Local models (Ollama on a laptop GPU) break on requests over ~19k tokens. Model-server runs now send only the tools and skill the job needs and summarise the conversation at ~16k tokens. Publish, schedule, reply and DM tools are removed from every run. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Social Manager
Metatroncube's AI social media manager: Claude + client brand files + live Meta data + team task briefs.
Setup (Windows)
- Install Node.js 20+ from nodejs.org.
- Open
.envand paste the system user token afterMETA_SYSTEM_USER_TOKEN=. - Test the connection:
cd connector npm install npm run test:meta - Use it from Claude Code in this folder (
.mcp.jsonregisters thesocial-managertools automatically), or add to Claude Desktop's config:"social-manager": { "command": "node", "args": ["D:\\MCS_Repos\\2026\\claude_Metatron_Social_Media_Manager\\connector\\src\\server.js"] }
Folder map
CLAUDE.md— the agent's job description and rulesclients/<client>/— brand, goals, constraints, history, calendars, reportsplaybooks/— platform best practicesteam/— roster and weekly task files.claude/skills/— account-audit, weekly-plan, assign-tasks, weekly-reportconnector/— MCP server for Facebook Pages + Instagrambranding/— Metatroncube logos + report stylingdashboard/— Social Command web app (clients, AI runs, PDFs, Odoo team board). Seedashboard/README.mddocs/— Claude Desktop setuptools/report-pdf/— turns report .md files into branded PDFs (node tools/report-pdf/build.mjs <file.md>)
Safety
- Publishing is off until
ALLOW_PUBLISH=truein.env, and even then each post needs explicit approval. .envis git-ignored. Never share the token.
Roadmap
- Meta Pages + Instagram read + publish (approval-gated)
- Meta Ads (separate app, ads_read first)
- YouTube Data + Analytics API
- LinkedIn Community Management API
- Google Ads (MCC + developer token)
- Push tasks to Odoo Project
Description
Languages
JavaScript
79.4%
Python
12.7%
HTML
4.1%
CSS
2.5%
Shell
0.6%
Other
0.7%