返回目錄

open-managed-agents

維護狀態: 活躍

openma-ai/open-managed-agents

開源實現 Claude 託管智慧體 API 的自託管執行時,相容 Claude Tag 風格,可運行於 Cloudflare Workers 或 Node.js。

安裝
$ git clone https://github.com/openma-ai/open-managed-agents.git && cd open-managed-agents && cp .env.example .env

243

星數

32

Fork

TypeScript

語言

Apache-2.0

授權條款

2026-04-10

建立於

2026-08-19

最近推送

Apache-2.0 自託管 Agent 執行平台,相容 Claude Managed Agents API,提供工作階段、沙箱、記憶庫與 MCP 工具;README 未記載 DeepSeek Harness 專屬整合。

DSH 整合

生態系相關

自動推斷

安全稽核

未稽核

最後核實

2026-08-21

授權條款

Apache-2.0

01它能幫你完成什麼?

  • Self-host a managed-agent runtime that runs agents with durable sessions, sandboxes and built-in tools

    A running /v1/agents + /v1/sessions API with Console UI on port 8787 (Docker self-host in ~2 min or Cloudflare Workers deploy), verified via the /health endpoint

    Platform and infra engineers who want self-hosted, BYOK agent infrastructure under their own deployment boundary

  • Automate workflows by publishing agents into Slack, GitHub and Linear

    Agents acting as real teammates — assignable on issues, @mentionable, replying in threads — with per-publication capability gates and isolated bot identities

    Engineering and product teams that want agent-driven workflow automation inside the tools they already use

  • Give long-running agents persistent, versioned memory

    Memory stores mounted into the sandbox at /mnt/memory/<store_name>/, accessed with standard file tools, with version history, redaction and CAS updates

    Builders of long-lived agents that need durable cross-session memory under their own deployment

02如何將外掛接入 DeepSeek Harness?

先決條件

  • Docker (self-host path) — the README quick start runs `docker compose up -d`
  • Two locally generated secrets: BETTER_AUTH_SECRET (openssl rand -hex 32) and PLATFORM_ROOT_SECRET (openssl rand -base64 32)
  • Cloudflare path only: Workers Paid plan (required for Durable Objects + Containers)

安裝步驟

  1. 01

    $ git clone https://github.com/openma-ai/open-managed-agents.git && cd open-managed-agents && cp .env.example .env

  2. 02

    Set BETTER_AUTH_SECRET and PLATFORM_ROOT_SECRET in .env; optionally ANTHROPIC_API_KEY so the first agent can run without a Model Card

  3. 03

    docker compose up -d (SQLite + LocalSubprocess sandbox, default) — or docker compose -f docker-compose.postgres.yml up -d for the Postgres backend

驗證整合成功

  • curl localhost:8787/health returns {"status":"ok","backends":{"db":"sqlite ..."}, ...}; the Console UI opens on http://localhost:8787
  • Smoke test the harness end-to-end: POST /v1/agents to create an agent, POST /v1/sessions to start a session, then POST a user.message event

03DSH 整合程度與能力邊界

DSH 整合生態系相關

No DeepSeek Harness integration is documented in the README; OMA is a self-hosted, Claude Managed Agents API-compatible meta-harness platform that runs agents and their workflows as standalone infrastructure

  • Claude Managed Agents-compatible API

    existing clients/SDKs of the Claude Managed Agents APIsame /v1/agents, /v1/environments, /v1/sessions, /v1/vaults endpoints and event types, drop-in compatible

  • Built-in agent toolset (agent_toolset_20260401)

    an agent configured with the built-in toolsetbash, read, write, edit, glob, grep, web_fetch, web_search, cron-style scheduling, and opt-in headless browser

    bash executes commands in the sandbox; web_fetch and web_search make outbound network requests (web_search requires TAVILY_API_KEY)
  • Encrypted vaults with outbound credential injection

    vaults holding static_bearer / mcp_oauth / cap_cli credentials bound to request hostsreal credentials injected at the network layer on matching outbound requests — tokens never enter the sandbox

    credentials are AES-GCM encrypted at rest under PLATFORM_ROOT_SECRET; each forward emits a structured op:"mcp_proxy.forward" log
  • MCP server registry

    up to 20 MCP servers per agent — HTTP/SSE (hosted) or stdio (npm/PyPI packages)upstream tools surfaced to the model as mcp__<server>__<tool>

    stdio servers are spawned inside the sandbox container (talks to 127.0.0.1:port/sse)

04適合誰?何時不該用?

適合

  • Platform and infra engineers who want self-hosted, BYOK agent infrastructure under their own deployment boundary
  • Engineering and product teams that want agent-driven workflow automation inside the tools they already use
  • Builders of long-lived agents that need durable cross-session memory under their own deployment

不適合

  • PLATFORM_ROOT_SECRET is mandatory to boot and encrypts all credentials, model-card API keys and integration tokens at rest — losing it makes every encrypted row unreadable, so it must be backed up.
  • The Cloudflare deployment path requires a Workers Paid plan because it relies on Durable Objects + Containers; the free self-host (Docker/Node) path is the alternative.
  • The default sandbox provider is subprocess with no isolation; an isolated backend (litebox, daytona, e2b or boxrun) should be used when running untrusted agents.

05相容性、維護與安全提醒

  • PLATFORM_ROOT_SECRET is mandatory to boot and encrypts all credentials, model-card API keys and integration tokens at rest — losing it makes every encrypted row unreadable, so it must be backed up.
  • The Cloudflare deployment path requires a Workers Paid plan because it relies on Durable Objects + Containers; the free self-host (Docker/Node) path is the alternative.
  • The default sandbox provider is subprocess with no isolation; an isolated backend (litebox, daytona, e2b or boxrun) should be used when running untrusted agents.
2026-04-102026-08-19@openma/cli@0.5.0

Apache-2.0 · actively maintained (last push 2026-08-19, latest release @openma/cli@0.5.0)

06常見問題

它和 DeepSeek Harness 是什麼關係?如何接入?

README 未記載與 DeepSeek Harness 的直接整合。OMA 本身即是執行 harness 的平台:用 SDK 的 defineHarness 撰寫 harness,透過 `oma deploy --harness my-harness.ts --agent <id>` 部署,平台負責提供工作階段、沙箱、工具、事件日誌與故障復原。

自託管需要什麼前置條件?

需要 Docker,以及兩個本機產生的金鑰:BETTER_AUTH_SECRET(簽署主控台工作階段)和 PLATFORM_ROOT_SECRET(靜態加密憑證,務必備份)。Cloudflare 部署路徑則需要 Workers 付費方案(Durable Objects + Containers)。

如何驗證安裝成功?

執行 `curl localhost:8787/health`,回傳 `{"status":"ok",...}` 即成功,主控台 UI 同在 8787 連接埠。README 也提供端到端冒煙測試:POST /v1/agents 建立 Agent、POST /v1/sessions 開啟工作階段,再傳送一則 user.message 事件。

憑證和金鑰存放在哪裡?

憑證在 PLATFORM_ROOT_SECRET 之下以 AES-GCM 加密靜態儲存;權杖永不進入沙箱,由出站解析器在網路層剋除入站認證標頭並注入真實憑證。遺失 PLATFORM_ROOT_SECRET 會導致所有加密資料無法讀取。

支援 MCP 工具嗎?

支援。每個 Agent 最多可註冊 20 個 MCP 伺服器(HTTP/SSE 託管式或 stdio 套件,stdio 於沙箱容器內啟動),上游工具以 `mcp__<server>__<tool>` 形式暴露給模型。

08資料與來源

  • 自動推斷github.com92556f718f74…

    **Open-source alternative to Claude Managed Agents** — and a foundation for open-source, self-hosted Claude Tag-style ag…

  • 自動推斷github.com92556f718f74…

    Drop-in compatible with the Claude Managed Agents API; runs on Cloudflare Workers + Durable Objects, or `docker compose…

此頁面根據專案公開文件、儲存庫中繼資料與 DSH Plugins 的結構化解析所產生;最後核實於 2026-08-21。發現錯誤?提交更正。

DSH Plugins 是獨立的 DeepSeek Harness 外掛市集,與 DeepSeek 官方無關,也不代表官方背書。第三方外掛未經安全稽核,安裝前請審查原始碼。

每週取得最新的 DeepSeek Harness 外掛,絕不濫發。