AI Desk Tools

Local MCP tools for personal AI agents.

This repo is the executable MCP tool layer for:

ntaffzii/Skill-Agents

Skill-Agents  = skills, workflows, docs, examples, provider config
ai-desk-tools = MCP server, Python tools, tests, runtime safety policy

Use this repo when you want LM Studio, Claude Desktop, Claude Code, Codex, or a local agent to call real tools on your machine.

What It Includes

• MCP server over stdio: server.py
• MCP server over HTTP/SSE-style transport: server_http.py
• 50+ local tool groups
• skill routing and compact context building
• prompt improvement with local model fallback
• read-only/draft-only defaults for private data
• security policy and audit logging
• unit tests for tool behavior

Companion Skill Repo

For skills, workflows, local LLM prompts, provider config, and usage docs:

ntaffzii/Skill-Agents

Recommended pairing:

Skill-Agents/examples/local-llm-agent-prompt.md
Skill-Agents/docs/SKILL_RUNTIME_FLOW.md
Skill-Agents/docs/LOCAL_LLM_SETTINGS.md
Skill-Agents/docs/PROMPT_IMPROVER_LOCAL_MODEL.md
Skill-Agents/config.json

Complete three-repository setup:

Skill-Agents Complete Local AI System Guide

Quick Start

Clone:

git clone https://github.com/ntaffzii/ai-desk-tools.git
cd ai-desk-tools

Install:

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt

Optional Playwright browser support:

playwright install chromium

Run stdio MCP server:

python .\server.py

Run HTTP MCP server:

python .\server_http.py --host 127.0.0.1 --port 8765

Use stdio first unless your MCP client specifically supports HTTP MCP.

LM Studio Setup

In LM Studio:

Program
-> Install
-> Edit mcp.json

Add:

{
  "mcpServers": {
    "ai-desk-tools": {
      "command": "C:\\path\\to\\ai-desk-tools\\.venv\\Scripts\\python.exe",
      "args": [
        "C:\\path\\to\\ai-desk-tools\\server.py"
      ]
    }
  }
}

With prompt improver local model:

{
  "mcpServers": {
    "ai-desk-tools": {
      "command": "C:\\path\\to\\ai-desk-tools\\.venv\\Scripts\\python.exe",
      "args": [
        "C:\\path\\to\\ai-desk-tools\\server.py"
      ],
      "env": {
        "PROMPT_IMPROVER_API_URL": "http://localhost:1234/v1/chat/completions",
        "PROMPT_IMPROVER_MODEL": "LFM2.5-8B-A1B"
      }
    }
  }
}

LFM2.5-8B-A1B is the recommended starter model for prompt improvement when available. It is not required.

Claude Desktop Setup

{
  "mcpServers": {
    "ai-desk-tools": {
      "command": "C:\\path\\to\\ai-desk-tools\\.venv\\Scripts\\python.exe",
      "args": [
        "C:\\path\\to\\ai-desk-tools\\server.py"
      ]
    }
  }
}

Environment Variables

Only set what you use. Do not commit real tokens.

$env:NOTION_TOKEN="..."
$env:GITHUB_TOKEN="..."
$env:FIGMA_TOKEN="..."
$env:SLACK_BOT_TOKEN="..."
$env:POSTGRES_DSN="..."
$env:FIRECRAWL_API_KEY="..."

Prompt improver:

$env:PROMPT_IMPROVER_API_URL="http://localhost:1234/v1/chat/completions"
$env:PROMPT_IMPROVER_MODEL="LFM2.5-8B-A1B"

If PROMPT_IMPROVER_API_URL is not set, prompt improvement still works with a rule-based fallback.

Local LLM Flow

When paired with Skill-Agents, use this flow:

User request
  -> skill-runtime.route_request
  -> prompt-improver if unclear
  -> skill-runtime.build_agent_context
  -> recommended toolset/tools
  -> final answer with verification

Example:

Use skill-runtime first. Route this request, improve it only if unclear, then load selected workflow and skills:

Build today's plan from Notion, Obsidian, calendar, inbox, chat, memory, and open issues.
Draft only. Do not send or apply anything.

Tool Groups

Runtime And Routing

• registry - inspect tools, workflows, runtime capabilities, allowed roots, and policy.
• skill-runtime - index skills, route requests, load selected workflows/skills, and build compact local-LLM context.
• toolsets - recommend curated tool groups for job types.
• audit - inspect audit logs and policy denials.
• mcp-security-audit - classify MCP tools by risk and policy coverage.
• system - inspect environment and command availability.

Project And Code

• filesystem - list, read, search, and inspect files.
• project - detect stack, scripts, important files, and health.
• docs - find documentation and build context bundles.
• repo-index - build and search lightweight repo maps.
• package - inspect manifests, dependencies, and lockfiles.
• code-editing - write, patch, preview diffs, format, and run tests.
• validation - plan and run allowlisted validation commands.
• test-inspection - find tests and map source to test files.
• ci - inspect CI files and validation commands.
• structured-data - read, validate, and patch JSON/YAML/TOML.
• sandbox - create temporary safe workspaces and compile snippets.

Git And GitHub

• git - read-only Git status, diff, log, show, branch.
• git-control - create/switch branches, stage/unstage, commit.
• github - inspect local GitHub metadata and draft PR descriptions.
• github-api - read repo, issue, PR, changed files, and checks when token is configured.
• issue-tracker - parse issue references, draft issues, break down tasks, and plan updates.

Backend And Security

• api - inspect routes, endpoints, OpenAPI, and API config.
• database - inspect schema files, migrations, ORM models, and database config.
• postgres - plan and run read-only Postgres queries when configured.
• config - inspect config files, env keys, and secret hygiene.
• dependency-risk - inspect dependency risk signals.
• security-scanner - scan for likely secrets, dangerous commands, env exposure, and dependency risks.
• docker - inspect Dockerfile/Compose and plan Docker validation.
• release - inspect versions, changelogs, and release readiness.
• backup - plan/create/list zip snapshots inside allowed roots.

Personal Workspace

• notion - search/read Notion and draft page/block payloads.
• obsidian-notion-bridge - plan safe Obsidian/Notion conversions.
• calendar - summarize supplied events, build daily plans, draft meeting prep.
• email-inbox - summarize supplied email messages, extract action items, draft replies.
• slack-discord - search/summarize messages, draft replies, extract actions.
• memory - save/search/summarize local memories.
• memory-context - save typed decisions, preferences, lessons, and context packs.
• vector-memory - lightweight semantic memory search.
• rag-adapter - chunk text, plan RAG indexes, and draft embedding requests.

Browser, Web, Media, Finance

• browser - inspect browser readiness, static HTML, and localhost URLs.
• browser-page-map - map HTML headings, links, forms, buttons, and inputs.
• playwright - inspect live pages and capture screenshots.
• playwright-actions - click/fill/assert text, inspect console/network, accessibility snapshots, persistent sessions.
• figma - inspect Figma files and draft frontend implementation plans.
• web - search/fetch/extract/summarize sources.
• web-capture - provider-neutral public webpage capture with social-site safety limits.
• finance-market - quotes, crypto prices, finance-news plans, watchlists, position risk.
• media - inspect/process images, audio, and video.
• prompt-improver - analyze, rewrite, score, and template prompts.
• external-mcp-catalog - compare public MCP patterns and draft local adaptations.
• task - scan TODO/FIXME/HACK/BUG markers and roadmap files.
• user-runner - write command handoffs and user-run PowerShell scripts.

HTTP Server

python .\server_http.py --transport streamable-http --host 127.0.0.1 --port 8765

If your client expects SSE:

python .\server_http.py --transport sse --host 127.0.0.1 --port 8765

Keep 127.0.0.1 for personal use. Do not expose this server publicly without authentication, firewalling, per-user policy, and audit review.

Safety Model

Defaults are intentionally conservative:

• Email/chat/issue/Notion actions are draft or plan oriented.
• Postgres rejects mutating SQL and only runs read-only queries.
• Web capture does not bypass login, CAPTCHA, private accounts, or paywalls.
• Git push, force push, reset hard, merge, and rebase are not exposed.
• File and command access are controlled by config/tool_policy.json.
• Audit logs are configured in config/tool_policy.json.

Project Structure

ai-desk-tools/
  README.md
  requirements.txt
  security.py
  server.py
  server_http.py
  trusted_sources.json
  config/
    tool_policy.json
  prompt_engine/
  tools/
  tests/

Tests

python -m unittest discover -s .\tests

If this repo is inside the full Skill-Agents workspace:

powershell -NoProfile -ExecutionPolicy Bypass -File ..\scripts\validate-all.ps1

License

See LICENSE in the companion repo or add your preferred license file for this standalone tools repo.