dot-gateway
Enables ChatGPT Dot to connect through OpenAI Secure MCP Tunnel to local coding agents, providing multi-turn sessions and persistent context via session IDs.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
dot-gateway
让 ChatGPT Dot 通过 OpenAI Secure MCP Tunnel,跟本机的编码 agent(默认是 DeepSeek Harness dsh --profile acp)进行多轮会话。Dot 可以一轮接一轮地追问,agent 会记得上下文;网关重启、机器重启以后,用同一个 sessionId 还能接着聊。
架构
flowchart LR
Dot["ChatGPT Dot"] -- "MCP 工具调用" --> T["Secure MCP Tunnel"]
T --> GW["本机上的 dot-gateway"]
GW -- "ACP" --> DSH["dsh"]
GW -- "ACP" --> MORE["……"]Dot 只会 MCP,网关负责翻译:对上用 MCP 和 Dot 说话,对下用 ACP(Agent Client Protocol)和各个 agent 说话。工具照 ACP 会话动作命名,任何走 stdio 的 ACP agent 写进配置就能接入。
Related MCP server: pi-agent-mcp
快速开始
cd ~/Workspace/dot-gateway
uv sync
scripts/start.sh # 前台运行,监听 http://127.0.0.1:8765/mcp(Streamable HTTP)
uv run pytest -q # 单元测试,用假 agent,不花钱tunnel-client 用 --mcp-server-url http://127.0.0.1:8765/mcp 连过来。部署、开机自启、配置项和技术细节见 AGENTS.md。
工具
工具 | 作用 |
| 开一个会话;带 prompt 时发出去就返回,不等跑完 |
| 同一个会话的下一轮;一轮没结束再调会返回 |
| 轮询结果。默认 |
| 某次工具调用的完整输入和输出;有 Auto review audit 时附 |
| 会话概况; |
| 接回会话(dsh 用 |
| 中断这一轮 / 结束会话 |
| 切换模型等配置、回答权限请求、列出 agent |
返回值默认是一段紧凑的 JSON(不额外返回 structuredContent;部署设置 DOT_GATEWAY_STRUCTURED_OUTPUT=1 后会同时返回 structuredContent)。最终答复只投递一次,如果没收到,传 includeText=true 或调 get_session(detail="full") 再取一次;messages/full 也可以用交付时相同的 afterSequence 重试那一页。
网关不加安全限制,但权限有两层:permissionMode 是 agent 自己的沙箱(默认 workspace-write);permissionPolicy 只决定网关怎么回答 agent 自己发起的权限请求,不是每次写入/命令的审批开关。沙箱允许的操作直接执行,所以 operator 下 pendingPermissions 为空、permissionRequests 为 0 是正常的;策略每一轮都可以调整。
某个调用被沙箱拒绝时,dsh 可以对同一调用重试一次、申请严格更宽一档的沙箱(带 justification),执行前由 dsh-acp 桥接成 ACP 权限请求。网关策略:read_only(默认)自动拒绝这类请求;allow_all 自动允许;deny_all 自动拒绝;operator 挂起在 pendingPermissions 里等 resolve_permission(条目带 toolCallId 和一行回填的工具摘要,回填不到是 tool)。选 allow-once 只放宽该次调用,不改变会话的 permissionMode;要整会话换档用 prompt(permissionMode=...)。
限制:dsh-acp 的 ACP 请求只带 toolCallId,申请的模式/理由只能从工具输入回填(command、sandboxPermissions、justification);Auto review 的审查理由没有 ACP 字段,网关看不到。
部署可以选择启用 dsh 的实验性 Auto review(dsh --profile acp-auto + permissionMode=auto + permissionPolicy=operator):执行前由模型逐调用审查,低风险直接以 full access 运行,medium/high 被拒时挂起问人。安装与配置见 docs/auto-review.md;vendored profile 会把 reviewer 的 decision/risk/reason 写到本地 audit 旁路,网关在 toolReviews/get_tool_call.review 里返回;scripts/test_auto_review.py 提供 allow/ask/deny 的确定性测试入口。这是实验功能,会多花模型调用,也可能误放行/误拒绝。
许可证
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Connect ChatGPT and Claude to your Mac, Windows or Linux with secure MCP remote access.
Real-time chat for AI agents. Claude Code, Cursor, Cline and Codex join channels over MCP.
Persistent memory, knowledge graph, task management and more for AI agents via MCP.
Remote MCP server for supportsheep: run AI interviews and manage support content for your blog.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceEnables ChatGPT (or any MCP client) to delegate coding tasks to a local Hermes-backed agent with async job management, supporting read-only investigation, implementation, and continuation of sessions via secure MCP tunnel.1MIT
- FlicenseNot gradedqualityDmaintenanceEnables MCP hosts like Claude Code and Codex to spawn, manage, and interact with persistent, reusable Pi coding-agent sessions, supporting task dispatch, status checks, and session lifecycle control.7 npm-
- AlicenseNot gradedqualityCmaintenanceLets MCP clients create, resume, message, and stop Pi coding-agent sessions running on another machine, with conversations that survive client disconnects and server restarts. Sessions can be kept loopback-only or published privately to trusted callers over Tailscale HTTPS.1Apache 2.0
- AlicenseNot gradedqualityAmaintenanceLets MCP-compatible clients run reusable Markdown-defined coding agents with a chosen backend CLI and return their results to the same conversation. It supports delegating tasks such as code review, test writing, investigation, and documentation, with shared settings and optional session continuation.127 npmMIT