Cellium Agent
Self-Learning AI Agent
Based on microkernel architecture, supporting any OpenAI-compatible API. Self-learning Agent driven by Control Loop, integrated with Strategy Gene for experience evolution and failure self-optimization.
Use Cases
Versatile scenarios powered by Cellium Agent
Code Development
Automate code generation, review, refactoring, and debugging workflows
Daily Assistant
Intelligent chat, information retrieval, file management, and task automation
Scheduled Tasks
Create timed tasks via natural language, auto-execute and push results to channels
Multi-Platform Access
Unified access via QQ Bot, Telegram, WebUI, with consistent cross-platform experience
Core Architecture
Six foundational capabilities powering the Agent brain
Runtime Self-Awareness
Real-time perception of running state (progress, stagnation, loops, saturation), dynamically adjusting decision strategies
Control Loop Architecture
Closed-loop control of decision-execution-feedback-learning in each iteration, achieving adaptive decision optimization
Self-Learning System
Action selection based on Bayesian Bandit, continuously optimizing decision strategies
Three-Layer Memory
Personality memory + Session memory + Long-term memory, FTS5 full-text + optional API vector hybrid recall
Strategy Gene Evolution
Based on GEP paper, encodes experience into compact Gene objects, auto-extracts Avoid Cues from failures for self-evolution
Heuristic Decision Engine
Rule-based feature extraction + Bandit for tie-breaking, balancing interpretability and learning ability
Core Architecture
Microkernel architecture with self-learning, control loop and core layers
Extended Capabilities
More features to enhance your Agent experience
Tool Usage Control
Dynamic tool enable/disable and recommendation switching to avoid tool-call loops
Component Sandbox
Three-layer security protection: process isolation, transparent path mapping, dangerous method interception
Event-Driven Architecture
Pub/sub pattern based on EventBus for loose coupling between components
Flash Mode
Skip memory injection to accelerate simple tasks
Scheduled Tasks
Support interval, daily, weekly tasks created via natural language, auto-push results to channels
Background Component Events
Components can run in background and proactively trigger Agent execution, supporting real-time data push and auto-analysis
Skill Management
Complete Skill package management supporting install, list, search, and uninstall from archive files
Multi-Channel Access
Support external platforms like QQ Bot, Telegram, with unified message routing and file transfer
Component Self-Extension
Agent can dynamically create and inject new components at runtime through built-in ComponentBuilder, achieving self-modification and evolution without restarting the service
Sensitive Info Control
Auto-detect and redact sensitive info like private keys, tokens, passwords; supports write blocking
Why Cellium?
Three core advantages that make Cellium Agent stand out
Universal API Compatibility
Supports any OpenAI-compatible API out of the box — just fill in your API key, model, and endpoint, no vendor lock-in
Self-Learning Evolution
Strategy Gene encodes successful patterns from experience, automatically learns from failures with Avoid Cues, growing smarter over time
Modular & Extensible
Component-based architecture lets you dynamically inject new capabilities at runtime without restarts, with built-in sandbox security
Ecosystem Compatibility
Any OpenAI-compatible endpoint works out of the box — coding plans highlighted
Coding Plans
Major Chinese AI providers offer coding-specific plans with high prefix cache benefits
PromptDiff Prefix Cache
Auto-tracks message diffs between requests — stable prefixes hit API cache, reducing latency and cost
Quick Start
Get started with Cellium Agent in just a few steps
Install with pip
pip install cellium-agent
cellium-webConfigure Model
config/agent/llm.yamlEdit config file to set API key
Start Service
cellium-webAfter startup, visit http://localhost:18000 to open the chat interface