Wren runs 2–8 specialized AI agents on your own machine — Rust backend, security, testing, Python — coordinated through a single terminal. Your code never leaves. You approve every action.
Built for Web3 & crypto devs security researchers NDA-bound freelancers
// Why Wren
Cursor, Windsurf, and GitHub Copilot are brilliant tools — but they route your codebase through external servers. For fintech, Web3, defence, and biotech engineers, that's a legal problem. Wren runs entirely locally. Zero API calls required.
Agents don't act autonomously. Every proposed action enters an approval queue — you see the intent, the files to be changed, the risk level. You approve. The AWAITING_APPROVAL state is built into the fleet protocol.
Wren dispatches work to purpose-built terminals: T2 (Rust/backend), T4 (Python), T5 (testing), T6 (security). Each terminal knows its domain. Start with 2–3 agents on a workflow template. Scale to 8 as you trust the system.
// Built for
Cursor is excellent software. It's just illegal for a lot of people to use at work.
Smart contract code is your IP. One leaked private key, one exposed algorithm, and the damage is irreversible. Cloud-based AI tools aren't an option when the stakes are on-chain.
You work with zero-days, exploit PoCs, and client vulnerability reports. Uploading any of that to an external API is a breach of contract, ethics, and common sense.
Client contracts prohibit externalising codebases. You're locked out of the AI tooling that makes you competitive — unless the AI runs on your machine.
// How it works
Wren is a coordination surface, not an autonomous agent. You direct. The fleet executes. You approve. Everything is local.
Type a natural-language goal into Oracle. Wren maps it to the right terminals in the fleet and builds an execution plan.
Specialist terminals pick up their slice of the work. Security audits, unit tests, and backend fixes run simultaneously — all on your Mac, via Ollama.
Before any file is touched, the Agent Dashboard surfaces the intent as a diff. AWAITING_APPROVAL. You see the patch. You approve or reject. No surprises.
The evolution engine scans your codebase while you sleep, files its own improvement missions, and dispatches patches. Your build number increments every morning.
// Comparison
| Feature | Wren | Cursor | GitHub Copilot | Devin |
|---|---|---|---|---|
| Local inference (no API calls required) | ✓ Yes | ✗ Cloud only | ✗ Cloud only | ✗ Cloud only |
| Zero telemetry / zero data exfil | ✓ Verified in HUD | ✗ | ✗ | ✗ |
| Approval gate before execution | ✓ Built-in | ✗ Autocomplete | Partial | ✗ Autonomous |
| Multi-agent fleet (specialized terminals) | ✓ 2–8 agents | ✗ Single agent | Partial | ✓ Autonomous only |
| Usable under NDA / compliance policy | ✓ Yes | ✗ Usually not | ✗ Usually not | ✗ No |
| Self-improving overnight (evolution engine) | ✓ Yes | ✗ | ✗ | ✗ |
// Pricing
You own the AI inference. We charge for the orchestration layer.
// Accelerator — May 2026
We're raising a post-money SAFE to bring Wren to the 300,000+ compliance-blocked engineers in NYC fintech, defence tech, and Web3. If you believe the future of developer tooling is local-first and sovereignty-preserving, we'd love to build it with you.