Skip to main content
Glama
Sign In
deenesjakoruzh
la-caja-mcp

la-caja-mcp

by deviceargent
OverviewSchemaSourceRelated ServersDiscussionsAdmin
Python
Hybrid

la-caja-mcp

MCP system for accessing and debating La Caja (contextual memory). Repo B: consumer of La Caja (repo A). This repo does not touch the memory core; it communicates with it via API and exposes La Caja to agents.

What's Here

  • src/la_caja_mcp/protocolo.py — agent-agent-human debate protocol. Claims, interference requests, rounds with deadlines, human adjudication. Deterministic and replayable (event-sourcing, same discipline as La Caja).

  • src/la_caja_mcp/mcp_server.py — MCP server: debate tools + memory tools for La Caja (repo A, consumed via API), a single set of tools and two transports (local stdio / remote streamable HTTP).

  • tests/ — mini-falsification of the state machine + integration with real MCP client via stdio (debate and memory).

  • demo_debate.py — demo of debate that reaches consensus.

  • smoke_http.py — smoke test of streamable HTTP transport.

  • worker/ — portable ASGI host (uvicorn + Dockerfile) for remote MCP.

Related MCP server: Arena MCP Server

MCP Server

One set of tools, two transports:

# local (stdio): lo lanza el agente como subproceso
la-caja-mcp --transport stdio

# remoto (streamable HTTP): unico standard para MCP sobre red
la-caja-mcp --transport streamable-http --host 127.0.0.1 --port 8000

Debate: crear_sesion, mover (JSON payload), estado, ultimos_eventos, reproducir_sesion. Memory (requires pip install la-caja; repo A): procesar_consulta, declarar_relacion, consultar, contexto_primado, stats.

Memory is persistent with --caja-db <path> (SQLite, event-sourcing from La Caja) or LA_CAJA_DB; without these, purely in memory.

MCP Transports (Architecture Decision)

  • Local = stdio (the agent launches the server as a subprocess).

  • Remote = streamable HTTP (the only standard for remote MCP, + OAuth).

  • worker/ is an example of a deployable host (VPS/Docker); no public hosted MCP.

Test

$env:PYTHONPATH="src"; python -m pytest tests -q

Available Tools

10 tools
consultarA

Confianza de la relacion entre dos terminos: 1.0 observada, 0.5^puentes inferida por cierre transitivo, 0.0 sin relacion.

Inputs: a (required), b (required). Returns: object

contexto_primadoC

Contexto asociativo de un termino para inyectar en un modelo: las relaciones observadas primero, luego el vecindario de activacion, acotado al presupuesto. Mecanismo separado de la navegacion.

Inputs: termino (required), presupuesto. Returns: object

crear_sesionB

Crea una sesion de debate sobre un claim. Devuelve sesion_id y el estado inicial. El primer movimiento es proponer(actor).

Inputs: claim (required), participantes (required), arbitro, contexto, +1 more. Returns: object

declarar_relacionC

Declara una relacion entre dos terminos (co-ocurrencia explicita) sin pasar por texto crudo. Util para scripting e integraciones.

Inputs: a (required), b (required). Returns: object

estadoB

Estado actual de la sesion mas el log completo de eventos (replayable).

Inputs: sesion_id (required). Returns: object

moverC

Aplica un movimiento a la sesion. Ver docstring del modulo para el payload JSON que espera cada tipo. Devuelve el evento y el estado.

Inputs: tipo (required), actor (required), sesion_id (required), payload. Returns: object

procesar_consultaB

Ingesta: procesa una consulta humana completa en La Caja (tokeniza, normaliza a conceptos canonicos, filtra y crea/reforza las burbujas del contexto). Devuelve los terminos procesados y los eventos de la piscina.

Inputs: texto (required). Returns: object

reproducir_sesionA

Reconstruye una sesion replayando un log de eventos (JSON list). Verifica determinismo: el estado final debe coincidir con el que produjeron los eventos originales.

Inputs: log (required), claim (required). Returns: object

statsC

Dimensiones de la memoria: terminos, nodos, aristas.

Inputs: no named parameters. Returns: object

ultimos_eventosA

Eventos con seq mayor a desde_seq. Primitiva de sondeo para discusion en vivo: un agente pregunta que paso desde su ultimo corte.

Inputs: sesion_id (required), desde_seq. Returns: object

Tool Schema Changelog

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

Loading tool schema changelog…

    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

    C2.9/5.0

    Scored across 10 tools

    Disambiguation4/5

    Each tool has a distinct purpose: session creation, moves, state retrieval, event polling, replay, query processing, relation declaration, relation query, context retrieval, and stats. There is some overlap between `estado` and `ultimos_eventos` (both return events), but the descriptions clearly differentiate full state+log vs. incremental events, so agents can distinguish them effectively.

    Naming Consistency2/5

    Tool names follow snake_case but mix verb+noun (crear_sesion, procesar_consulta), bare verbs (mover, consultar), bare nouns (estado, stats), and noun phrases (ultimos_eventos, contexto_primado). This inconsistency in syntactic role makes the naming pattern hard to predict for an agent.

    Tool Count4/5

    With 10 tools, the set covers both session management and term/relation operations without feeling bloated. The scope is well-bounded; each tool adds a distinct capability. Slightly on the higher side but still appropriate for the domain.

    Completeness2/5

    Core operations are present (create session, move, query state, process queries, declare/query relations), but there are notable gaps: no session listing or deletion, no term/relation listing, no update or delete for relations. These missing CRUD operations would force agents into workarounds or failures.

    Resources

    Related MCP Connectors

    Related MCP Servers

    • multi-agent-debate-mcp

      albinjal
      A
      license
      A
      quality
      D
      maintenance
      Facilitates structured multi-agent debates with arguments, rebuttals, and judgments across multiple rounds, enabling diverse AI personas to engage in formal debate and collaborative problem-solving.
      1
      20 npm
      17
      MIT

    • Arena MCP Server

      tim101010101
      A
      license
      Not graded
      quality
      D
      maintenance
      Enables running position-driven adversarial debates and code reviews between AI agents via MCP tools, supporting custom positions, multiple rounds, and local CLI models like Claude, Codex, and Gemini.
      282 npm
      1
      MIT