One local hub for Claude / Cursor / Codex and more: onboard, trace, and route in one place—without changing how your tools work. Resources grow from your habits; idle capacity can earn credits.
macOS · Windows · Linux · Apache-2.0 open source
Works with the agents you already use
TRACE · ANALYTICS
Onboard agents; full trace of route, model, tokens, cost. Multi-device inventory with subscriptions vs PAYG side by side.
SMART ROUTING
Keep native model names; route to local or cheaper supply. Local-first chain, scene strategies, optional lossless compression.
ONE-CLICK HUB
One-click onboard/restore; multi-account CLI by directory; tray for status. Point tools at one local address.
PORTRAIT · FOR YOU
Mine a work portrait from real habits; discover, accumulate, and iterate MCP / Skill / Prompt / Agent that fit you.
COMMUNITY SHARING
Contribute idle capacity to community sharing for credits; circles and a network map; spend credits on shared models.
v0.5 stitches onboarded agents, multi-account CLIs, and community resources into one local control plane, without changing how your tools talk to models.
Set a main agent as the aggregation entry. It receives the task, plans steps, and dispatches to other onboarded agents. Tool streams stay visible; stop and resume when you need to.
Scan or add Claude / Codex instances. Dispatch by working directory so each project keeps its own account, quota, and route.
Mine a work portrait from real habits; discover, accumulate, and iterate MCP / Skill / Prompt / Agent that fit you.
Community catalog plus personalized recommendations; project onto runtimes; prompts via tokenbank-prompts MCP.
In the Gateway tab, installed tools appear automatically. Three clean states: track only · via gateway · reverted.
Swap to any third-party model in a click — no client changes.
Each app can bind its own scenario route; the global supply chain is the fallback. When local is unavailable it switches to community sharing — transparent to the agent throughout.
Desktop, CLI, and server gateways each register as a device; after sign-in, usage is reported and merged in the cloud.
Use your own free tiers, subscriptions and metered accounts first; tap community shared compute only when needed — both switch transparently inside one gateway.
Supply under your name: local Ollama, vendor free tiers, app/API subscriptions, and metered accounts billed at list price.
Members contribute idle quota, shared through a credits network. Spend credits to call models others supply — often cheaper than official.
Contribute quota for models you have, earn credits, and spend them on models you don't.
Idle quota becomes credits you bank — draw on them anytime later, across models and hours.
Open rules, open source. The contribution rate beats the consumption rate; higher quality earns a higher multiplier.
Contribute local Ollama, idle upstream quota, even private models on a corporate intranet — agents dial out over WebSocket, no inbound ports needed.
Steadier uptime, faster responses, and higher success rates raise your quality multiplier — long-term contributors come out ahead.
Create or join circles of like-minded AI users. Share contributed models inside the group, invite friends for mutual credit rewards, and contribute to multiple circles at once.
One local address fronts every AI tool and agent workflow. Double-click to install on macOS / Windows, or run from the CLI on Linux.
Your tools keep their native model names and configs — Token Bank sits in between as one local address. It traces every token, routes to the cheapest viable source, compresses requests losslessly, and converts protocols, all transparently.
No. Keys stay on your machine and are used only by the local gateway. Sign-in syncs usage statistics for cross-device aggregation — never your credentials.
Community routing is opt-in and only used as a fallback. Contributors dial out over WebSocket with no inbound ports exposed; settlement rules are open and credits are auditable, so contribution always out-earns consumption.
Anthropic, OpenAI, and Codex protocols are auto-converted, plus Ollama, Groq, GitHub Models, and any OpenAI-compatible endpoint. You can swap in third-party models per app without touching the client.
Yes — Token Bank is Apache-2.0 licensed and free to download for macOS, Windows, and Linux / Docker. Community credits are earned by contributing, not bought.
Circles are small communities inside Token Bank. Create or join via invite link, share models scoped to your circle, post messages to coordinate, and earn mutual credit rewards when friends join. You can contribute to multiple circles at once from the Contribute settings.
In Playground, pick a main agent as the aggregation entry. It receives your task, can plan steps, and dispatch work to other onboarded agents. Combined with multi-account CLI and the MCP / Skill / Prompt resource hub, Token Bank becomes a local agent control plane, not only a token proxy.