62680dd060
## Claude Code integration now fully opt-in (Fix 3) - `pbi connect` no longer writes to ~/.claude/ automatically - New `pbi-cli` entry point: `pbi-cli skills install/uninstall/list` - `pbi-cli skills install` shows exact paths before writing and requires y/N confirmation - `pbi connect` prints a one-line tip if skills are not yet installed - `pbi skills` subgroup removed from the `pbi` entry point ## DLL licensing compliance (Fix 1) - pyproject.toml updated to PEP 639 SPDX dual expression: MIT AND LicenseRef-Microsoft-AS-Client-Libraries - license-files declaration: LICENSE, THIRD_PARTY_LICENSES.md, NOTICE - THIRD_PARTY_LICENSES.md: full verbatim MS Analysis Services Client Libraries EULA - NOTICE: short-form attribution for wheel redistribution - src/pbi_cli/dlls/README.md: in-directory sentinel for the MS DLLs - setuptools requirement bumped to >=77.0 for PEP 639 support ## SECURITY.md rewrite (Fix 2) - Supported versions table updated to 3.10.x - Architecture section: no MCP server, no subprocess, direct pythonnet interop - Global Configuration Modifications section updated to reflect opt-in model - Bundled Binaries section references THIRD_PARTY_LICENSES.md ## Documentation - README.md, README.pypi.md: corrected 3-step setup flow - CHANGELOG.md: [3.10.3] entry - CONTRIBUTING.md: pbi skills -> pbi-cli skills - All 7 semantic model SKILL.md files: prerequisites updated to 3-step flow - New SVG/PNG marketing and documentation assets
1.6 KiB
MCP tools have a hidden cost most people don't talk about: 𝘀𝗰𝗵𝗲𝗺𝗮 𝗯𝗹𝗼𝗮𝘁.
Every MCP tool definition (name, parameters, descriptions) gets loaded into your context window. Not when you call it. On 𝗲𝘃𝗲𝗿𝘆 𝘀𝗶𝗻𝗴𝗹𝗲 𝘁𝘂𝗿𝗻. A single tool schema costs anywhere from 400 to 3,500 tokens just to be available.
Real benchmarks back this up: → A 5-server MCP setup burns ~55K tokens before you type anything (JD Hodges) → GitHub's MCP server with 43 tools: a simple repo query costs 44K tokens vs 1,365 for CLI (Scalekit) → Anthropic themselves measured 134K tokens of tool definitions internally
Now imagine a Power BI MCP server with 20-40 tools for all the TOM operations. At 2,000-3,000 tokens per schema, that's 𝟲𝟬𝗞-𝟭𝟮𝟬𝗞 𝘁𝗼𝗸𝗲𝗻𝘀 of overhead eating your context window before you do any actual work.
pbi-cli avoids this entirely. CLI commands cost ~30 tokens to invoke, and only when used. No schemas loaded. No per-turn tax. Skills load on-demand, not on every turn.
That's why I built it as a CLI with skills instead of wrapping an MCP server. The architecture isn't just a preference. Real benchmarks show CLI is 4-32x more efficient per task, and for schema overhead specifically, the difference is even larger.
Check the chart. The difference speaks for itself.
GitHub: https://github.com/MinaSaad1/pbi-cli Details: mina-saad.com/pbi-cli
#PowerBI #ClaudeCode #MCP #TokenEfficiency #OpenSource #VibeModeling #AI #DataModeling