Token Bank

Run one local gateway and every AI coding agent you use shares the same address, key, and usage ledger. Token Bank onboards Claude Code, Codex CLI, Kimi Code, OpenCode, Claude Desktop, OpenClaw, and WorkBuddy in one click by injecting base-URL variables or patching configs, and Revert restores the official settings. Agents keep native model names while the gateway quietly reroutes requests to Ollama, Groq, Cerebras, or DeepSeek, converting between Anthropic Messages, OpenAI Chat, and Codex Responses. Route chains step from local models to free APIs, subscriptions, and pay-as-you-go keys with failover, scene and task presets, and optional lossless JSON compression. Every call is traced with route, model, tokens, latency, and cost, local session logs are imported, and signed-in devices merge into one dashboard. An assets library holds MCP servers, skills, prompts, and agent templates, projects them onto each installed runtime, and a rule-driven butler repairs broken projections. The unified agent API dispatches tasks to local apps, professional templates, external agents joined by pairing code, or agents hired per task, returning results synchronously, in the background, or as a stream. A credit-based market lets you rent out idle quota while keys stay on your machine. The container build serves the desktop interface in a browser on port 11431, with the gateway on 11430 and access control left to your firewall or proxy. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. Apache-2.0 licensed.

Token Bank
Token Bank
Token Bank
Token Bank
Token Bank
Token Bank

Benefits

  • See every token you spend
  • Every request logs its route, resolved model, tokens, latency, and cost, and local Claude and Codex session logs are imported and de-duplicated, giving one dashboard sliced by app, provider, model, and device.
  • Cut costs without touching agents
  • Agents keep native names like claude-sonnet-4-6 while the gateway reroutes them to Ollama, Groq, or DeepSeek, converting protocols automatically and failing over to the next source when one is down.
  • Try it and revert in one click
  • Onboarding injects base-URL variables or patches config files, and Revert restores the official configuration, so trying the gateway never locks you out of your existing Claude Code or Codex setup.
  • Keep every AI asset in sync
  • Skills, MCP servers, prompts, and agent templates live in one library and are projected onto each installed runtime, so a single edit updates every managed agent instead of scattered copies.

Features

  • One-click app onboarding
  • CLI shims inject base-URL variables for Claude Code, Codex CLI, OpenCode, and Kimi Code, while config patches cover Claude Desktop, OpenClaw, and WorkBuddy; Cursor uses an OpenAI-compatible base URL.
  • Protocol-converting router
  • Anthropic Messages, OpenAI Chat, and Codex Responses interoperate, and route chains fall from Ollama to free APIs, subscriptions, and pay-as-you-go keys with failover, scene presets, and optional lossless compression.
  • Unified agent API
  • The /v1/agent endpoint dispatches tasks by harness_id to local apps, professional templates, external agents joined by pairing code, or hired agents, with synchronous, background, and server-sent-event results.
  • Assets library and butler
  • One library holds MCP servers, skills, prompts, and agent templates for projection onto runtimes, while a rule-driven butler repairs broken projections, flags risky skills, and learns from corrections.
  • Usage dashboard and traces
  • Dashboards slice requests, tokens, free-hit rate, and estimated cost by app, provider, model, and device, and step-level session traces show tool calls, skills used, and token breakdown.
  • Credit-based sharing market
  • Rent out idle subscription quota, local models, or agents for credits, or hire rare models and expert agents per call, with keys, skills, and prompts never leaving your machine.

Apps Similar to Token Bank