MPL-2.0 的 Go 语言 LLM 网关，对 DeepSeek 提供深度兼容：npx/Docker 一键安装，编程智能体接入其 OpenAI/Anthropic 兼容端点，本地 Web UI 管理路由。
DSH 适配
兼容
作者声明
安全审计
未审计
最后核验
2026-08-21
许可证
MPL-2.0
01它能帮你完成什么？
Give coding agents a unified model endpoint with DeepSeek best compatibility
OpenAI- and Anthropic-compatible endpoints (http://localhost:12580/tingly/openai/v1 and /tingly/anthropic) with protocol translation, Thinking-workflow reasoning support, cache-hit optimization, vision proxy and web search for DeepSeek
Developers running DeepSeek models in coding agents who want one local endpoint with broad protocol compatibility
Wire coding agents to the gateway with one click
One-click configuration for Claude Code and OpenCode (manual guides in-app for Xcode and others), with transparent proxying for SDKs and CLI tools
Teams using coding agents or custom AI tools that need a quick, low-friction model endpoint setup
Control AI agents remotely from IM apps
Remote task execution and monitoring through Telegram, DingTalk, Feishu, Lark, Weixin, WeCom, Slack and Discord bots configured in the web UI's Remote section
Builders who want to run and check on agents from a phone or share agent access across a team
02如何接入 DeepSeek Harness？
前置条件
- Node.js LTS with npx for the recommended install path (the README provides install steps via nvm for macOS/Linux and winget or nodejs.org for Windows); or Docker for the container paths
- Go 1.26+ only applies to building from source (source build is documented under the Contributing Guide)
安装步骤
- 01
Run `npx tingly-box@latest` (or `npx -y tingly-box@latest`) — installs and runs the binary, auto-restarts, migrates and opens the Web UI when run without args
$ npx tingly-box@latest
- 02
If you hit network issues, use the bundled binary: `npx -y tingly-box-bundle@latest` (CN npm mirrors are supported, e.g. `npx --registry=https://registry.npmmirror.com -y tingly-box-bundle@latest`)
$ npx -y tingly-box-bundle@latest
- 03
Or run from Docker: `mkdir tingly-data` then `docker run -d --name tingly-box -p 12580:12580 -v `pwd`/tingly-data:/home/tingly/.tingly-box ghcr.io/tingly-dev/tingly-box`
- 04
Or use Docker Compose for an isolated environment: `docker-compose up -d` (Web UI at http://localhost:12581; `docker-compose down` to stop)
- 05
Open the Web UI at http://localhost:12580 and configure providers, routes and agent integrations there
验证接入成功
作者未说明
03DSH 适配与能力边界
Runs as a standalone local LLM gateway beside DSH (not installed inside the harness): agents connect to its OpenAI/Anthropic-compatible endpoints, while the repo is tagged as a dsh plugin and the README documents DeepSeek best compatibility
Agent-first model gateway
model providers (API keys and OAuth providers like Claude.ai / Codex)→unified OpenAI/Anthropic-compatible endpoint for agents with automatic protocol translation; adds typically < 1ms of overhead
proxies requests to remote model providers (network egress); stores per-user configuration and usage data locallySmart routing engine
models, tokens and custom policies→requests routed across models and tokens based on cost, speed, or custom policies — beyond simple load balancing
Visual control plane (Web UI)
browser on http://localhost:12580→manage providers, routes, aliases, models and remote bots at a glance — no config files needed; includes client-side usage analytics (token consumption, latency, cost estimates)
writes configuration locally instead of config filesLocal stdio MCP web tools
an MCP client→local stdio MCP server exposing `web_search` / `web_fetch` tools
fetches web content over the network when called
04适合谁？何时不该用？
适合
- Developers running DeepSeek models in coding agents who want one local endpoint with broad protocol compatibility
- Teams using coding agents or custom AI tools that need a quick, low-friction model endpoint setup
- Builders who want to run and check on agents from a phone or share agent access across a team
不适合
- Known issues have been recorded in a public fault record; the maintainers ask users to update to the latest version to resolve them, so staying on the latest release is effectively required.
05兼容性、维护与安全提示
- Known issues have been recorded in a public fault record; the maintainers ask users to update to the latest version to resolve them, so staying on the latest release is effectively required.
- The gateway serves local agents but forwards requests to remote model providers, so it needs network access and valid API keys or OAuth authorizations for the providers you route to.
- The recommended install path requires Node.js LTS with npx; container paths require Docker, and Docker Compose uses port 12581 to avoid conflicting with a host install on 12580.
MPL-2.0 · actively maintained (latest release v0.260819.0, 2026-08-19)
06常见问题
Tingly Box 如何与 DeepSeek Harness 配合？
Tingly Box 作为独立本地网关运行在 harness 旁边，而不是安装到 harness 内部：智能体连接它的 OpenAI/Anthropic 兼容端点（http://localhost:12580/tingly/openai/v1 或 /tingly/anthropic）。仓库标记为 dsh plugin，README 也说明了 DeepSeek 最佳兼容性——协议适配、推理思考链、缓存优化、视觉代理、网页搜索和 1M 上下文。
安装前需要准备什么？
推荐的 `npx tingly-box@latest` 安装方式需要 Node.js LTS 与 npx；容器方式需要 Docker / Docker Compose。只有从源码构建时才需要 Go 1.26+。
请求和数据会流向哪里？
网关在本机 12580 端口运行，把请求透明代理到你通过 API Key 或 OAuth（Claude.ai、Codex 等）配置的模型服务商；配置和每用户用量数据保存在本地。Guardrails 文档介绍了基于策略的安全检查与敏感凭据脱敏。
支持 MCP 吗？
支持——Tingly Box 附带一个本地 stdio MCP 服务器，提供 `web_search` / `web_fetch` 工具，详见 docs/mcp-web-tools.md。
有哪些限制？
已知问题记录在公开的 fault record 中，维护者要求更新到最新版本来解决。此外网关需要联网，并且你路由到的模型服务商要有有效的 API Key 或 OAuth 授权。
07相关的 DSH 工作流
openviking
作者 volcengine
为 AI 智能体打造的自进化上下文数据库，统一智能体记忆、知识 RAG 与技能。
colleague-skill
作者 titanwings
将冰冷的离别化为温暖的 Skill，欢迎加入数字生命1.0！Transforming cold farewells into warm skills? It's giving rebirth era. Welcome to Digital Life 1.0. 🫶
archify
作者 tt-a1i
为编码智能体生成美观可验证的架构图、时序图与数据流图，输出自包含 HTML，支持动效与清晰导出。
learn-harness-engineering
作者 walkinglabs
Harness 工程新手教程，从 0 到 1 系统学习智能体工作流框架。
08数据与来源
DeepSeek is optimized for mainstream agent workflows, offering broad compatibility across protocol adapters, agent clien…
Tingly Box **serves agents, coordinates AI models, optimizes context, and routes requests** for maximum efficiency — wit…
页面基于项目公开文档、仓库元数据和 DSH Plugins 的结构化解析生成；最后核验于 2026-08-21。发现错误？提交更正。