MCP 服务器

PLwC

Governed MCP gateway for Claude Desktop providing workspace, document, sandbox, profile, reflection, and audit controls through a single visible server.

README

PLwC — Personality Layer with Conscience

PLwC is a governed MCP gateway for Claude Desktop and compatible local MCP clients. It exposes one visible MCP server, routes all capabilities through policy and governance checks, and provides workspace, document, sandbox, profile, reflection and audit-oriented controls.

Status

Current status: Open Beta, based on v0.2.0-rc18.dev9.
This repository is the privacy-filtered public Open Beta snapshot. Private development history, local evidence, real profiles and workspace data are intentionally not included.
v0.2.0-rc18.dev9 combines PR-005 privacy-filtered packaging with RC18-ALIAS-001: reflection marker/trust inputs are English-first while preserving PBA2 canonical storage.
Readiness audit result: RC_READY_WITH_NOTES.
This is not a final public release.
The latest package-, Desktop- and Odysseus-smoked open beta package is v0.2.0-rc18.dev9 with verdict PASS.
The packaged MCPB is not signed.

PLwC is not production-certified. It is local infrastructure under active development.

What PLwC does

Exposes exactly one visible MCP server: plwc-gateway.
Provides exactly eight public facade tools (see below).
Controls workspace operations (read / write / search / move / rename / exact replace / batch read).
Creates and reads DOCX, XLSX, PPTX and PDF artifacts.
Inspects, extracts, merges, splits and rotates PDFs.
Inspects, extracts and creates ZIP archives with bounded safety limits.
Reads workspace raster images and returns them as MCP image content blocks (PNG, JPEG, WEBP, first-frame GIF).
Runs sandboxed Python and Shell snippets through Docker, with no silent fallback to a host shell.
Manages profiles, reflection writes, memory and persona governance, and Governor plan / apply flows including reflection_condensation with plan_id apply.
Blocks protected profile and governance paths from direct workspace writes.
Emits structured local audit events for high-risk operations.

Public tools

PLwC v0.2 exposes exactly these eight public facade tools:

Tool	What it does
`plwc_status`	Reports gateway, sandbox, and profile runtime state; generates a ready-to-paste Claude Desktop config snippet.
`plwc_describe`	Returns a self-description of all supported tools, operations, plan types, and capability boundaries — the single source of truth for what PLwC can do in the current session.
`plwc_profile`	Loads, inspects, and activates profiles; compiles the configured profile into the active session context.
`plwc_reflection`	Writes governed reflection entries to `reflection.md` with semantic validation — rejects pure technical logs and requires a reusable user, collaboration, or system insight.
`plwc_governor`	Two-stage plan / apply flow for promoting insights to `memory.md` or `PERSONA.md`, condensing stale reflection candidates, and creating new profiles — all requiring explicit confirmation before any write.
`plwc_sandbox_run`	Executes Python or Shell snippets in an isolated Docker container with no network access and no silent fallback to a host shell.
`plwc_workspace_operation`	Reads, writes, lists, searches, copies, moves, and renames files within configured workspace roots; includes binary and base64 paths and an exact-replace mode. Delete is not public.
`plwc_document_operation`	Creates and manipulates DOCX, XLSX, PPTX, and PDF files; inspects, merges, splits, rotates, and extracts text from PDFs; handles ZIP archives; reads workspace images as MCP image content blocks.

The 19 individual public tool names from earlier scaffolds (plwc_compile_profile, plwc_write_workspace_file, plwc_governor_plan, plwc_write_reflection, ...) are no longer public in v0.2. Their behavior is reachable through the eight facade tools above via operation / scope / lang dispatch parameters.

Document capabilities

DOCX Creation V2 (layouts, styles, runs, lists, tables, images, page breaks).
XLSX Creation V2 (multi-sheet, formatting, formulas-as-written (not executed), freeze panes, boolean auto_filter, merges).
PPTX Creation V2 (five slide layouts, content elements, tables, images, slide sizes, plain-text speaker notes).
PDF Creation V2 (layout-PDF builder with A4/A5/landscape/custom page sizes, headings, paragraph runs, bullet/numbered lists, tables, images, explicit page breaks).
read_image (governed workspace image reads).
ZIP create / inspect / extract.
PDF inspect / extract / merge / split / rotate / extract_pdf_text.

Security model

One visible gateway server. No bypass MCP servers are exposed.
Protected profile and governance files (PERSONA.md, memory.md, reflection.md, governance/config.yaml, ...) cannot be edited through workspace operations.
Workspace operations are scoped to configured roots; parent traversal, absolute host paths, UNC paths and protected segments are rejected.
Docker-backed Python and Shell sandbox; no silent host-shell fallback.
Fail-closed behavior on missing engines, missing Docker image, or policy denial.
No external URL fetching for document assets (PDF V2, PPTX V2 and DOCX V2 image paths must be workspace-relative).

Known limitations

The following are intentionally not implemented in v0.2:

No OCR.
No PDF redaction.
No digital signing.
No form filling or form creation.
No PDF/A claim.
No LibreOffice or Pandoc conversion.
No macro execution (no .docm / .xlsm / .pptm generation).
No external URL fetching, no network access at runtime.
No JavaScript in PDFs.
No HTML/CSS rendering pipeline.
No final-release claim yet.

Installation

For step-by-step installation, see docs/QUICKSTART_CLAUDE_DESKTOP.md.

For full installation context, prerequisites and configuration options, including local GPT and Odysseus stdio setup plus hosted ChatGPT web/custom-app status, see docs/INSTALLATION.md.