Skip to main content
Glama
deenesjakoruzh
lluvr

frame-check-mcp

by lluvr
OverviewSchemaSourceRelated ServersDiscussionsAdmin
Python
Local

Frame Check

Descubre lo que cualquier documento no te muestra.

Frame Check es una herramienta de análisis de encuadre estructural. Identifica qué perspectivas adopta un documento, cuáles omite y cómo posiciona al lector. Las afirmaciones numéricas se contrastan con fuentes autorizadas cuando existe cobertura.

Inicio rápido (servidor MCP)

El paquete PyPI frame-check-mcp es el servidor del Protocolo de Contexto de Modelos (MCP). Se ejecuta localmente y proporciona a cualquier cliente de IA compatible con MCP (Claude Desktop, Cursor, Cline, Continue.dev, etc.) un análisis de encuadre estructural determinista como herramienta.

pip install frame-check-mcp

Luego, apunta tu cliente MCP al punto de entrada instalado. Para Claude Desktop, añade lo siguiente a claude_desktop_config.json:

{
  "mcpServers": {
    "frame-check": {
      "command": "frame-check-mcp"
    }
  }
}

Reinicia el cliente. Después, en cualquier conversación: "¿Puedes hacer un frame-check de este documento?". Detalles completos de instalación y verificación en MCP_SERVER.md.

Related MCP server: Kluster.ai Verify MCP

Qué hace

Proporciona un documento y Frame Check devolverá:

  • Un perfil de encuadre estructural: cuál de las cinco perspectivas analíticas (causas, riesgos, partes interesadas, tendencias, incertidumbre) cubre el documento, cuáles omite y la densidad de cada una.

  • Voz y postura epistémica: cómo posiciona el documento al lector y qué porcentaje de las afirmaciones se atribuyen a fuentes.

  • Orientación temporal: si el documento basa sus conclusiones en datos históricos, en el estado actual o en proyecciones.

  • Coincidencias candidatas del Estándar de Vocabulario de Encuadre: patrones de encuadre nombrados cuyas señales basadas en reglas se activan en el texto, cada uno con pistas de identificación y ejemplos prácticos. Las coincidencias son a nivel de candidato; la precisión frente al etiquetado de múltiples fuentes es una cuestión de investigación activa.

  • Verificación de red de fuentes: afirmaciones numéricas verificadas frente a SEC EDGAR, FRED, Banco Mundial, REST Countries, Alpha Vantage y Wolfram Alpha donde dichos proveedores tienen cobertura.

  • Una narrativa de IA opcional que interpreta el encuadre a nivel de prosa. Etiquetada claramente para que los lectores no confundan la interpretación del modelo de lenguaje con la medición determinista.

Enfoque

La medición estructural es la base. Cada afirmación de encuadre que hace la herramienta se calcula a partir de comparadores de patrones deterministas y siempre devuelve el mismo resultado para la misma entrada. La interpretación asistida por IA está disponible como enriquecimiento cuando se configura una clave API, pero está etiquetada como tal y nunca se oculta detrás de la capa estructural.

La verificación está limitada. La herramienta solo verifica afirmaciones numéricas frente a proveedores con cobertura genuina para el tipo de afirmación, y muestra sus propios resultados de calibración (precisión, exhaustividad, F1 por proveedor) en lugar de afirmar veredictos sin pruebas.

La detección de patrones nombrados es una capa de fiabilidad separada del perfil estructural. El F1 del detector es 0.36 frente a etiquetadores expertos en una validación pre-registrada, por debajo del umbral útil de 0.4. El giro hacia la superficie de honestidad de constructo (marcadores de sub-detección, advertencias de densidad, estados de confianza) en lugar de etiquetas seguras es la afirmación principal en la v0; un estudio de ayuda al lector (Track B, pre-registrado) prueba si esta superficie realmente ayuda al lector a ver el encuadre que de otro modo pasaría por alto.

Los límites honestos y las lecturas adversarias anticipadas están catalogados en ANTICIPATED_CRITIQUES.md.

Documentación

  • METHODOLOGY.md — documento de metodología completa (borrador v0.2)

  • MCP_SERVER.md — referencia del servidor MCP (herramientas, recursos, prompts)

  • MCP_CLIENT_CONFORMANCE_v1.md — 32/32 viajes de ida y vuelta de conformidad frente al wheel instalado

  • FRAME_DIVERGENCE_v1.md y FRAME_DIVERGENCE_v2.md — arquitectura de divergencia

  • data/frame_library/ — catálogo FVS de 20 entradas con ejemplos prácticos

  • ANTICIPATED_CRITIQUES.md — lecturas adversarias auto-enumeradas

  • VALIDATION_PROGRAM.md — planes de validación observacional + formal

Ejecución de pruebas

pip install -r requirements.txt
python3 run_tests.py

49 suites de pruebas, ~3 minutos. Incluye 63 pruebas de despachador adversario en test_mcp_adversarial.py además de la cobertura del motor V4.2 + clasificador.

Licencia

Apache-2.0 para el código; CC-BY-4.0 para la biblioteca FVS y los ejemplos prácticos (ver NOTICE para la enumeración por directorio).

Citación

Si Frame Check es útil en tu trabajo, consulta CITATION.cff para ver la forma citable. Frame Check es obra de Lovro Lucic; la autoría nombrada es el principal activo de credibilidad del proyecto según la disciplina de honestidad de constructo.

Contribución

Se requiere Sign-off-by-DCO según CONTRIBUTING.md. Gobernanza según GOVERNANCE.md (modelo BDFL con funciones de forzamiento nombradas para decisiones de promoción de canon). Compromiso de revisores externos según REVIEWERS.md.

Problemas

https://github.com/lluvr/frame-check-mcp/issues

Available Tools

2 tools
frame_checkA

Deterministic structural framing analysis. Returns analysis (measurements) + agent_guidance (composition discipline, scope-regime guidance, faithfulness rules) + provenance (versions, license, citation). When source_text is provided, also runs Layer 4 source_fidelity and Layer 11 grounding_decomposition with a Monte-Carlo-verified scope regime. When include_divergence=true (default at 0.8.0), the response carries a top-level divergence block sorted by signal_strength. The agent's role is to compose ONE insight grounded in the cited measurements (a reading the user could not see by reading their own document), not to walk the measurements one by one. The measurements are Frame Check's; the reading is the agent's. Cite measurements as Frame Check's; frame the reading as a reading ('the pattern reads as X'), never as a verdict ('the document is X'). Repeated calls with identical inputs return identical measurements; the agent's insight is a composition over them.

ParametersJSON Schema
NameRequiredDescriptionDefault
document_textYesThe document to analyse. English. 300-10,000 words. Markdown accepted. This is the text whose framing you want named.
source_textNoOptional. The source material the document is supposed to ground in (research report, filing, primary source). Unlocks Layer 4 source_fidelity (digit-level match) and Layer 11 grounding_decomposition (sentence-level G/F/P) with a scope regime telling you which layer to trust on number-dense sources. Without this, only structural framing analysis runs.
prefer_contract_versionNoOptional. Coverage contract version the client prefers. 1 (default): emit both v1 coverage and v2 coverage_v2 (Phase 1 compatibility window). 2: emit only coverage_v2 and omit v1 (for clients that have migrated and want to avoid payload duplication). See MCP_CONTRACT_V2_PROPOSAL.md and MCP_SERVER.md 'Contract versions' section. When Phase 3 activates, v1 will stop emitting regardless of this parameter.
include_divergenceNoOptional. Frame divergence output per FRAME_DIVERGENCE_CONTRACT_v1 Part 2. When true (default at 0.8.0; was opt-in at 0.7.x), the response carries a top-level `divergence` block (absent_frames array sorted by signal_strength tier, FaithfulnessEnvelope with divergence_summary prose) and two `agent_guidance` additions (how_to_render_divergence, absence_is_not_prescription). MCP surface does not invoke any LLM for divergence; the caller's agent model completes the composition using the guidance + library resources. Set explicitly to false to receive the v0.7.x-shape response with no divergence block. Default: true.
domain_hintNoOptional. Hint about the document's domain used to filter absent frames for domain relevance. Only meaningful when include_divergence=true. If omitted, the envelope reports domain_inferred='unfiltered' and all absent frames are returned. Domain-metadata-based filtering is a future contract minor version; the current implementation echoes the hint to the envelope without field-level filtering (documented in envelope.limitations).
divergence_renderingNoOptional. How the caller wants absent-frame records decorated. Only affects AbsentFrameRecord decoration (teaching_question field present in 'teaching_questions' mode); all other modes return the same data. The caller's agent model renders per this preference with faithfulness guarantees from agent_guidance.how_to_render_divergence. Default: 'list'.
catalog_version_pinNoOptional. Pin the FVS catalog version used for absent-frame set difference. If omitted, the latest stable catalog version is used (currently library_v3 per commit 9abeb3d). Unsupported pins are coerced to library_v3 with a limitation note in envelope.limitations.
user_contextNoOptional. The user's situation, role, or decision context in plain prose (e.g., 'I'm a startup founder making a hire decision in healthcare AI', 'reviewing a research paper on language model alignment', 'drafting a Substack post on agent safety'). When provided, agent_guidance.how_to_render_divergence is extended to instruct the caller's model to filter divergence relevance for this context. The MCP does NOT echo the value back into the response (privacy posture); the caller's agent has it from the call args. Discipline: the context personalizes RELEVANCE FILTERING; never PRESCRIPTION. The absence_is_not_prescription guarantee extends to contextual surfacing. Default: omitted (no contextual filtering). Maximum length 2000 chars.
include_frame_opportunitiesNoOptional. Opt-in flag for LLM-augmented frame-opportunity composition (Item 12 of the substrate-side composition roadmap). When true, divergence.frame_opportunities carries up to 3 document-specific questions composed by the LLM from absent-frame teaching questions plus the document's content. Each opportunity carries model_provenance with model name, cost_usd, and is_deterministic=false. The deterministic substrate (clusters, patterns, absences) is unchanged. Cost is bounded at ~0.001 USD per invocation (3 Gemini Flash calls max). Falls back to empty opportunities list with available=false if GEMINI_API_KEY is not set or the google.genai library is unavailable. Default: false (deterministic substrate only).
user_goalNoOptional. The user's stated goal for invoking Frame Check. One of 'decide', 'brainstorm', 'persuade', 'learn', 'audit'. When provided, absent_frames carry a goal_relevance dict for frames load-bearing for the chosen goal, and the absent_frames sort promotes goal-relevant entries within their signal_strength tier (goal precedes genre in the within-tier ranking). 'audit' is the default-equivalent posture: no goal-specific override is applied; the existing catalog/coverage/genre ranking stands. When omitted, behavior matches 'audit'. Substrate-side composition Item 11.
compose_budgetNoOptional. Bound the substrate's output volume so an agent in a tight working-memory budget can request a compact reading without losing structural shape. 'minimal' = top-3 absent_frames, top-1 absence_cluster, top-1 frame_pattern. 'standard' = top-5 absent_frames, all clusters, all patterns. 'full' (default) = unfiltered. The envelope.tier_counts always reflects PRE-slice counts so the agent sees the truncation honestly; divergence.compose_budget_applied carries per-layer returned/total counts. Substrate-side composition L5 interface UX. Backwards-compatible: omit to preserve current behavior.

TDQS

A4.4/5.0
Behavior5/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

With no annotations, the description carries full burden and excels: it declares determinism, output structure, agent's role, and limitations (e.g., 'The measurements are Frame Check's; the reading is the agent's'). No contradictions.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness3/5

Is the description appropriately sized, front-loaded, and free of redundancy?

The description is front-loaded with purpose but becomes verbose with detailed instructions and repeated guidance. While well-structured, it could be trimmed by ~30% without losing meaning.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness4/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

Given the complexity (11 params, no output schema), the description covers response components and behavioral nuances well. Missing explicit field-level structure of divergence block, but agent_guidance hints at rendering instructions.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters4/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

Schema description coverage is 100%, so baseline is 3. The description adds value by explaining defaults, version behavior, and interaction between parameters (e.g., domain_hint only meaningful when include_divergence=true). This exceeds the schema alone.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description clearly states 'Deterministic structural framing analysis' and lists output components (measurements, agent_guidance, provenance). It explicitly distinguishes from sibling 'frame_compare' by focusing on single-document analysis.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines4/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

The description provides guidance on when to use source_text (for layers 4 & 11), when include_divergence is default, and agent role. However, it does not explicitly exclude use cases or contrast with frame_compare, so it lacks explicit when-not advice.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

frame_compareA

Deterministic structural comparison of two documents on the same subject. Returns analysis (per-document summaries plus the cross-document comparison: shared blind spots, unique coverage gaps, voice / temporal / epistemic deltas, and a structured framing-differences narrative with per-dimension reader implications) + agent_guidance (what comparison tells and does not tell you, how to cite without implying a ranking) + provenance. Repeated calls with identical inputs return identical results. No LLM is invoked.

ParametersJSON Schema
NameRequiredDescriptionDefault
document_a_textYesThe first document to compare. English. 300-10,000 words.
document_b_textYesThe second document to compare. English. 300-10,000 words. Should be on the same subject as document_a_text for the comparison to be meaningful.
document_a_labelNoOptional short label for document A (e.g. 'Industry view' or 'Gemini response'). Used in the comparison narrative. Defaults to 'Document A'.
document_b_labelNoOptional short label for document B. Defaults to 'Document B'.

TDQS

A4.2/5.0
Behavior4/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

With no annotations, the description carries the full burden. It discloses that the tool is deterministic, returns identical results for identical inputs, and does not invoke an LLM. It also explains what the output includes (analysis, agent_guidance, provenance), providing transparency beyond basic behavior.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness4/5

Is the description appropriately sized, front-loaded, and free of redundancy?

The description is well-structured, starting with the core purpose, then listing outputs, and finally additional traits (deterministic, no LLM). It is slightly long but each sentence adds value, and it front-loads the most important information.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness5/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

Given no output schema, the description thoroughly explains what the tool returns (analysis, agent_guidance, provenance). It covers parameter constraints (word count, language, same subject) and behavioral traits. The context is complete for an agent to understand when and how to use the tool.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters4/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

Schema coverage is 100%, so baseline is 3. The description adds meaningful context: documents must be English, 300-10000 words, and on the same subject. It also clarifies that labels are optional and used in the narrative, which goes beyond the schema descriptions.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description clearly states the tool does 'deterministic structural comparison of two documents on the same subject' and lists specific outputs (per-document summaries, cross-document comparison, etc.). It uses a specific verb ('compare') and resource ('two documents'), distinguishing it from the sibling tool 'frame_check'.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines3/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

The description implies when to use (documents on same subject, deterministic comparison) but does not explicitly state when not to use or compare with alternatives. It notes that repeated calls return identical results and that no LLM is invoked, but lacks direct guidance on choosing this over the sibling tool.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

  1. 2 tool updatesv0.1.0
    • First observedframe_check
    • First observedframe_compare
Deploy Server
One endpoint for every MCP toolRayrun runs every MCP server behind one URL and holds the credentials. Each call is decided per client, tool, and argument, then recorded.ray.runAd

TDQS

A4.3/5.0

Scored across 2 tools

Disambiguation5/5

The two tools have clearly distinct purposes: one analyzes a single document, the other compares two documents. There is no ambiguity.

Naming Consistency5/5

Both tools follow the consistent 'frame_<verb>' pattern (frame_check, frame_compare). The naming is predictable and uniform.

Tool Count3/5

With only two tools, the server feels minimal. While the tools are well-scoped, a server for structural analysis might benefit from additional tools for tasks like batch comparison or result export.

Completeness4/5

The server covers the two core operations (single analysis and pairwise comparison). Minor gaps exist, such as lacking a tool for multi-document analysis or aggregation, but the main use cases are addressed.

Maintenance

ActivityMaintained
ResponsivenessNo issues

Resources

Related MCP Connectors

Related MCP Servers

  • BigContext MCP

    Rixmerz
    A
    license
    B
    quality
    D
    maintenance
    Enables working with large documents of any size by intelligently segmenting them and using TF-IDF search to retrieve only relevant fragments, preventing context window saturation. Provides 31 domain-agnostic tools for document ingestion, semantic analysis, epistemological validation, and extraction verification across formats like PDF, EPUB, and HTML.
    31
    MIT

  • clearon-source-audit

    kwgoodwin
    A
    license
    B
    quality
    C
    maintenance
    Enables revision-bound source audits with exact article fingerprinting, claim-to-source mapping, quotation verification, and immutable JSON evidence reports for prepublication review.
    9
    MIT