Features & Technical Capabilities
Pak-LLM Coder is engineered by SBF Consultancy (Pvt.) Ltd. (sbf-consultancy.net), combining autonomous agent execution with developer productivity tooling.
Enterprise Mission Control — Four Pillars
Tri-surface verification (static, sandboxed browser, terminal/WASM) with immutable Plan, Atomic Patch Set and Verification Matrix artifacts; architecture graph with Pak-LLM BPE context slicing; AST-guarded unified diffs with hunk-level accept/reject and self-healing; concurrent workers in isolated virtual branches or real Git worktrees with a human sign-off merge gate.
Decoupled SDK — Lifecycle Hooks & Subagents
TypeScript (@pak-llm/sdk) and Python (pakllm-coder) SDKs with three hook classes: read-only Inspect telemetry, fail-closed Decide authorization, and ordered Transform rewriting. Exit verification blocks completion while tests or linters fail, and ephemeral subagent trees run deep searches in fresh context windows with isolated toolsets.
Smart Paste Transpilation & Autocomplete Snooze
Clipboard detection auto-transpiles alien stack syntax (Python dicts, CSS rules, JSON, SQL DDL) into target TypeScript/JSX dialects with 1 click. Quick status bar snooze micro-toggle pauses AI suggestions for 15m or 1h for uninterrupted typing.
Visual Element Inspector & Bidirectional JSX Sync
Inspect live preview DOM elements with real-time class name and dimension tooltips. Click any element to immediately jump to the exact JSX component and line number in the Monaco code canvas.
Side Chats (/side, /btw) & Transcript Search (Cmd+K)
Fork isolated exploration threads that investigate hypotheses without polluting your active workspace plan. Instant inverted-index search (Cmd+K) allows searching all past conversation turns, prompts, and diff proposals with TF-IDF ranking.
Cross-Platform Biometric & Passkey Lock Suite
Fail-closed device locking across Web, Windows (Hello), macOS (Touch ID), iOS (Face ID), Android (BiometricPrompt), and Linux with 6 modalities: Passkeys (WebAuthn/FIDO2), PIN, Password, 2FA TOTP, and Gesture Pattern under NIST SP 800-63B, ISO 27001, and SGS standards.
Universal Workspace & Past Thread Management
Full CRUD lifecycle (Add, See, Review, Edit, Delete) for past workspaces and conversation threads across Web IDE, Desktop, CLI, SDK, and Mission Control, synchronized into the central Admin Profile for manual review.
Autonomous CodeAct AI Agent Engine
Full-cycle autonomous software engineer implementing the CodeAct paradigm. Emits real-time reasoning events, executes shell commands, computes unified diffs, tracks plan milestones, and measures token costs.
Multi-Language WASM Runtime Interpreters
Embedded in-browser execution kernels supporting Python 3.12 (Pyodide WASM with scope evaluator), Node.js v22 (V8 Sandbox), TypeScript 5.5, Rust 1.78, Go 1.22, SQLite 3.45, and POSIX Bash.
Terminal Multiplexer & Sovereign Shell Suite
Integrated multi-tab terminal suite with dual split panes (Columns view), CPU/RAM resource telemetry, interactive shell commands (ls, cat, python, node, git, agent), and Code Runner stdout/stderr streaming.
NIST CSF 2.0 & OWASP Security Sentinel
Real-time static code analysis, automated vulnerability auditing, and pre-inference heuristic scrubbing that redacts high-entropy API secrets, JWTs, and private credentials before reaching the router.
Visual Git Diff Inspector & Source Control
Command Palette (Ctrl+Shift+P / F1) for rapid action execution, Git Source Control panel with AI-generated commit messages and staging, and Side-by-Side Diff Reviewer for visual inspection before applying AI code.
Sovereign Token Metering & Auto-Replenish
Real-time token consumption metering, automatic quota replenishment formula (Monthly Quota / 30 added upon exhaustion), SMTP email notifications at 75%, 50%, 25%, and 0%, and 0% Riba Swich payment gateway.
Bilingual Urdu & English AI Pair Programmer
Native bilingual prompt processing and reasoning. Generate complex software architectures, explain algorithms, and comment code in both English and fluent technical Urdu.
Multi-Platform Native Desktop Installers
12 verified standalone native installers for Windows (.exe setup wizard, .msi enterprise), macOS (.dmg disk image, .zip bundle), and Linux (.AppImage portable, .deb Debian/Ubuntu) with SHA-256 integrity verification.
Pak-LLM CLI Coding Agent & Terminal Shell
Decoupled terminal client with a reactive TUI (side-by-side diff review, bilingual English/Urdu logs), in-memory snapshot rollback via undo or double Esc, and headless CI execution (-p, --auto, --json-report). Human-in-the-loop bash safety gates remain enforced.
Sovereign Extension Marketplace
Curated ecosystem of extensions across AI Agents, Languages, Linters/Formatters, Git/DevTools, and Themes with instant search, category filtering, and zero-telemetry verification.
Model Context Protocol (MCP) Server
Standardized MCP JSON-RPC server mode (`pak-llm --mcp`) plus a SDK adapter that dynamically loads the official @modelcontextprotocol/sdk when present and falls back to a dependency-free JSON-RPC 2.0 dispatcher for air-gapped and browser installs.
Report Bug & User Feedback Sync
Integrated bug reporting and feedback system across Web, Desktop, and CLI. Direct database persistence, real-time webhook broadcast to the Pak-LLM Admin Dashboard, and automated two-way SMTP email notifications.
Bespoke Android & iOS Mobile Applications
Native mobile & tablet engineering studio adapted to top 50 popular screen sizes (phones, foldables, tablets). Features touch-optimized Monospace editors, floating 44px+ touch bottom bars, dropdown menus, offline WASM runtimes, biometric authentication, and zero ad telemetry.
Technical Specifications & Architectural Matrix
Verified engineering parameters of the Pak-LLM Coder system
| Subsystem | Technology / Engine | Execution Boundary | Security & Compliance |
|---|---|---|---|
| Foundation Coder Model | saadfarrukh191/PakLLM_Coder (32B MoE) | Dedicated Sovereign Cluster | Zero-Retention Policy |
| In-Browser Python Runtime | Python 3.12 (Pyodide WebAssembly) | Client RAM Sandbox | Zero Server Exposure |
| In-Browser Node.js Sandbox | Node.js v22 (V8 Sandbox Engine) | Isolated Web Worker | No Host Access |
| Prompt Secret Sanitizer | Prompt Security Sentinel Heuristics | Edge & Client Pre-Inference | OWASP A03 / NIST CSF 2.0 |
| Tri-Surface Verification | Static AST + Sandboxed Browser + Terminal/WASM | Client Sandbox & Sovereign Shell | ISO/IEC 25010 / NIST SI-2 |
| Architecture Semantic Graph | Tree-sitter-inspired lexer + BPE context slicer | In-Memory Graph (Browser/Node) | ISO/IEC 25010 Performance |
| AST-Guarded Patch Engine | jsdiff Myers hunks + Language Lexer Guard | Pre-Write Virtual Filesystem | NIST SI-7 / OWASP LLM02 |
| Multi-Worker Isolation | Virtual Branch Snapshots / Opt-in Git Worktrees | Per-Worker Isolated Snapshot | NIST AC-3/AC-6 / ISO/IEC 42001 |
| Domestic Payment Gateway | Swich Fintech (JazzCash / EasyPaisa / 1Link) | State Bank of Pakistan Rails | 0% Riba / Shariah Compliant |
| Statutory Tax Settlement | 1% Statutory Tax Handled via Swich | Zero Duplicate FBR Submission | Statutory Tax Verified (Swich Settled) |
| Mobile Studio & Apps | Android (Universal APK/AAB) & iOS (Signed IPA/OTA) | Sandboxed Storage & Secure Enclave | OWASP Mobile Top-10 / Zero Ad Telemetry |
Architecture & Features FAQs
Detailed technical answers regarding the autonomous loop and runtimes.
Mission Control now runs four coordinated pillars: (A) tri-surface verification with immutable Plan, Atomic Patch Set and Verification Matrix artifacts; (B) a living architecture graph that indexes imports, routes, schemas, components and services and injects BPE-budgeted context slices; (C) an AST-guarded unified-diff engine with hunk-level accept/reject and self-healing; and (D) concurrent workers isolated in virtual branch snapshots or opt-in Git worktrees. Worker branches merge only after milestone completion, a passing verification matrix and explicit user sign-off.