We let developers run more agents while keeping your token bill low. We train custom models to determine how we can maintain performance while reducing your token bill.
Integrate into your own apps, or use our router with Claude Code, Codex, or any agent of your choice.
Active Founders
Avyay Varadarajan
Founder/CEO
Building dari.dev. Previously CS @ Caltech, SWE @ Uber, ML @ Lakers, startups.
Avyay Varadarajan
Founder/CEO
Building dari.dev. Previously CS @ Caltech, SWE @ Uber, ML @ Lakers, startups.
Benjamin Hong
Founder/CTO
Building dari.dev
Benjamin Hong
Founder/CTO
Building dari.dev
Company Launches
dari.dev - The easiest way to deploy your agents to production
dari.dev turns an agent project into a versioned, stateful API endpoint. We’re building on top of the best open-source harnesses starting with Pi. Support coming soon for Hermes Agent. Book a call to deploy your proprietary harness.
How it works:
Drop a dari.yml + tools + skills at your repo root
“dari deploy .” uploads a snapshot and publishes a new version behind a stable agent URL
Clients open a durable session and stream multimodal events (text, files, images)
The platform runs the harness, spins up a sandbox workspace per session, and handles logs, retries, and credentials
New publishes create new versions. In-flight sessions stay pinned to the version they started on
The problem:
If you're building an agent right now, you're probably:
Stitching together durable execution + a sandbox provider + Postgres + a secret store yourself
Losing state when a turn crashes mid-execution
Shipping “agents” that can’t survive a 30-minute task
Rebuilding the same infra (auth, retries, workspace, logs) on every project
Durable sessions: Long-running tasks survive crashes, restarts, and redeploys. Resume from the same session_id days/months/years later.
Versioned deploys: Every publish is a new version behind one stable URL. Roll forward without breaking live sessions.
Bring your own sandbox: Sandboxes run under your provider account and API key. You control limits and data. E2B today, Daytona / Blaxel / Modal coming soon.
BYOK with OpenRouter / any OpenAI-compatible API endpoint: Use your own key, pick your model, control spend, keep logs.
Custom tools: Drop Python or TypeScript tools under tools/<name>/. We run them with retry policies you define.
The API:
Publish once. Then, create a session and send messages + stream responses.
We provide full observability, session-level event logs, per-turn state, and workspace file outputs attached to the session lifecycle.
Why we built this
We started building browser automation agents and kept hitting the same missing primitive, a durable control plane underneath.
We built it for ourselves, then realized every other team shipping agents (coding, research, ops) is rebuilding the same layer.
If you’ve wired together durable execution + a sandbox provider + Postgres + a secret store just to run one agent in prod, this is for you.
Install the CLI, run dari auth loginthen publish from your repo
Pricing: Either you pay your providers directly (E2B + OpenRouter/OpenAI) or use our managed Anthropic/OpenAI/OpenRouter offerings. No markup from us initially.