Why your code should never leave your machine
Local-first isn't a privacy nicety bolted onto an agent tool — it's the whole architecture. Here's what "your code never leaves your box" actually buys you.
Every “agentic” coding tool on the market starts from the same hidden assumption: to let an agent work on your code, you first have to give your code to the agent’s cloud. Upload the repo, mount the workspace, stream the diffs — and trust that whatever runs over there is doing only what it says.
Tandem starts from the opposite assumption. The agent comes to your machine. Your code stays put.
One inbox for every model — locally
The pitch is simple: one inbox for every model. Claude, local LLMs running on your own GPU, or any provider you bring a key for — all of them route into the same review surface. But the part that matters architecturally is where the work happens:
- Prompts and completions flow directly from your device to whichever provider you chose.
- Generated code lands in a local git worktree, not a vendor sandbox.
- The orchestration layer — task graph, approval inbox, audit trail — is a process on your box.
Nothing about that pipeline requires Tandem’s servers. In the default install, zero bytes of your code leave the machine.
The moat is the architecture
It’s easy to add a “privacy mode” toggle to a cloud product. It’s very hard to retrofit local-first onto one. That asymmetry is the moat:
- Spawn a fleet of agents in parallel — one prompt, N isolated worktrees.
- Review every agent’s diff from a single inbox; comment, approve, or reject.
- Merge what you ratify. Plans route through a human approval gate before anything runs.
Each of those steps is a local operation. The audit trail is append-only and lives in your vault as plain markdown — readable in Obsidian, greppable from your shell, yours forever.
Zero telemetry is a number, not a vibe
We publish it as a stat because it’s measurable: telemetry — 0 bytes. No crash uploads, no usage beacons, no “anonymous” diagnostics phoning home by default. If you opt into a hosted workspace later, that’s an explicit choice with its own encryption and a deliberately tiny sub-processor list — not the default you have to discover and turn off.
Pay for the app, not for the inference. No token markup, no per-seat surveillance.
That’s the whole bet: a real fleet of agents, running on your hardware, under your review — and your code never leaving the box.
Want to see it move? Go Pro or read the feature breakdown.