Pipali: The Open-Source Desktop AI Coworker From Khoj AI's YC Team
Most "AI coworkers" still live inside a chat window. You paste, they answer, you copy. The work — the actual files, the actual browser, the actual Linear ticket — never gets touched without you doing the touching. Pipali, the new beta from Y Combinator–backed Khoj AI, is one of a small handful of tools trying to fix that gap, and it does it with an Apache 2.0 license and a Tauri shell that runs on macOS, Windows, and Linux.
The pitch is simple enough that it almost sounds boring: let the model use your computer the same way you do. In practice, that means Pipali reads and writes files in a sandbox, drives your browser, runs commands, and talks to your SaaS stack through MCP — Anthropic's Model Context Protocol. The interesting part is that Khoj AI shipped the whole thing as open source, and let you bring whichever frontier model you trust.
What Pipali actually is
Pipali is a desktop app, not a hosted product. The repo at khoj-ai/pipali ships a Tauri shell with a TypeScript-heavy codebase (87% TypeScript, 2.3% Python, 1.9% Rust by GitHub's language stats) and is licensed under Apache 2.0. The latest release on the day of writing is v0.5.2, tagged May 12, 2026, and the repo currently sits at 134 stars and 30 forks with 494 commits in main.
The agent has four core competencies, in Khoj's words:
| Capability | What it does |
|---|---|
| Research | Reads your local files and the live web, summarizes, cross-references |
| Create | Produces briefs, decision memos, spreadsheets, emails, personal micro-apps |
| Automate | Runs scheduled or event-triggered routines (weekly reports, mail triage) |
| Integrate | Talks to Linear, Jira, Slack, GitHub and other tools through MCP |
It is, by design, not another wrapper. The app gives the model a sandboxed shell, a file system handle, and a browser driver. You watch the work happen in a side panel that streams Pipali's plan, asks for permission on dangerous actions, and lets you intervene mid-task.
Bring-your-own-model — but you don't have to
A nice touch: Pipali ships with model access through the Pipali Platform via single sign-on. No API keys to provision before the first run. The maker, Debanjum, confirmed on Product Hunt that the platform routes requests to Claude, GPT, Kimi, GLM, DeepSeek, Minimax, and Gemini. If you'd rather pay providers directly, the architecture is open enough that you can wire your own keys in.
"We built Pipali because we wanted AI to move beyond chat — not just answer questions, but actually operate your computer with you and finish useful work." — Debanjum, Khoj AI
That choice matters because the agent's quality is upstream of Pipali. Pipali is the harness; the brain is whichever model you point at it. The same task delegated to Claude Opus and to a cheaper Kimi K2.6 will produce very different results — and Pipali lets you choose per-task, which is the right design.
Skills and Routines: where it gets sticky
The two features that separate Pipali from the long list of "Claude on your desktop" wrappers are Skills and Routines.
Skills are reusable, scoped instructions you teach Pipali once and reuse forever. The repo's README lists examples like "which accounting method to follow", "where to find project documents", and "my email organization policy." Skills are Pipali's answer to the prompt-engineering tax — instead of re-explaining your context every chat, you bake it into a named procedure the agent loads on demand. (Khoj points users to the public agentskills.io registry for community-shared skills.)
Routines are scheduled or event-triggered runs. Examples from the README:
- "Draft my weekly project update email"
- "Sync my ledger on the 1st of every month"
- "Mark all marketing emails as spam"
Combine the two and Pipali starts to look less like a chatbot and more like the script you've been meaning to write but never quite did.
The sandbox is the whole point
Letting a model run commands on your laptop is the kind of idea that makes anyone with a git push --force scar in their muscle memory tense up. Khoj's answer is a local sandbox that restricts file and network access by default, plus an explicit approval step for anything that escapes those bounds.
The README puts it plainly: "Commands that need broader access require your explicit approval. You can configure these permissions yourself." It's not a novel pattern — Claude Code, Goose, and Codex all do versions of this — but having it open-source means you can actually read the sandbox boundary instead of trusting the marketing page.
The repo also includes a SECURITY.md with a vulnerability disclosure process, and the drizzle/ directory hints at a local SQL store rather than a hosted database. Conservative defaults, and the kind you want.
How it stacks up
Pipali is launching into a crowded space. The obvious neighbors:
- OpenClaw / OpenYak / Clawdi — the Claw-family of desktop agents that exploded after the OpenClaw rebrand drama earlier this month.
- Goose by Block, which crossed 35K stars and emphasizes local-first execution.
- Hermes Agent and Raindrop Workshop, both also riding the "agent that learns your work" angle.
Pipali's differentiator isn't raw capability — it's the combination of open-source, multi-model, MCP-native, and YC-backed. Most of the alternatives pick two of those. Khoj is gambling that being honest about all four buys them developer trust faster than a closed-source incumbent can buy adoption.
Getting it running
The path to first run is short:
- Grab the installer from pipali.ai/download for macOS, Windows, or Linux.
- Sign in from the desktop app.
- Hand it a starter prompt — the README suggests "Summarize the last 5 PDFs in my Downloads folder into a professional 1-page brief" as a smoke test.
If you'd rather build from source, the repo uses Bun as the package manager (bun.lock is checked in), Tauri for the shell, and Drizzle for local persistence. A CONTRIBUTING.md and Playwright test config are in the tree, which is more setup discipline than most agent projects ship with in their first month.
What's missing
Pipali is beta. The version number — 0.5.2 — is the company telling you so out loud. A few caveats worth keeping in mind:
- Star count is still small (134 at time of writing), so community-contributed Skills will be thin until the catalog grows.
- Billing is gated through the Pipali Platform if you use the bundled model access. Individual users and team admins need to set up billing before the initial signup credits run out.
- Browser control on Linux tends to be the rough edge in this class of tool. Worth testing before committing a workflow.
The Bottom Line
Pipali is one of the cleanest open-source bets on the "AI coworker on your computer" thesis that 2026 has produced. It's not the most starred project in the space and it's not the most polished, but it's the one that picked the most defensible combination: Apache 2.0, multi-model, MCP-native, sandboxed by default, and YC-backed. If you've been waiting for a Claude-Code-style agent that wasn't tied to one vendor, this is the project to bookmark and watch through the next few releases. The repo is at github.com/khoj-ai/pipali — clone it, run it for a week, and see whether Skills and Routines actually save you the hours Khoj is promising.
