One switchboard for every AI workflow.

Route models, track plan usage, and supervise coding agents from one local control plane.

Explore the product Built for macOS · Runs on your machine
⌁ SwitchBoard 127.0.0.1:8420

Overview

Routing, usage, and Studio in one place.

INTERACTIVE PREVIEW
Plan usageIllustrative
Claude Code
exact
OpenAI API
observed
Gemini API
estimated
5h · 7d · monthly windows, with reset times
Routing profileGlobal
Your toolclient key
→
SwitchBoardloopback gateway
→
AnthropicOpenAIOllama
Explicit 429 fallback · provider caps
Studio · local agent runDemonstration
⌘
Plan → approve → executeLocal git worktree · gateway agent + ACP agent
Awaiting approval
01 Dispatch requires a human decision before the ACP task starts.
⇄

Route

One OpenAI and Anthropic compatible endpoint for local tools. Set provider and model by scope, with explicit fallback rules.

How routing works ↗
▥

Meter

See plan windows and request usage together. Every signal says whether it is exact, observed, or estimated.

Explore usage ↗
⬡

Run

Keep coding workflows local, with git worktrees, checkpoints, an audit trail, and human approval gates.

Meet Studio ↗
01 / THE CONTROL PLANE

Your tools keep their workflow.
SwitchBoard handles the route.

Connect an editor, CLI, or agent to a single loopback endpoint. Each client has its own key; your upstream credentials stay in macOS Keychain.

LOCAL CLIENTSEditorCLIAgent
→
⌁
SwitchBoard gateway127.0.0.1:8420
Scoped routingUsage ledgerApproval gates
→
YOUR PROVIDERSCloud APIsLocal modelsVendor CLIs via ACP exploratory
The management API stays on a private Unix socket. Upstream keys are held in gateway memory only while it runs.
02 / USAGE INTELLIGENCE

Know the window
before you hit it.

SwitchBoard brings five-hour, seven-day, weekly, monthly, and request windows into one view when a source provides them. Thresholds at 50, 75, and 90 percent help you plan the next route.

Exact · source reports the window Observed · measured request data Estimated · calculated from available data
Usage overviewILLUSTRATIVE DATA
Claude CodeFive-hour window
72%
OpenAI APIMonthly spend estimate
38%
Gemini APIRequest usage
54%
Source fidelity and reset time are shown per window.
03 / SUPERVISED AGENTS

Give agents a workspace.
Keep the decision.

Studio is being built around local git projects. A run can pause at dispatch or tool approval, then resume from a checkpoint after your decision.

01Choose a project

Work begins in an isolated local git worktree.

02Route the roles

Gateway and ACP agents can share a supervised run.

03Review the action

Approve or reject at the gate; inspect the run journal.

Current foundation: a local Studio workflow and sidecar are in the source build. Broader agent orchestration and cloud runners are roadmap work.

04 / CONNECTIONS

Bring your own models.

The current gateway supports API keys and local Ollama. Vendor CLI connections via ACP are exploratory; sanctioned third-party OAuth awaits provider support.

API keys · foundation

AnthropicOpenAIxAIGemini AI StudioMoonshot / KimiDashScope / QwenNVIDIA buildOpenRouter

On your machine

OllamaOpenAI compatible toolsAnthropic compatible toolsClaude Code statusLine

Connection paths being explored

Claude Code ACPCodex ACPGemini CLI ACPGitHub Copilot ACPVendor sanctioned OAuth
05 / BUILD STATUS

Building steadily.
Honest about the edges.

SwitchBoard is an early macOS build. The local gateway, routing, usage ledger, menu bar preview, and supervised Studio foundation are present. Distribution still needs a signed, notarized app and a bundled runtime.

Foundation

Loopback gateway, API key and local-model routes, SQLite usage ledger, macOS menu bar app.

IN SOURCE
Desktop experience

Settings, usage views, Studio workflow, durable configuration, release packaging.

IN PROGRESS
Expanded connections

WidgetKit, sanctioned OAuth, richer ACP adapters, cloud runners, signed distribution.

PLANNED

Make every model
work on your terms.

See what is available today and what comes next.

See build status