Launching today
Tempest
Run AI agents in parallel with 64% fewer tokens
15 followers
Run AI agents in parallel with 64% fewer tokens
15 followers
You run multiple AI coding agents, but they all waste tokens understanding the same repository. Tempest fixes this. Index your repository once, then run every agent in its own isolated workspace while sharing the same understanding. Up to 64% fewer tokens for any agent.
Hello Product Hunt Community! 👋
Have you ever tried running multiple AI coding agents in parallel, only to realize every agent reads the same repository before it even starts working?
Frustrating, right? 😤 That's exactly why we built Tempest.
I'm Prharsha, founder of Tempest. And I believe we're optimizing for the wrong thing.
Instead of making every agent rediscover your codebase, Tempest indexes your repository once, so every agent shares that understanding while working in its own isolated workspace. Every agent gets its own space to work safely without stepping on another agent's changes, while using up to 64% fewer tokens.
🚀 What is Tempest?
Tempest is an open-source agent orchestrator for AI coding agents. It supports Claude Code, Gemini CLI, Codex, OpenCode, GitHub Copilot, Cline, Goose, and more. It runs natively on Windows, macOS, and Linux, so you can use the same workflow regardless of your development environment.
💡 Why we built Tempest
Most agent orchestrators are obsessed with one metric:
"How many agents can we run?"
We think that's the wrong goal.
The future of agentic engineering isn't about running 100 agents. It's about getting the most work done for the least cost.
If 100 agents each spend thousands of tokens rediscovering the same repository, you've built an expensive system...not an efficient one.
We'd rather run 50 well-informed agents that share repository understanding, work in isolated workspaces, and finish the same amount of work using far fewer tokens.
We don't think the future is more agents. We think the future is smarter agents.
What can you do with Tempest?
⚡ Run multiple AI coding agents in parallel.
🧠 Share repository understanding across every agent to reduce token usage by up to 64%.
🌿 Give every agent its own isolated workspace so they never step on each other's changes.
💬 Comment on code and send your feedback directly back to an agent.
🌐 Preview your app without leaving Tempest.
💻 Works on Windows, macOS, and Linux.
We're building Tempest in the open, and we'd genuinely love your feedback. Whether you agree with our vision or think we're completely wrong, we'd love to hear your thoughts. Every conversation helps us build a better product.
Thanks for checking out Tempest and supporting open source! ❤️
A specific number like 64% is more interesting than "fewer tokens" would be, so the question is what it's measured against — same task, same model, same success rate? Token reduction is easy if you're allowed to let quality slip.
And where does the saving come from? Shared context between parallel agents, more aggressive pruning, or not re-sending the same files each turn? Those have different failure modes: pruning is where I'd expect an agent to quietly lose the constraint you gave it forty turns ago.
@ark_y_k Hey Yuki! Thanks for the question
The 64% comes from running the same coding tasks with the same models and comparing Tempest against the baseline where each parallel agent independently loads repository context. We measured token consumption while keeping task quality the same in our testing.
The savings come from a shared knowledge graph. Tempest indexes the repository once into a shared knowledge graph that every coding agent can use via a MCP, instead of each agent repeatedly rebuilding repository understanding from scratch. That eliminates unnecessary file reads while keeping every agent grounded in the same codebase. We're also validating quality parity on harder tasks to make sure the savings don't come at the cost of missing subtle dependencies.
Everything stays local. Tempest is completely open source