Use any model with any AI agent — safely.
Thomas is the plug-and-play model hub that connects your agents to your models, with security guardrails, quotas, fallback, and cost controls.
Thomas is the universal adapter between AI agents and model providers. Install it once, and any agent on your machine — Claude Code, Codex, OpenClaw, Hermes Agent — can talk to any provider — Anthropic, OpenAI, OpenRouter, Kimi, DeepSeek, Groq, or your own OpenAI-compatible endpoint — without editing each agent's own configuration.
Roadmap order: connect → control → optimize → protect.
| Stage | Capability | Status |
|---|---|---|
| connect | Discover installed agents; route any agent to any provider; cross-protocol translation (Anthropic ↔ OpenAI) including streaming | ✅ v0.1.0 |
| control | Per-agent quotas, allowed-models policies, audit log | 🚧 planned |
| optimize | Multi-provider fallback on failure; cost-aware routing; latency-aware routing | 🚧 planned |
| protect | Prompt-injection / PII / secret detection; tool-call guardrails | 💼 commercial (planned) |
The open core is licensed under Apache-2.0. Security, governance, and team/enterprise features ship as a separate commercial product (thomas-cloud) that thomas can talk to over HTTP.
Audience: individual users, developers, solo and small teams. Not aimed at enterprise procurement.
| Approach | Problem |
|---|---|
Manually edit each agent's config (~/.claude/settings.json, ~/.codex/auth.json, …) |
Brittle. Five agents = five drift surfaces. |
| Use a profile-switcher that rewrites those configs | When you uninstall the switcher, the agent stays pointed at whatever it last wrote. |
Use a router that requires ANTHROPIC_BASE_URL exported into your shell |
When the router stops, your agent breaks. |
| thomas | Installs a transparent shim earlier in PATH. Original config is never touched. Uninstall thomas → shim disappears → every agent reverts. |
npm i -g @trustunknown/thomas
thomas doctorRequires Node 20+.
# 1. See what agents and credentials you already have on the host
thomas doctor
# 2. Wire an agent through thomas (installs a transparent PATH shim)
thomas connect claude-code
# 3. Add a provider key (only if thomas didn't import one for you)
thomas providers add openrouter sk-or-v1-...
# 4. Switch which model that agent uses (without touching its own config)
thomas route claude-code openrouter/anthropic/claude-sonnet-4.5
# 5. See current state
thomas list
# 6. Optional: supervise the proxy with launchd / systemd so it survives reboot
thomas daemon install
# Revert at any time
thomas disconnect claude-codeAfter thomas connect, add ~/.thomas/bin to your PATH (the command prints the exact line).
Agents (CLIs that thomas knows how to detect, import credentials from, and shim):
- Claude Code
- Codex CLI
- OpenClaw
- Hermes Agent
Providers (built-in routing targets):
anthropic(Anthropic API)openai(OpenAI API)openrouterkimi(Moonshot AI)deepseekgroq
Plus any OpenAI-compatible or Anthropic-compatible endpoint via thomas providers register.
Both directions of /v1/messages ↔ /v1/chat/completions are translated, including streaming SSE — so a Claude Code (Anthropic-shape) agent can talk to OpenRouter / Groq / Kimi (OpenAI-shape), and a Codex (OpenAI-shape) agent can talk to Anthropic. System prompts, tool definitions, tool calls, tool results, image inputs, stop reasons, and SSE events are all mapped.
This repo ships a SKILL.md in the root. AI agents that find this repo on GitHub can read it directly to drive thomas on a user's behalf.
To install the skill into your local Claude Code skill directory:
thomas skill install claude-codeAfter this, your Claude Code session can answer "switch claude code to kimi" / "show me which agents are connected" autonomously.
v0.1.0 — public alpha. The connect/route stage of the roadmap is complete and tested. Quotas, fallback, cost-aware routing, and security guardrails are not yet implemented.
This project is licensed under the Apache License 2.0 (Apache-2.0).
This means:
- ✅ You can use, modify, and distribute this software
- ✅ You can use it for commercial purposes
- ✅ Includes an explicit patent-license grant from contributors
- ✅ You can build closed-source products on top (e.g. ship your own agent adapter as a private package)
⚠️ You must preserve copyright and license notices, and include the NOTICE file when redistributing⚠️ "thomas" is a trademark of Trust Unknown; the license does not grant trademark rights
See the LICENSE and NOTICE files for the full text.
The agent and protocol adapters in this repo (the L1 and L2 layers — what makes thomas talk to Claude Code, Codex, OpenClaw, Hermes, etc.) live closer to ecosystem infrastructure than to product IP. Apache-2.0 lets the community fork, contribute, and embed those adapters without copyleft friction; meanwhile the proprietary value (policy intelligence, multi-tenant control plane, security analysis) lives in a separate commercial product (thomas-cloud) that communicates with thomas over HTTP.