Open-source AI gateway · MPL-2.0

Any agent.
Any provider.
One box for your team.

Tingly Box plugs Claude Code, Codex, OpenCode, Xcode and every other agent into any model provider — and gives your team shared keys, routing rules, usage and guardrails in one place.

npm i -g tingly-box install

tb start run

Any agent ⇄ any provider

30+model providers, cloud and self-hosted

10+agents with one-click or guided setup

3protocols translated both ways: Anthropic, OpenAI, Gemini

< 1 mstypical gateway overhead

How it works

From one laptop to a whole team, on one endpoint.

01 · Your agents

Keep the tools you already use

Claude Code, Codex, OpenCode, Cursor, Xcode, VS Code — or any app that speaks an OpenAI or Anthropic API.
02 · One endpoint

Point them at Tingly Box

One-click config writes the base URL and key for you: localhost:12580/tingly/<scenario>.
03 · Any provider

Route to any model

Rules map a model name to providers, with smart routing and fallback. Protocols are translated both ways, and OAuth lets you reuse existing quotas.
04 · Your team

Hand out sharing keys

Each member gets a dedicated key on a shared team endpoint. Model rules are managed centrally; data stays isolated per user.
05 · Power-ups

Switch on more power

Guardrails, MCP tools, Usage and Remote come built in. Turn on what you need, when you need it.

Any agent → any provider

Bring your own agent. Pick any model behind it.

Agents keep their native protocol. Tingly Box converts between Anthropic, OpenAI and Gemini formats, so Claude Code can run on DeepSeek and Codex can run on Claude.

Agents

Providers

One-click agent setup

Claude Code, OpenCode and Codex are configured for you; Xcode, VS Code and others get a step-by-step in-app guide.

Claude Code profiles

Switch between profiles that map each task tier to a different model — Opus on one provider, Haiku on another.

API keys and OAuth

Add providers by key, or sign in with OAuth and reuse the quota you already pay for, in any compatible tool.

Smart routing

Route across models and tokens by cost, speed or your own policy, with load balancing and fallback built in.

Self-hosted models

vLLM, SGLang, Ollama, LM Studio, LocalAI and Jan connect like any cloud provider, with defaults pre-filled.

Image API

Route image generation and editing through the same gateway at /tingly/imagegen, with a built-in playground.

Agents page in Tingly Box listing Claude Code, Claude Desktop, Codex, OpenCode, Xcode, VS Code and SDK scenarios

Team AI management

Give everyone AI. Keep one place to run it.

One team endpoint backed by centrally managed model rules. Members get their own keys; you get per-user usage, isolation and policy.

1

Set the rules once

Open Team and decide which models the team can call, and which providers serve them.
2

Hand out sharing keys

Create a key per member under Sharing Keys. Revoke or rotate one without touching anyone else.
3

Track who uses what

Dashboard → Team usage shows tokens, latency and cost estimates per member and per model.

Team page with base URL, plugins and model rules routing each model to a provider

Guardrails

Policy checks on requests and responses, built-in protections, and masking for protected credentials.

Guardrails guide

MCP gateway

Register MCP servers once and share them with every agent — per source or aggregated — with built-in web_search and web_fetch.

Remote control

Drive agents from Telegram, WeChat, WeCom, Feishu, Lark or DingTalk — from any device.

Usage analytics

Tokens, latency, cost estimates and model selection for every request, across people and scenarios.

Inside the box

A UI that makes the complex parts visible.

Quick start

Running in a minute.

macOS, Linux and Windows. Install from npm — the platform binary ships inside the package, so an npm mirror is all you need behind a firewall.

# install (recommended)
npm install -g tingly-box@latest

tb start   # runs in the background (--no-daemon for foreground)
tb open    # opens the web UI

# update: reinstall, then restart
npm install -g tingly-box@latest && tb restart

# behind a firewall, any npm mirror works
npm install -g tingly-box@latest --registry=https://registry.npmmirror.com

# one-off run without a global install
npx -y tingly-box@latest

# with a mirror
npx --registry=https://registry.npmmirror.com -y tingly-box@latest

mkdir tingly-data
docker run -d --name tingly-box \
  -p 12580:12580 \
  -v "$(pwd)/tingly-data:/home/tingly/.tingly-box" \
  ghcr.io/tingly-dev/tingly-box

# any OpenAI-compatible client works the same way
from openai import OpenAI

client = OpenAI(
    api_key="your-tingly-model-token",
    base_url="http://localhost:12580/tingly/openai/v1",
)

# any Anthropic-compatible client works the same way
from anthropic import Anthropic

client = Anthropic(
    api_key="your-tingly-model-token",
    base_url="http://localhost:12580/tingly/anthropic",
)

1
Open the UI
tb open, or http://localhost:12580
2
Connect AI
Add a provider by API key, OAuth or a self-hosted URL.
3
Pick an agent
Claude Code, Codex, OpenCode… and apply the config.
4
Invite your team
Optional: set rules and hand out sharing keys.

Learn more

Docs and community

User manualInstallation, configuration and operations Docker guideCompose, custom images, troubleshooting GuardrailsPolicies, protections, credential masking IssuesReport a bug or ask for a feature DiscussionsQuestions, ideas and setup help ContributingBuild from source and send a PR

Plug in your first agent in a minute.

Open source, runs on your machine, adds under a millisecond.