jeumpallm
AI Gateway — Unified LLM provider aggregation with Acehnese module, CLI, and sandboxed agent
jeumpallm
Running (v2.6)
jeumpallm is a Python-based AI gateway that aggregates 30+ LLM providers under a single OpenAI-compatible API. Runs on a single Android device (Termux-Proot) with autonomous provider discovery, triangle-mode failover, auto-path detection, config auto-merge, Acehnese language module, CLI streaming chat, sandboxed agent, and PWA dashboard.
What It Does
jeumpallm is not a model. It is a gateway that routes API requests to 30+ LLM providers (OpenAI, Anthropic, Google, Groq, and free/no-auth providers discovered automatically). Built to run on low-resource devices (1.1GB RAM) with zero recurring cost.
- Routes requests to best available provider (5 strategies: round-robin, priority, latency, random, fallback)
- Auto-discovers new free LLM providers by scanning GitHub + HuggingFace + probing endpoints
- Triangle mode: 2-instance SO_REUSEPORT failover, zero-downtime if one crashes
- Auto-path: all paths via jeumpa_paths.py, no hardcoded /sdcard paths anywhere
- Config auto-merge: discovered providers auto-written back to config.yaml, auto-delete after 5 failures, discovery every 60s
- AI-powered health monitor diagnoses provider status (down/rate_limited/slow/unstable)
- Auto-integrates discovered providers into routing tables and task fallbacks
- Serves OpenAI-compatible API, streaming SSE, and MCP over SSE (Claude Desktop compatible)
- Acehnese language module: 5 endpoints (detect, lookup, translate, enrich, glossary) with auto-detect middleware
- CLI with streaming chat, model search, provider probe, server management, and voice mode
- Sandboxed agent with command allowlist (30 safe commands) for code execution
- Provides multi-tier access: anonymous, free, santri (Aceh students), donor
Modules
Core
Provider Registry
Request Router
Task Router
MCP SSE
MCP JSON-RPC
Autonomous Systems
Auto Path Detection
Auto Tunnel Detection
Provider Discovery
AI Health Monitor
Provider Intelligence
Auto Integration
Access & Billing
x-api-key Auth (email:password)
Rate Limiting (4 tiers)
Donation System
Santri Program
Agent & Memory
Tool-Using Agent
Code Sandbox (allowlist)
Conversation Memory
Vector Search (Jina AI)
Shared Memory (3 scopes)
Acehnese Module
Language Detection (150+ word frozenset)
Google Translate Free API
kata.web.id Dictionary Lookup
Parallel Dual-Source Translation
Auto-detect Middleware
Glossary Injection
CLI
Chat (streaming SSE)
Models list & search
Providers list & probe
Server management
Voice mode
PWA Dashboard
Single-page app (10 files, vanilla JS)
20 pages / views
Acehnese toggle in Settings
File upload + image preview
PWA (Manifest + SW), installable
Responsive (320px–2560px)
Theme toggle (dark/light)
file:// compatible, no build step
History
v1.0 (Node.js — Archived)
- Express.js server (port 3333)
- 22 providers
- Full engine suite: MCP gateway, A2A protocol, orchestration, marketplace
- 4-state circuit breaker
- Cost estimator, model sync
Source archived in archive/proxygatellm/
v2.6 (Python — Running)
- aiohttp server (port 3456), single process
- 30+ providers + auto-discovery
- 6 autonomous modules (paths, tunnel, discovery, monitor, intelligence, integrator)
- 54 Python modules, ~13K lines
- 31+ API endpoints + Aceh module (5 endpoints)
- Acehnese language module: detect, lookup, translate, enrich, glossary
- CLI: streaming chat, models search, providers probe, server mgmt, voice
- Sandboxed agent with command allowlist (30 safe commands)
- PWA dashboard: 20 views, vanilla JS, file:// compatible
- Runs on Android Termux (1.1GB RAM)
- Triangle mode: 2 instances on same port (SO_REUSEPORT), kernel load-balanced, zero-downtime if one crashes
- Auto-path: all paths via jeumpa_paths.py, no hardcoded /sdcard paths anywhere
- Config auto-merge: discovered providers auto-written back to config.yaml
- Auto-delete: dead providers pruned after 5+ consecutive failures, discovery every 60s
Next
- Multi-payment integration (QRIS, crypto) for donor tier
- Telegram bot integration for remote administration
- RAG engine port from Seulanga (TypeScript → Python) for local knowledge retrieval
- Multi-payment integration (QRIS, crypto) for donor tier
- Telegram bot integration for remote administration