la-caja-mcp

MCP system for access and debate over La Caja (contextual memory). Repo B: consumer of La Caja (repo A). This repo does not touch the memory core; it talks to it by API and exposes La Caja to agents.

Español: README.es.md

What's here

• src/la_caja_mcp/protocolo.py — agent–agent–human debate protocol. Claims, interference requests, deadline rounds in turns, escalation and human adjudication. Deterministic and replayable (event-sourcing, same discipline as La Caja). The state machine is a contract: in disputed the author can only respond or escalate; unresolved (deadlock) is the ONLY terminal the human can adjudicate.
• src/la_caja_mcp/mcp_server.py — MCP server: debate tools + La Caja memory tools (repo A, consumed by API), one tool set and two transports (local stdio / remote streamable HTTP).
• src/la_caja_mcp/install.py — one-command installer: detects the installed MCP agents and writes the correct config entry for each (opencode, Claude Code, Cursor, VS Code, Claude Desktop).
• tests/ — falsification of the state machine + integration with a real MCP client over stdio (debate and memory) + multi-agent SSE push (46 tests).
• experiments/uso_real.py — real use case: two LLM agents as MCP clients (shared memory, debate, replay, push).
• experiments/interrupcion_etapas.py — the protocol's basis: an agent mid-reasoning accepts being interrupted and yields within the round.
• worker/ — portable ASGI host (uvicorn + Dockerfile) for the remote MCP.

Validated with real LLMs (OpenRouter, gpt-4o-mini, over streamable HTTP)

Validation	Result
Real use case (2 MCP agents, shared memory)	OK — consensus, replay, push
Full interference request (deadline → escalate → human)	OK
Interruption mid-reasoning	OK — proponer → manifestar → interferir → responder → aceptar

The protocol basis works: the author exposes its reasoning in stages with manifestar (the interruption medium), checks the state between stages, the interferer requests interferir mid-way, and the interrupted yields by responding within the round (vence_en_turnos intact). Design detail and limits in the La Caja writeup (experiments/writeup.md, repo A).

Installation

Standard Python package (la-caja-mcp), like its dependency la-caja (repo A):

# Direct from the repository (works today)
pip install git+https://github.com/deviceargent/la-caja-mcp.git

# Published (PyPI)
pip install la-caja-mcp

# Local development
pip install .

Installs the la-caja-mcp executable and the dependencies (fastmcp, la-caja). The remote worker needs the hosting extra:

pip install "la-caja-mcp[host]"    # or  pip install .[host]  in the repo

One-command install (agents)

For the average user who wants "better memory" without hand-editing JSON:

la-caja-mcp install                # detects agents and registers everywhere
la-caja-mcp install --agent opencode
la-caja-mcp install --scope global # user-wide config instead of project
la-caja-mcp install --name caja    # entry name (default: caja)
la-caja-mcp install --caja-db <path>   # persistent memory for the server
la-caja-mcp install --list         # only list detected agents

It detects the installed agents and writes the config entry each one reads: opencode (opencode.json), Claude Code (.mcp.json / ~/.claude.json), Cursor (.cursor/mcp.json), VS Code (.vscode/mcp.json), Claude Desktop (claude_desktop_config.json, including the MSIX/UWP case, where the app redirects %APPDATA% into the package). The server command is resolved portably with the running Python (python -m la_caja_mcp.mcp_server --transport stdio). Restart your agent to pick the config up.

MCP server

One tool set, two transports:

# local (stdio): launched by the agent as a subprocess
la-caja-mcp --transport stdio

# remote (streamable HTTP): the standard for MCP over the network
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 (needs pip install la-caja; repo A): procesar_consulta, declarar_relacion, consultar, contexto_primado, historial (dormant trace, inert layer), stats.

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

Live discussion (push)

Over streamable HTTP the server also exposes an SSE endpoint:

GET /caja/push?sesion_id=<id>

Emits an estado event (snapshot: ultimo_seq + estado) on connect and then a sesion_actualizada per successful mover(). An agent subscribes with a separate HTTP connection (any stack) and uses ultimos_eventos to fetch the detail. Needs a shared server: over stdio each agent has its own process, there liveliness comes from polling with ultimos_eventos.

MCP transports (architecture decision)

• Local = stdio (the agent launches the server as a subprocess).
• Remote = streamable HTTP (the standard for remote MCP, + OAuth).
• worker/ is an example of a deployable host (VPS/Docker); there is no public hosted MCP.

Test

$env:PYTHONPATH="src"; python -m pytest tests -q
python experiments/uso_real.py                    # needs OPENAI_API_KEY
python experiments/interrupcion_etapas.py         # needs OPENAI_API_KEY

License

MIT — see LICENSE.