Download
Download Tepegöz.
Signed builds for Windows, macOS and Linux. No account, no telemetry, no backend — add your own AI key and it works.
Downloads
Pick your platform.
| Platform | Format |
|---|---|
| Windows | Installer (.exe) |
| macOS | Disk image (.dmg) |
| Linux | .deb · .rpm · .tar.gz |
Windows 11 is the primary target. macOS and Linux build and pass the full test suite on every push, but receive less hands-on testing.
Integrity
Verify what you downloaded.
Every release publishes a checksum alongside the binaries. Builds are code-signed on Windows and notarized on macOS, so your operating system will not warn you — but the signature tells you the file came from us, and the checksum tells you it arrived intact. Both are worth thirty seconds.
If you would rather not run a binary at all, build from source — the instructions below produce the same application from code you can read.
Then add a key
The browser works without one. The agent needs a model.
Open Settings and add a key from Anthropic, OpenAI, Google or Kimi — or point it at a local model and run entirely offline.
Your key is encrypted through your operating system's keychain and stays in the privileged process. You pay your provider directly; there is no Tepegöz account and nothing is proxied through us.
Build from source
Three commands.
Five minutes, and genuinely simple — there is no compiler step and no native database to rebuild.
- Node.js 24 or newer — the same runtime Electron 43 embeds, so the app and its tests run on identical ground
- pnpm 10 or newer
Terminal
git clone https://github.com/kuraykaraaslan/tepegoz-browser.git
cd tepegoz-browser
pnpm install --frozen-lockfile
pnpm devThat is the whole thing. No build tools, no Python, no C++ toolchain, no database to compile.
What you are getting
Pre-release software.
Things will break, and two specific gaps are worth knowing before you trust it with anything that matters.
- No independent security audit has been performed. The threat model is published and the architecture is readable, but no outside party has reviewed it.
- The automation has not been independently benchmarked. The adversarial battery and the head-to-head comparison are written and pre-registered; the runs have not been paid for, so there is no measured attack-success-rate or task-success number, and we will not quote one until there is.
What you can rely on: it is AGPL-3.0, the entire source is public, there is no account, no telemetry and no backend — and the problems we already know about are written down instead of discovered by you.
Known issues are published at known issues. Phase-by-phase status is at the roadmap.