SimpleAgent: A Lightweight Local AI Programming Assistant for Edge Devices

Title: SimpleAgent: A Lightweight Local AI Programming Assistant for Edge Devices Abstract: A terminal AI programming assistant designed specifically for small models. Through structured workflows, secure patch application, and human-machine collaboration, it enables local models with 4B parameters to perform practical code editing tasks. Original author/maintainer: weirenong Source: GitHub (Link: https://github.com/weirenong/simpleagent) Release time: May 24, 2026 Core idea: Compensate for small model weaknesses through strong structure, high-quality context, security patches, and human review, allowing local models with 4B parameters to execute useful code editing work. This proves that excellent system design can enable small models to create value in specific scenarios.

SimpleAgent's design targets common failure modes of small models:

1. Format errors: Understands the task but outputs patches with incorrect formats
2. Format wrapping issues: Generates correct code but wraps it in unusable formats
3. Whitespace sensitivity: Forgets precise whitespace characters leading to Python patch failures
4. Framework detail omissions: Creates components but forgets framework-specific layout calls
5. Partial code masquerading as complete: Outputs partial code but pretends it is the entire file
6. Multi-task degradation: Performs well on single instructions but degrades when mixing multiple goals

Core design principles of SimpleAgent:

1. Explicit and refreshable file context: Use /attach to load files, and context is automatically refreshed after editing
2. Conservative patch application: Displays a review interface before writing to files, does not blindly apply model outputs
3. Differences as safety signals: Uses differences as safety signals, supports safe (boundary-only changes)/risky (all changes) application modes (F2/F3/Esc shortcuts)
4. Robust parser: Supports multiple editing formats (SEARCH/REPLACE, full-file/partial fenced code, unified diffs, etc.), normalizes filename lines
5. Maintain human-machine collaboration: Emphasizes developer review of differences; tools assist efficiently but do not replace human judgment
6. Prioritize locally inspectable workflows: Workflows are pure Markdown files, easy to edit, inspect, and version control

Features

• Code editing workflow: Model output → Parse edits → Generate diffs → Classify safe/risky → User review → Apply (F2/F3)
• Role system: Define different role personas, customize prompts and behavior patterns
• Attachments and local RAG: /attach loads files, supports web content crawling as context, enhances with embedded content retrieval
• Terminal experience: Clean interface, real-time streaming output, shortcut operations, workflow debugging mode

Practical Application Scenarios

• Edge device programming: Runs locally without cloud calls, low memory usage, suitable for offline or privacy-sensitive scenarios
• Rapid prototyping: Workflow templates start tasks, secure patches reduce trial-and-error costs, human-machine collaboration ensures quality
• Programming learning assistant: Small models provide concise explanations, diff reviews help understand code changes, local operation has no API fees

Installation and Usage

• Quick installation: pipx install weirenong-simpleagent
• Launch: simpleagent
• Configure Pollinations API: /api-pollinations

Recommended Models

• Local Ollama models: 4B-level small models like nemotron-3-nano:4b
• Pollinations.ai models: Support more complex workflows
• Ollama cloud models: For heavier tasks

Test environment: MacBook Pro M2, 16GB RAM

SimpleAgent's slogan: "Work work. Ship ship. Poor man's Claude Code for tiny local models." Project significance:

1. Lowering the threshold: Allows developers without high-end hardware or API budgets to use AI-assisted programming
2. Pragmatic design: Does not pursue model capabilities; compensates for small model weaknesses through system architecture
3. Privacy-first: Runs locally, code does not go to the cloud, suitable for sensitive projects
4. Human-machine collaboration: Emphasizes human review, avoids blind trust in AI outputs

Against the backdrop of the large model arms race, it provides a different path: making existing small models more useful through better system design rather than pursuing larger models.