Salvaguardas portátiles para clientes de codificación agénticos que hablan el Model Context Protocol (MCP).
El proyecto pretende reducir el radio de impacto de los flujos de trabajo de codificación rápidos y altamente autónomos, proporcionando a los clientes un punto de decisión de políticas compartido. El servidor MCP no ejecuta por sí mismo la acción propuesta.
Este repositorio es un andamiaje inicial. La herramienta guard_check actual demuestra un contrato estable de decisión de políticas; aún no es una adaptación completa de las políticas de opencode-workflow-guard.
Lo más importante es que conectar un servidor MCP no lo convierte en un interceptor de todas las herramientas nativas que un cliente de codificación puede ejecutar. La aplicación estricta de las políticas depende del soporte de integración del host. Consulta Compatibilidad y Plan.
Related MCP server: Laguarde
guard_check: evalúa una acción propuesta de tipo shell, file_write, git o network y devuelve allow, deny o ask con un ID de política legible por máquina.
guard_status: informa del modo de aplicación de políticas actual del servidor. Identifica explícitamente este andamiaje como un asesoramiento de políticas dependiente del host.
Requiere Node.js 20 o superior.
npm install
npm test
npm run typecheck
npm run buildEl transporte inicial es stdio porque tanto Claude Code como Codex admiten servidores MCP stdio locales. Se puede añadir Streamable HTTP sin cambiar la API de políticas.
La promesa pública es deliberadamente más limitada que «este MCP aísla a tu agente de codificación en un sandbox». Centraliza las políticas. Los adaptadores del cliente aplican esa política allí donde el cliente exponga un mecanismo de interceptación fiable; de lo contrario, el resultado sigue siendo orientativo y debe combinarse con el sandbox nativo del cliente y sus controles de aprobación.
El diseño de compatibilidad se contrastó con la documentación oficial y el código fuente actuales el 27 de agosto de 2026:
Anthropic Claude Code MCP: https://docs.claude.com/en/docs/claude-code/mcp
Anthropic Claude Code hooks: https://github.com/anthropics/claude-code/blob/main/plugins/plugin-dev/skills/hook-development/SKILL.md
OpenAI Codex: https://github.com/openai/codex
Implementación de la configuración MCP de Codex: https://github.com/openai/codex/blob/main/codex-rs/config/src/mcp_types.rs
MCP TypeScript SDK: https://github.com/modelcontextprotocol/typescript-sdk/blob/v1.29.0/docs/server.md
MIT
Evaluate a proposed coding-agent action. Advisory unless the host wires the result into enforcement.
| Name | Required | Description | Default |
|---|---|---|---|
| path | No | ||
| action | Yes | ||
| command | No |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description carries the full burden, and it does disclose an important non-obvious trait: the result is advisory unless the host wires it into enforcement. The wording 'Evaluate' also implies analysis rather than execution, giving the agent a reasonable safety picture, though side effects and return behavior are not explicitly addressed.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Two short, front-loaded sentences with no filler: the first states the purpose, the second states a critical caveat about enforcement. Every sentence earns its place.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a tool with no output schema, no annotations, and zero parameter documentation, the description is too thin. It leaves the return value undefined, does not clarify the semantics of path/command/action, and does not say what a typical verdict looks like. An agent would likely need additional probing to call it correctly.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 0%, and the description adds no meaning for `path`, `action`, or `command`. The enum in the schema hints at action types, but the description does not explain how the parameters relate to the evaluation, what values mean, or when each is relevant, so it fails to compensate for the schema's lack of descriptions.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description uses a specific verb ('Evaluate') and a clear object ('a proposed coding-agent action'), which makes the tool's basic function understandable. It does not explicitly contrast with the sibling guard_status, but 'Evaluate a proposed action' versus 'status' is a strong enough distinction for an agent to pick the right tool.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description implies the tool is meant to be used when there is a proposed coding-agent action to assess, but it does not explicitly say when to use guard_check vs guard_status or when not to use it. The advisory/enforcement caveat gives useful context about reliance, but there is no routing guidance or exclusion criteria.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Report guard capabilities and the enforcement boundary.
| Name | Required | Description | Default |
|---|---|---|---|
No parameters | |||
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
No annotations are provided, so the description bears the full burden. The verb 'Report' implies a read-only, informational operation, which is a useful behavioral signal, but the description does not explicitly state side-effect freedom, authentication needs, or any operational constraints.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is a single, front-loaded sentence with no filler. It communicates the core purpose in nine words, and every word earns its place.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a zero-parameter, no-output-schema tool, the description adequately states what the tool reports ('guard capabilities' and 'enforcement boundary'). It is minimal but sufficient for invocation, though some domain-specific meaning of 'enforcement boundary' is left implicit.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The tool has zero parameters, so the input schema is fully complete. Per the baseline for zero-parameter tools, the description need not add parameter-level detail.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description uses a specific verb ('Report') and identifies a clear resource ('guard capabilities and the enforcement boundary'). It is not a tautology and gives the agent a concrete sense of what the tool does, though it does not explicitly contrast with the sibling 'guard_check'.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides no guidance on when to use this tool versus its sibling 'guard_check'. There is no mention of preferred scenarios, exclusions, or alternative routing, leaving the agent to infer usage from the name alone.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
v0.1.0guard_checkguard_statusScored across 2 tools
guard_check and guard_status have clearly distinct purposes: one evaluates an action, the other reports capabilities and enforcement boundaries. There is no overlap or ambiguity between them.
Both tools follow a consistent guard_ prefix with lowercase underscore naming. The pattern is uniform and predictable.
With only 2 tools, the server feels thin even for its narrow purpose. The count is borderline, sitting at the low end of what might be considered a minimal but workable surface.
The two tools cover the core guard workflow (evaluate action, check status) with only minor gaps like listing specific rules or explaining decisions. Agents can likely work around these by using the provided tools.
Deterministic allow/require_approval/deny verdicts for agent actions, before they happen.
Fail-closed policy guardrails for AI agents running kubectl, terraform, helm, and argocd.
Git-native policy layer for AI agents: check_action verdicts against rules approved via PR.
Pre-execution governance for AI agents. Deterministic PASS/FAIL/REVIEW verdicts, replayable proof.
