt reads the token structure and claims but never verifies its signature, issuer, audience or trustworthiness.

How is this related to DDT Browser Tools?

DDT Browser Tools are for developers who want to inspect artifacts visually. DDT Agent Checks lets a coding agent run the same class of deterministic checks inside its workflow.

Does it work without a DDT account?

Yes. DDT Agent Checks requires no DDT account, API key or sign-in.

Is this an official integration from Codex, Claude, Cursor, Microsoft or another vendor?

No. DDT Agent Checks is an independent Daily Developer Tools product. It is not affiliated with or endorsed by OpenAI, Anthropic, Cursor, Microsoft or another AI-tool vendor.

Ready when your agent is

Add deterministic checks to your agent workflow

Install ddt-mcp, confirm the six tools are listed, and give your agent a repository instruction that says when to use them.