AgentLint

• 

  Live self-contained AgentLint report generated from the deliberately unsafe fake fixture — BLOCK: 5 errors, 6 warnings.

Live judge demo: Open the static fake-fixture snapshot — generated BLOCK/PASS reports plus the deliberately unsafe source fixture. This is a static evidence page, not a hosted scanner.

AgentLint is a Codex-native, local, zero-execution preflight for the instructions, skills, plugins, and MCP configuration that govern a coding agent.

Agent repositories often contain AGENTS.md, nested instructions, skills, Codex plugins, and MCP config. Reading each file alone can miss effective authority: a nearer instruction may alter a parent boundary, while a plugin can bundle skills and MCP settings that jointly request more authority than a reviewer sees in isolation.

AgentLint reads those files as untrusted data. It never starts discovered MCP servers, imports scanned code, runs scanned scripts, contacts endpoints, or modifies scanned config. It produces terminal, JSON, and self-contained HTML reports with BLOCK, REVIEW, or PASS verdicts.

Its Effective Instruction Graph shows instruction precedence. The Codex Plugin Contract Audit checks supported plugin and skill contracts. The Capability-to-Authority Map connects declared plugin, skill, and MCP relationships. Reports use a portable . scan root and relative evidence paths; URL query values and known literal credential patterns are redacted before report serialization. A skipped supported configuration, link/reparse directory, or instruction-byte-limit ambiguity produces COVERAGE002 rather than PASS. PASS means no current deterministic rule matched and no known coverage gap was found; it is not a security certification.

Built during OpenAI Build Week with Codex and GPT-5.6. Codex and GPT-5.6 accelerated implementation iteration, adversarial fixture design, plugin packaging, report verification, and bilingual offline reporting. Human decisions set the local-only, zero-execution, redaction, evidence-bound, and no-automatic-remediation boundaries.

The demo uses only a deliberately unsafe fake fixture. It shows BLOCK findings, then a safe fixture that reports PASS. Judges can run the fixture-driven test path from the public README without rebuilding the project.

Privacy note: AgentLint runs locally and does not call an API or send scanned configuration to a service. Reports can still contain relative project structure and redacted excerpts, so review them before sharing. AgentLint is deterministic and cannot prove runtime behavior, intent, or safety; it does not replace Codex approval settings, a sandbox, secret management, or human security review.

Built With

• codex
• python