AGENT SYSTEMS, ZERO COST. M4ST://
I'm Anuj — a solo systems architect from Bareilly, India. I wire 21 MCP servers + a voice-enabled bridge core across 11 free LLM providers into agent systems that run on a ₹0 cloud budget.
One brain, many providers.
A task-intent router decides which of 11 free LLM providers handles each prompt — quality, speed, or free — with SMART_KEY auto-detect and a 3600s semantic cache.
20 real MCP servers. Live.
Every server below exists in the MAST repo — real Python MCP servers, not mockups. ★ marks the genuinely novel ones.
The glue between servers.
bridge_core/ in the MAST repo — voice, healing, brain and routing that hold the 20 servers together.
Bidirectional Voice
Vosk wake word + Groq Whisper STT, pyttsx3 / gTTS speak-back, and live barge-in — interrupt the agent mid-sentence. Hinglish-optimized.
voice.pySelf-Healing Engine
Errors are classified, fixed, retried up to 3×, then routed to an alternative — every incident logged to a healing log that makes the system smarter.
self_healing.pySmart Key Brain
11 providers with auto key-prefix detection and live .env reload. BRAIN=auto/fast/quality/free picks the right model per task.
smart_brain.pyTask Chain Router
9 task chains, 3600s cache, Hinglish routing hints — "research this then summarize in Hindi" lands on the right provider chain automatically.
brain.py28 hot-reloadable skills.
Every skill is a SKILL.md the agent can load, learn from task sequences, and even write new skills — without a restart.
It runs while you read.
A live feed of the autonomous loop — tasks routed, servers responding, memory written. This is what M4ST does every second.
Built to be shipped, not shown.
Everything here runs — locally, on free tiers, and in the agent stack itself.
MAST Framework
The whole stack — 21 MCP servers, bridge_core with voice + self-healing + smart brain, 28 skills, agents and commands. The system this site advertises is the system that runs it.
mcp_servers/ + bridge_core/ →mast-llm-router
An MCP server exposing 8 tools: chat, stream, batch, multi-turn, task detection, cache control, provider listing and router status. Built with fastmcp, pinned to a stable protocol.
src/server.py →Universal MCP Config Gen
One source of truth generates MCP configs for 6 IDEs — Cursor, Windsurf, Claude Desktop, Antigravity and Codex — so the whole stack connects everywhere with zero drift.
config_generator.py →Let's build agent systems.
Open to systems architecture roles, agentic AI work, and collaboration on autonomous stacks. Based in Bareilly, working anywhere.
hello@m4st.in