Your autonomous second brain that plans, researches, edits code, and executes system commands.
A local-first, streaming-first desktop AI assistant with 13 specialized tools, 4 LLM providers, a visual workflow builder, an autonomous browser agent, and a full code editor — orchestrated by a persistent Central Intelligence loop.
At the heart of Contragraviton is a persistent agentic reasoning loop built on LlamaIndex AgentWorkflow. Give it a complex goal and it will plan, execute, self-correct, and persist through failures — up to 100 iterations deep.
Breaks your request into a multi-step execution checklist with loop-prevention rules and progress tracking.
Selects the best tool from 13 options — grep first for precision, then semantic search, then web fallback.
If a tool return is insufficient, it modifies search terms, pivots strategy, or escalates to a different approach.
Synthesizes all gathered information into a coherent answer or action — streamed live via SSE with full thought visibility.
A hybrid retrieval engine combining a PARA-structured filesystem, a SQLite brain index (brain.db), and a ChromaDB vector store with all-MiniLM-L6-v2 embeddings. Six search strategies cascade from SQL queries to grep to semantic search to live web — so nothing is ever lost.
Built on React Flow (@xyflow/react), the visual workflow canvas lets you drag nodes, connect data flows, and build intelligent automations. Wire a Web Search node into an LLM Brain, pipe results to a Display — and watch it execute in real-time with live data flowing between nodes.
Every building block you need: Web Search, LLM Brain, Code Execution, Display, Discord, Slack, Email, Webhooks, and more. Design once. Run anytime.
A separate agentic pipeline powered by Playwright Chromium with its own AgentWorkflow, dedicated SSE stream, and 8 browser tools — browser_navigate, browser_click, browser_type, browser_screenshot, browser_scroll, and more. Headless or headed mode. Screenshots after every action.
Goes to any URL via page.goto(), follows links, handles redirects
Clicks elements by CSS selector, fills forms, interacts with dropdowns
Captures PNG after every action. Live mode polling via /screenshot/live
Discovers all clickable/typeable DOM elements, sorted top-to-bottom
A split-pane IDE with Monaco Editor, file tree, and xterm.js terminal. The agent uses replace_file_content for surgical single-block edits and multi_replace_file_content for complex multi-point refactors. run_shell_command executes builds in the background while command_status monitors progress.
"Create a Flutter to-do list app and build the APK."
120+ files generated. APK compiled. README written. One prompt.Google Gemini, OpenAI, OpenRouter, or LM Studio (local). Modular provider pipeline with @register_provider decorator — zero if provider == branching. Switch mid-conversation, takes effect on next message.
Every reasoning step, tool call, and result streams live via Server-Sent Events through a producer/consumer asyncio.Queue. The Thoughts Accordion shows Tool Calls, Results, Reasoning, and Approval gates in real-time.
SQLite brain.db, ChromaDB vectors, thread JSON files, and your PARA-structured knowledgebase — all on your machine. No cloud storage. No telemetry. Your data never leaves your device.
File writes and shell commands trigger InputRequiredEvent for explicit approval. The agent cannot execute destructive actions without your confirmation in the UI.
Native desktop app via Tauri 2 (Rust + WebView2) or browser SPA served by FastAPI. Same React 19 + TypeScript codebase, same API, same Zustand stores. Google OAuth built in.
The generate_image tool creates visual assets using Flux-scale lightning models. Diagrams, UI mockups, creative assets — generated directly inside the agent's reasoning loop.
grep_knowledgebase for exact substring matches. search_knowledgebase for semantic vector similarity via ChromaDB. read_file_range for surgical line-level zoom. Six strategies, one goal: find it fast.
search_web queries DuckDuckGo for real-time data. read_url_content ingests and summarizes external articles. Toggle web search mode to force prioritization.
run_shell_command executes PowerShell with background support. command_status monitors long-running tasks. Deploy, install, build, test — all from the chat.
list_knowledgebase_filesExplore files and folders with auto-discoveryread_knowledgebase_fileDeep-dive into specific documentswrite_knowledgebase_filePersist notes, logs, and project filesmove_knowledgebase_itemReorganize your second braingrep_knowledgebaseHigh-speed substring search across all filesread_file_rangeSurgical line-range extraction (zoom-context)replace_file_contentTargeted single-block code/text editmulti_replace_file_contentComplex multi-point edits in one passsearch_webLive internet queries via DuckDuckGoread_url_contentIngest and summarize external articlesgenerate_imageVisual assets via Flux lightning modelsrun_shell_commandPowerShell execution with background supportcommand_statusMonitor long-running background tasks
Contragraviton is open source and ready to run on your machine.
Get Started on GitHubA one-click installer for Windows, macOS, and Linux is in active development. Download, launch, and start building — no Python, no terminal, no dependencies.
Follow me on LinkedIn for updates and early access:
Follow on LinkedIn