Token Buffet: CLI LLM Proxy Router

Expose your flat-rate CLI subscriptions (Claude, Codex, Grok, agy) as a local OpenAI/Ollama/Anthropic-compatible API. Route requests by semantic tier (fast/medium/reasoning), smart load balance by quota, and power your downstream tools without paying per token.

Learn the Concept

For an overview of the LLM Wiki pattern and how this fits in, watch:

Karpathy LLM Wiki Overview

(Note: The video uses Claude Code as the example, but explains the LLM Wiki concepts and patterns well. Token Buffet applies the same pattern to any CLI-based LLM subscription (Claude, Codex, Grok, agy) and powers Obsidian vaults, community plugins, and agent frameworks.)

This shows how to turn your subscriptions into a local knowledge-synthesis engine for tools like Obsidian, chatbot clients, and AI agents.

Features

• Small runtime: The proxy, router, and dashboard use Python's standard HTTP server. Extractor capture needs its optional pseudo-terminal dependencies, and the Responses endpoint needs the OpenAI SDK.
• Semantic Routing: Translates generic model requests (fast, medium, reasoning) into specific provider models based on remaining usage quotas.
• Smart Load Balancing: Reads exported CLI quota files and routes requests to the provider with the most remaining quota.
• True Parallelism: Spawns sandbox directories to bypass SQLite locking limits in native CLIs.

Architecture

Quick Start

Token Buffet always works in the folder you run it from. Make one folder for it:

mkdir TokenBuffet && cd TokenBuffet
pip install token-buffet               # Python 3.12+; installs everything it needs
tbuffet init                           # scaffold, register the provider CLIs found on PATH, check app integrations
tbuffet demo init                      # optional: copy the demos, add them to their apps, list which can run
tbuffet start                          # proxies + router (8090) + dashboard (8080); Ctrl+C stops this session's process trees
tbuffet start --background             # explicitly leave services running in the background
tbuffet status | stop | restart | doctor

Log in to each vendor CLI (claude, codex, grok, agy) once. init doesn't check logins; a provider that isn't signed in reports its error at tbuffet start.

Demos (pre-configured)

tbuffet demo init copies ready-to-use demos into demo/. Each one already points at the router, so after tbuffet start:

• Obsidian LLM Wiki (needs Obsidian installed): run tbuffet open obsidian conference-notes (or daily-notes), then choose Trust author and enable plugins. See the demo README.
• Chatbots and Claudia: see demo/README.md.

Providers and your own apps

tbuffet add-provider claude --validate        # register a built-in provider + login/quota check
tbuffet add-provider mine --from config.json  # import a config authored with the register-provider skill
tbuffet list-providers | remove-provider NAME
tbuffet add-app obsidian my-vault --path "/path/to/vault" --vault-type daily-notes
tbuffet list-apps | reconfigure-app TYPE NAME

Directory Structure

• src/tokenbuffet/: proxy, router, dashboard and launcher
• src/tokenbuffet/extractors/: provider usage and model capture/parse engine
• src/tokenbuffet/default_templates/: what tbuffet init copies into your folder:
  • config/
  • integrations/ (built-in app types, and providers/ with the built-in provider configs)
  • scripts/
  • demo/ and the demo registrations in registered_apps/ are copied by tbuffet demo init

In your Token Buffet folder:

• integrations/providers/: built-in provider configs (what Token Buffet supports)
• registered_providers/: the providers registered on this machine
• data/: runtime state (providers_status/, provider_homes/, temp/)
• logs/: proxy, router and extractor logs

From a repository checkout

Install editable with pip install -e ., then run tbuffet init and tbuffet start from the repository root.

tbuffet start launches every provider proxy, the router (port 8090) and the dashboard (port 8080); tbuffet stop stops them. tbuffet dashboard starts the dashboard alone.

The router listens on port 8090, and the dashboard is at http://127.0.0.1:8080/. See the documentation index for the bridge guides.