For the multi-agent mess
Replace the pile of terminals, tabs, screenshots, copied stack traces, and setup guesses with one workspace built for parallel agent sessions.
Local-first AI agent workstation
The control room for AI coding agents.
Keep agents, shells, localhost previews, routed error context, and explicit permissions in one native workspace. Your tools stay yours. The control surface stays local.
TokenburnerProject: northstarBranch: feat/navigation
Local workspace Send-to-AI walkthroughFailure detected locally
v0.5.2 current releaseBYO-AI use your accountsWindow capture shipsLocal no desktop telemetry
The local loop
From a failing test to the right agent, without losing the room.
Tokenburner keeps the work visible, makes routing explicit, and leaves consequential actions under your control.
- 01 / Watch
Keep the local work visible.
Run AI CLIs, regular shells, prompt pads, localhost web panes, and direct provider chat panes in one native split workspace.
- 02 / Catch
Detect failures where they happen.
PTY output is scanned locally for error patterns so broken test runs and stack traces surface without tab hunting.
- 03 / Route
Send the right context to the right agent.
Route captured window context and terminal snippets to a chosen AI pane. Pane and region capture are planned, not shipped.
- 04 / Govern
Let agents coordinate with explicit grants.
A loopback MCP server lets agents list panes, read permitted context, send messages, spawn panes, and call granted tools with an audit trail. The Settings UI now includes paste-ready MCP config snippets and a connection doctor.
- 05 / Cost
See cost where Tokenburner records it.
Direct Chat panes get local cost estimates for completed replies. Terminal tracking attaches only when a pane launches claude, codex, or openai, and records cost only if that command prints a usage line Tokenburner recognizes. Current Claude Code and Codex CLI output does not include a recognized usage line, so those panes record no cost. Codex Agent panes are not recorded yet.
Current capture boundary: window screenshot capture ships today. Pane and region capture remain planned.
Built around your tools
One workspace for the pieces you already use.
Tokenburner is the local cockpit between AI CLIs, logs, previews, screenshots, and permissions.
For local feedback loops
Error badges and routing shortcuts turn failing output into a directed debugging prompt without sending the whole project away.
For agent control
Agents can coordinate through Tokenburner tools only after explicit local grants. The app now helps generate MCP config snippets instead of making users spelunk dotfiles.
For people who already chose their tools
Keep your editor and your model accounts. Tokenburner is the workstation around them, not another editor or model provider.
Clear boundaries
The workstation stays narrow on purpose.
Local control works only when the product is direct about what it does, what it does not do, and which paths are still incomplete.
Not a cloud agent platform
No hosted shells, no remote sandboxes, no background cloud workers.
Not a code editor
You keep Cursor, Zed, VS Code, JetBrains, Vim, or whatever already works.
Not model billing
You pay AI providers directly. Tokenburner shows local cost estimates for supported panes; it does not meter inference.
Not fully done
Pane/region capture, budget kill switches, automated MCP config editing, and the fully verified self-serve paid path are still planned.
Privacy posture
Sensitive work stays visible to you.
Terminal output, screenshots, local cost records, file paths, and project metadata are treated as sensitive workspace data.
PTY content stays local
Terminal output is used inside the app for rendering, error detection, and optional local diagnostics.
Screenshots stay local
Screenshot bytes remain on the machine unless the user deliberately shares or routes them.
Cost data stays local
Token and dollar estimates are local app data, not remote analytics. The cost history CSV export (Pro or trial in this beta) saves a local file; it is not telemetry.
MCP is loopback only
The MCP server binds to localhost. Every MCP tool call must match an active per-agent tool-and-scope grant. A matching grant authorizes repeated calls while active; new grants default to 60 minutes. Clearing the expiry prevents a timeout, but the grant is still dropped at the next app launch unless you enable restart persistence.
No telemetry SDK
The desktop app has no analytics package or hidden phone-home path.
Update checks are explicit
The app checks dl.tokenburner.net at startup by default unless you disable it. The check includes the app version, OS target, and architecture. Like any web request, it also gives Cloudflare the IP address, user-agent, and timestamp; on Tokenburner's current Cloudflare Free plan, Cloudflare retains Worker request logs for 3 days. GitHub receives ordinary request metadata when a release-bearing check follows the signed-manifest redirect. If you accept the update, the archive request reaches Cloudflare before redirecting to GitHub, so both receive ordinary request metadata for the download.
License keys verify locally
The app can activate minisign-signed license keys without phoning home. Checkout and license delivery run through separate Cloudflare Worker routes.
Current status
Useful now. Honest about the gaps.
The latest release is v0.5.2. It includes guided setup, New Project Wizard v2, MCP connection guidance, agent health checks, cost model v2, local license activation, paid-path plumbing, and Windows release assets. Mainline work is hardening the local loop and release path.