Multi-phase cognitive reasoning, tree-sitter AST indexing, subagent fleet orchestration, and autonomous compiler error diagnostics — powered by Gemini & Vertex AI.
# Install latest Automergent binary + 'amt' alias
$ curl -fsSL https://automergent.github.io/install.sh | bash
# Export your Gemini API Key
$ export GOOGLE_API_KEY="your-api-key"
# Launch interactive terminal session
$ automergent # or: amt
# Requires Go 1.25+ with CGO enabled for tree-sitter AST
$ CGO_ENABLED=1 go install github.com/iSundram/Automergent/cmd/automergent@latest
$ git clone https://github.com/iSundram/Automergent.git
$ cd Automergent && make build && sudo make install
Built from the ground up for software engineers who require precision, transparency, safety, and sub-millisecond responsiveness.
Powered by Charm v2 (bubbletea/v2, lipgloss/v2). Works seamlessly in standard terminals, tmux, and SSH sessions with instant startup and smooth rendering.
Native first-class support for Google AI Studio & Vertex AI, plus any OpenAI- or Gemini-compatible endpoint (Anthropic, DeepSeek, Ollama, vLLM) with automated community model catalog sync.
Comprehensive toolset spanning atomic file edits, persistent PTY shell processes, tree-sitter AST symbol queries, live web documentation scraping, secret scanning, and SQL execution.
Spawn concurrent specialist subagents (Explore, Review, Build, Contexter) with isolated scratchpads, background execution, live steering notifications, and autonomous goal workflows.
Adaptive token estimation, AST file dependency tracking, proactive auto-compaction, session checkpoint rewind, and persistent graph-backed memory stores.
Hot-swappable 16-slot color palettes (Modern, Catppuccin, TokyoNight, Dracula, Nord, Gruvbox, OneDark, Monokai) with syntax highlighting and diff viewers.
Automergent processes complex requests with a disciplined four-phase pipeline that eliminates hallucinations and ensures verifiable correctness.
Decomposes user prompts into atomic intents. Selects optimal execution strategy (Direct / Explore / Plan / Build) and initial token budget.
Tree-sitter AST discovery, symbol indexing, ripgrep matching, and issue confirmation without risk of unwanted modifications.
Constructs a structured execution plan with verification milestones, file boundaries, risk mitigation, and interactive approval gates.
Performs atomic edits, dispatches parallel subagents, compiles code, runs test suites, and executes automated compiler recovery.
Clean, declarative configuration with hierarchical precedence: Defaults < Global < Project < Profile < Env < Session < CLI Flags.
# AI Provider & Model Selection
provider: "google-aistudio" # google-aistudio | google-vertex | custom
model: "gemini-3.6-flash" # default model
temperature: 0.2
mode: "accept-edits" # manual | accept-edits | auto | plan
theme: "modern" # modern | catppuccin | tokyonight | dracula | nord ...
# Provider API Credentials
providers:
google-aistudio:
apiKey: "${GOOGLE_API_KEY}"
google-vertex:
project: "my-gcp-project-id"
location: "us-central1"
custom:
baseUrl: "https://api.openai.com/v1"
apiKey: "${CUSTOM_API_KEY}"
defaultModel: "gpt-4o"
# Context Window Management & Auto-Compaction
maxContextTokens: 1048576 # 1 Million token budget
autoCompressAt: 0.55 # compact context when 55% full
maxToolOutputChars: 32768
# Security Sandboxing & Safety Guards
sandbox: auto # auto | bubblewrap | seatbelt | none
blockedWritePaths:
- "/etc/**"
- "/usr/**"
allowedWritePaths:
- "."