An open-source desktop environment for designing and running autonomous multi-agent teams locally with Ollama. Built in Rust & Tauri v2 with an interactive node-graph designer, SQLite WAL event log, and human-in-the-loop diff approval.
Hey Product Hunt! 👋
I’m Abhay, a student developer passionate about systems programming, distributed systems, and local AI.
Over the past several months, I’ve been building Trans4mers as an open-source research project to explore what a desktop-native environment for local multi-agent systems looks like.
### Why I Built Trans4mers
I wanted to work with multi-agent swarms on my local workstation without needing a browser-based SaaS or command-line scripts. I chose to build a native desktop application in Rust and Tauri v2 to keep the memory footprint modest (~78 MB idle RAM) and explore how classical systems engineering patterns apply to autonomous agents:
- Visual Swarm Designer: A node-graph canvas to configure supervisor-worker, adversarial debate, and fan-out topologies.
- Hardware-Bounded Concurrency: Tokio semaphores manage execution permits so multiple background workers don't cause thermal throttling or memory exhaustion on consumer hardware.
- Event-Sourced SQLite WAL: Every domain event is written to an append-only log in an embedded SQLite database (trans4mers.sqlite) before state projections update tables, enabling state inspection and startup replay if interrupted.
- Human-in-the-Loop Diff Gate: High-risk actions (modifying files outside scratch, executing shell scripts) produce an LCS line-level diff with SHA-256 argument verification, pausing execution for operator review.
- 4-Tier Cognitive Memory: Combines working sliding context, episodic task history in SQLite, hybrid semantic search (LanceDB vectors + SQLite FTS5 BM25 fused with Reciprocal Rank Fusion), and procedural learned rules.
- Local-First Inference: Runs directly against local Ollama daemons, with optional BYOK support for frontier models (Anthropic, OpenAI, Google) stored in OS credential vaults.
### Brutal Honesty & Current Project State
- Solo Student Alpha: This is an early-stage alpha project built solo while studying.
- Bugs Expected: Because it's actively evolving and hasn't been tested across thousands of hardware setups, there will be unhandled edge cases, UI quirks, and bugs.
- Security Caution: While the capability system and diff reviewer exist in code, the application has NOT undergone third-party commercial security audits. Treat autonomous shell execution with caution and test in sandboxed or backed-up folders.
The project is 100% open-source under MIT:
👉 GitHub: https://github.com/abhayzangir1/...
💖 Support a student builder: https://github.com/sponsors/abha...
I would love your feedback, bug reports, and thoughts on the architecture!