@aldus-palace/mcp

You told your assistant something important. It remembered — in a place you cannot read, scoped to that one tool, and gone the moment you switch client.

npm install -g @aldus-palace/mcp

The moment. You use Claude Desktop at your desk and Claude Code in the terminal. Neither knows what the other was told. You mention, once, that a UI felt noisy — six weeks later the assistant is confident you prefer minimalism, and there is no way to see where that came from or correct it. Meanwhile the decision you actually made is nowhere.

This puts the same context in a file you own — and gives you the tools to inspect it: what it remembers, why, and what you replaced. It works in Claude Desktop, Claude Code, Cursor or any MCP client, and every client shares one database.

Why not the assistant's built-in memory

Built-in memory is a black box: scoped to one tool, unreadable, unexportable, and it cannot tell you why it believes something. Here every memory says where it came from, contradictions surface as a question, replacing a belief keeps the old version readable, and any memory can be archived.

• Portable — one file, or one Worker. Change client or model freely.
• Auditable — every memory carries its evidence, its confidence, and a note that says whether you stated it, confirmed it, or the system inferred it.
• Reversible — archive any memory from the list, including one the system stored on its own.
• Queryable — “what is planned today”, “what did I promise”, “what is at risk”.

Tools

| Tool | What it does | |---|---| | capture | records a thought, commitment or decision — local-first, model if available | | list_today | what to do now, what is next, what is at risk | | list_commitments | the full list, filterable by status | | list_work_streams | grouped view of the same commitments | | list_memories | active memories, pending candidates, and what you have replaced — each with a note on why it is active | | confirm_memory | activates a candidate, optionally replacing an older belief | | reject_memory | archives a memory, including one the system stored on its own | | list_actions | what the Action Gate is holding, and what it decided | | decide_action | approve or reject a waiting action; critical needs two approvals | | revoke_action | take back a proposal or an approval |

capture, list_today and the action tools return a short card, with the full payload kept in structuredContent:

capture  "Ship the onboarding page next week"

Captured · 1 commitment
  commitment  Ship the onboarding page next week  ·  window 2026-09-19 → 2026-09-26

list_actions

Actions · 1 waiting
  high      memory_deleted  ·  waiting  ·  replaces_a_belief
  Trust 0.50 · level 2 · 0 approved / 0 rejected

Two prompts ship as slash commands:

/mcp__aldus-palace__capture 下周三前把 v0.2 的技术文发出去
/mcp__aldus-palace__today

Tool sets

Install once, enable what you need. A smaller tool surface is selected more accurately and costs fewer tokens.

| Profile | Tools | |---|---| | full (default) | all ten | | capture | capture | | today | list_today, list_commitments | | memory | list_memories, confirm_memory, reject_memory | | workstreams | list_work_streams | | actions | list_actions, decide_action, revoke_action |

ALDUS_PALACE_PROFILE=memory aldus-palace-mcp            # via env
aldus-palace-mcp-today                                  # via a focused binary

Focused binaries: aldus-palace-mcp-capture, -today, -memory, -workstreams, -actions.

Claude Desktop

Settings → Developer → Edit Config:

{
  "mcpServers": {
    "aldus-palace": {
      "command": "npx",
      "args": ["-y", "@aldus-palace/mcp"],
      "env": { "ALDUS_PALACE_PROFILE": "full" }
    }
  }
}

Restart, then check that the tools are listed.

Claude Code

claude mcp add aldus-palace -- node "$(npm root -g)/@aldus-palace/mcp/dist/index.js"
claude mcp list          # aldus-palace: ✔ Connected
claude -p "/mcp__aldus-palace__capture 周五前把发布说明写完"

Point it at a server instead

Local SQLite needs no server. If you already run one (Docker or Cloudflare), use it so every client shares one database:

{
  "command": "npx",
  "args": ["-y", "@aldus-palace/mcp"],
  "env": {
    "ALDUS_PALACE_API_URL": "http://127.0.0.1:8787",
    "ALDUS_PALACE_API_TOKEN": "your-token"
  }
}

Configuration

| Variable | Default | Meaning | |---|---|---| | ALDUS_PALACE_PROFILE | full | tool set | | ALDUS_PALACE_API_URL | — | point at a running server, using its database | | ALDUS_PALACE_API_TOKEN | — | bearer token for that server | | ALDUS_PALACE_DB | ~/.aldus-palace/aldus.db | local SQLite path | | ALDUS_PALACE_USER_NAME / _TIMEZONE / _LANGUAGE | Local User / UTC / en | local user defaults | | LLM_PROVIDER | auto | dev runs fully offline; anthropic / deepseek / openai-compatible for full quality |

Notes

• One writer per file. SQLite allows a single writer, so run one server per database — use profiles to expose a subset of tools from the same process. If you need several processes, point the extra ones at ALDUS_PALACE_API_URL.
• Everything is logged. Captures made through MCP appear in the action log exactly like captures made anywhere else.
• Databases are interchangeable. The MCP server applies the same migrations as the reference server, so a file can move between them.

Documentation

• Capability 9: MCP server
• Integration · Deployment

Apache-2.0.