AI-Powered
Product Engineering Assistant

Beyond code generation — gt ships production-ready features with design systems, tracking pixels, health checks, and quality guardrails built-in. Bring Your Own Keys. Run with Groq, OpenAI, or locally with Ollama.

Get Started Explore Skills
$ gptcode do "add user authentication"
AI AGENT EXECUTION FLOW VISUALLY EXPLAINED

Input Analysis · Transformers

Watch how the transformer processes your command through self-attention to generate intent predictions.

Embedding
Transformer Block · Multi-head Self Attention
MLP Intent Probabilities
Embedding
chu
do
"add
user
auth"
Transformer Block
Key (K)
command structure
operation type
target entity
Query (Q)
add user auth
security feature
Value (V)
authentication context
user management
Attention
MLP
chu
do
"add
user
auth"
Intent Classification
authentication_add
60%
user_create
18%
security_config
8%
AI AGENT EXECUTION FLOW VISUALLY EXPLAINED

Intelligence Routing

Based on the intent classification, the system routes to the appropriate execution mode.

Request Features
Intent: Modify Code
Complexity: High
Context: 14 Files
Decision Factors
λ
Model Recommender
Profile: Balanced
Estimated Complexity
Autonomous Agents
Analyzer
Active
Scans codebase & builds dependency graph
→ src/auth/index.ts
→ src/middleware/jwt.ts
Planner
Planning
Creates structured implementation plan
1 Define JWT schema
2 Update middleware
Editor
Editing
Applies code changes to files
+ export const authMiddleware...
Validator
Testing
Runs tests & linters
✓ All tests passed
Execution Mode
Orchestrated
92%
Single Agent
6%
Direct LLM
2%
Ensemble Optimization
XGBoost
(Speed)
KAN
(Logic)
Unbiased Sampling
Calculating...
AI AGENT EXECUTION FLOW VISUALLY EXPLAINED

Agent Execution · Orchestrated

Multiple agents collaborate sequentially to analyze, plan, edit, and validate changes.

Execution Timeline
Analyzer 2.1s
Planner 1.8s
Editor 4.1s
Val 1.2s
Analyzer
• Scan codebase
• Identify patterns
• Extract context
Planner
• Design approach
• Define steps
• Set constraints
Editor
• Apply changes
• Generate code
• Refactor files
Validator
• Run tests
• Check lints
• Verify syntax
How Agents Collaborate
Analyzer Role
Scans the entire codebase using AST parsing to build a dependency graph. Identifies relevant files, extracts context, and provides structural insights for subsequent agents.
Planner Role
Takes high-level goals and breaks them into executable steps. Uses context from Analyzer to create structured implementation plans with constraints and dependencies.
Editor Role
Executes the plan by generating code snippets, applying changes to files, and refactoring existing code. Includes auto-recovery for tool failures.
Validator Role
Ensures correctness by running automated tests, linters, and syntax checks. Validates that changes meet quality standards before completion.
AI AGENT EXECUTION FLOW VISUALLY EXPLAINED

Response Generation · Transformers

The final response is decoded from execution context using the transformer.

RAG Context Transformer Decoder
Generated Response
RAG Context
Retrieved Files
→ auth/index.ts (342 lines)
→ middleware/jwt.ts (89 lines)
→ tests/auth.test.ts (156 lines)
Agent Results
✓ Files modified: 3
✓ Tests passed: 12/12
✓ Lints clean
Context: 4,102 tokens
Window: 128K tokens
Transformer Decoder
1. Context Embedding
RAG retrieves code and agent results as vector embeddings.
2. Self-Attention
Cross-attention captures relationships between code and results.
3. Token Generation
Softmax sampling with temperature control.
Model Configuration
Model: llama-3.1-8b
Params: 8B
Context: 128K tokens
Temp: 0.7
Top-p: 0.9
Performance Metrics
Tokens/sec: 142
Latency: 48ms
Cost: $0.0012
Cache hit: 73%
Generated Response
Token-by-token generation with softmax sampling
Done!
95%
I've
92%
added
88%
authentication
97%
Final Response
"Done! Added authentication. All 12 tests passing."
28 tokens · 196ms
Model claude-3-5-sonnet
Cost $0.000000
Time 0.0s
Status IDLE

Autonomous Execution

gt do "task" orchestrates 4 specialized agents: Analyzer → Planner → Editor → Validator. Auto-retry with model switching when validation fails.

Validation & Safety

File validation prevents creating unintended files. Success criteria auto-verified. No surprise scripts or configs. Supervised vs autonomous modes.

Intelligent Context

Dependency graph + PageRank identifies relevant files. 5x token reduction (100k → 20k). ML intent classification (1ms routing).

Radically Affordable

$0-5/month vs $20-30/month subscriptions. Use Groq for speed, Ollama for free. Auto-selects best models per agent.

Interactive Modes

gt chat for conversations. gt run for tasks with follow-up. Context-aware from CLI or Neovim plugin.

Manual Workflow

Break down complex tasks: gt research → gt plan → gptcode implement. Full control when you need it.

🎯 15 Product Skills

Beyond code: Design Systems, Product Metrics, Production Guardrails, QA Automation. Language idioms + product best practices built into every prompt.

Agent Orchestration: Analyzer → Planner → Editor → Validator

Fast routing, focused context, safe edits, and verified results — end to end

Analyzer

Scans the codebase, builds dependency graph and selects only the relevant files

gt do "add authentication"

Planner

Creates a concrete plan with success criteria and allowed file list

gt do "add authentication"

Editor

Applies changes incrementally with file validation and auto-recovery

gt do "add authentication"

Validator

Runs tests and checks success criteria before finishing

gt do "add authentication"

Learn the Architecture See Validation & Safety

Quick Start

1. Install CLI

# One-line install (creates gptcode and gt commands)
curl -sSL https://gptcode.dev/install.sh | bash

# Or using go install
go install github.com/gptcode-cloud/cli/cmd/gptcode@latest

# Setup
gt setup

2. Add Neovim Plugin

-- lazy.nvim
{
  dir = "~/workspace/gptcode/neovim",
  config = function()
    require("gptcode").setup()
  end,
  keys = {
    { "<C-d>", "<cmd>GPTCodeChat<cr>", desc = "Toggle Chat" },
    { "<C-m>", "<cmd>GPTCodeModels<cr>", desc = "Profiles" },
  }
}

3. Start Coding

# Use gptcode or the short alias gt
gt do "add user authentication with JWT"
gt chat
gt research "best practices for error handling"
gt plan "implement rate limiting"

Core Capabilities

Three Ways to Work

Special Modes

Intelligence & Optimization

Developer Experience

Why GPTCode?

GPTCode isn't trying to beat Cursor or Copilot. It's trying to be different—and yours.

Read the full positioning → · Original vision →
Agent routing vs tool search → · Intelligent model selection →
Dependency graph → · Chat REPL →
🆕 Universal Context Management →

Documentation