← Back to Home
AI & automation

Otonom Dosya-Yönetim Asistanı

Autonomous CLI assistant that uses tools on the local machine

PythonFunction callingDeepSeek APIDockerSubprocess Isolation

Screenshots & Schematics

Autonomous Agent Security & ExecutionReasoner · Chat · Subprocess · SSRF Jail
User Goal / CLINatural LanguageDual-Model Braindeepseek-reasoner (Planlama)deepseek-chat (Function Call)Security & Isolation JailNo-Shell Subprocess · Command Allowlist · Path GuardsSSRF Prevention · DNS Pinning · Manual Confirm on High-RiskHost File/Cmd Toolsread_file, write_fileAudit Ledger (SQLite)Masked API Keys/Secrets

Problem & Challenge

The danger of autonomous LLM agents executing destructive terminal commands, falling prey to prompt injection, leaking API keys, or triggering SSRF via unchecked downloads.

How It Works

01

User goals are routed through 'deepseek-reasoner' for chain-of-thought analysis and step-by-step safety plan generation.

02

The execution phase leverages 'deepseek-chat' with structured tool calling (read_file, write_file, execute_command, download_image).

03

Terminal commands run without a shell wrapper via direct subprocess invocations restricted to a strict command allowlist.

04

Image and asset downloads pass SSRF validation with DNS pinning to block private IP space attacks.

05

Every operation is logged into an auditable database with secrets and bearer tokens redacted in real time.

Architecture & Technical Decisions

Python 3.12 runtime, function calling layer, DeepSeek API dual-model hybrid architecture, Docker container sandbox, allowlist security jail, and redacted audit logging.

An assistant that reads and writes files, downloads images and runs terminal commands, using separate models for planning and execution. The real work sits in the security layer: commands run without a shell and only from an allowlist, secrets and system directories are closed to both reads and writes, downloads are SSRF-protected with DNS pinning, and risky operations demand manual approval even in auto-approve mode. Every command lands in an audit log with its secrets redacted.