Apache-2.0 MCP server and Python SDK that generates, audits, tests and evolves agent workflows from natural language, with an optional read-only DeepSeek Harness viewer plugin.
DSH integration
Compatible
Author-claimed
Safety audit
Unaudited
Last verified
2026-11-13
License
Apache-2.0
01What can it help you accomplish?
Turn a described process into a workflow you can run, instead of a one-off answer
A workflow graph, node documentation, audited and tested Python nodes and a FastAPI backend under `workspace/workflows/<workflow_id>`
Teams that want an editable, reusable pipeline rather than a chat answer, driven from an MCP-capable agent session
Change a running workflow without rebuilding the nodes that did not change
An evolved workflow or a single evolved node, with unchanged nodes reused
Developers iterating on an existing agent pipeline while keeping the reviewed parts intact
Watch generated workflows from inside DeepSeek Harness
A read-only workflow viewer for DeepSeek Harness Web and Desktop, next to the skill that creates, evolves and runs the workflows
DeepSeek Harness users who want the generated artifacts visible in the app while the build/run path stays on MCP or the SDK
02How to install into DeepSeek Harness
Prerequisites
- Python 3.10 or later, plus Git to fetch FlowX and its workflow runtime dependency
- An OpenAI-compatible LLM provider — DeepSeek is the default
- An MCP-capable agent client for the local stdio path — the README lists this as a prerequisite
Installation steps
- 01
Clone the repo: `git clone https://github.com/AIpRoBuilder/FlowX.git`
$ git clone https://github.com/AIpRoBuilder/FlowX.git
- 02
Create a virtual environment and install FlowX with its runtime dependencies: `python -m pip install -e .`
- 03
Copy `.env.example` to `.env` and set `FLOWX_LLM_PROVIDER`, `FLOWX_LLM_MODEL`, `FLOWX_LLM_API_KEY` and `FLOWX_DEFAULT_WORKSPACE`
- 04
Register `flowx-mcp` in your MCP client using the absolute path of the virtualenv binary — MCP hosts may not inherit your shell `PATH`
- 05
Optional: install the DeepSeek Harness viewer plugin from `plugins/deepseek-harness-flowx` separately in each profile
Verify the integration
- Run `flowx-mcp` once by hand to confirm the stdio entry point starts (an MCP host normally starts it itself)
- After registering it, ask the agent to create a workflow and poll `get_build_status` until `result.ok` is true, then inspect it with `get_workflow` and `list_nodes`
03DSH integration and capability boundaries
Outside-in integration: a local stdio MCP server (`flowx-mcp`) plus a Python SDK and a non-streaming A2A endpoint drive workflow generation; the optional DeepSeek Harness plugin adds a read-only workflow viewer and is installed separately in each profile
Workflow generation (graph, nodes, backend)
a natural-language process description→`workflow.json`, `node_docs/*.md`, `nodes/*.py` and `main.py` under `workspace/workflows/<workflow_id>`
generation sends the requirement, and possibly logs, to your LLM providergenerated tests may run during the buildNode audit, test and bounded repair
a generated node and its audit report or failing test log→a repaired node, or a paused build asking for guidance — failing nodes are not published
a failed node test pauses the build until an answer, `test: <guidance>`, `skip`, or an abort is suppliedWorkflow and node evolution
a change request plus an optional target: the whole workflow, a step id or a node id→an evolved workflow that reuses the unchanged nodes
`start_build_workflow` and `start_evolve_workflow` modify existing workflows in placeRun as a service with step inputs
a built workflow, then one step input at a time→a connectable `base_url`; `run_step` reports `awaiting` when the workflow pauses, `reset_session` cancels a paused run, `stop_service` stops it
`run_service` starts or restarts the generated FastAPI service on the hostPython SDK and A2A delegation
SDK calls such as `build_workflow_json` / `build_nodes` / `evolve`, or an A2A `message/send`→`AgentResult` objects with per-stage outcomes, errors, logs and artifact paths; A2A tasks with the result attached
the SDK keeps services alive until the `with` block exitsthe SDK and A2A server read process env vars and do not load `.env` automatically
04Who is it for? When not to use it?
Good for
- Teams that want an editable, reusable pipeline rather than a chat answer, driven from an MCP-capable agent session
- Developers iterating on an existing agent pipeline while keeping the reviewed parts intact
- DeepSeek Harness users who want the generated artifacts visible in the app while the build/run path stays on MCP or the SDK
Not for
- Only one build can be active across workflows and a paused build holds that slot; build status and paused builds live inside the MCP server process, so they cannot be resumed after that process exits.
- Generation calls your LLM provider with the requirement (and possibly run logs) and can execute generated tests, so the provider key and network access are required and generated code should be reviewed before its service is run.
05Compatibility, maintenance and safety notes
- Only one build can be active across workflows and a paused build holds that slot; build status and paused builds live inside the MCP server process, so they cannot be resumed after that process exits.
- Generation calls your LLM provider with the requirement (and possibly run logs) and can execute generated tests, so the provider key and network access are required and generated code should be reviewed before its service is run.
- Node builds pause by default on a failed node test and wait for an answer, which is why the MCP build interface is background-only: poll `get_build_status`, then resume with `submit_background_build_input` or abort.
- The `.env` file is not loaded automatically outside the MCP server path: the SDK and A2A server read process environment variables, and an installed `flowx-mcp` needs `FLOWX_CONFIG_ROOT` to find a `.env` outside the checkout.
Apache-2.0 · README evidence only, latest release v1.0.2, safety not audited
06Frequently asked questions
How does DeepSeek Harness connect to FlowX?
Through the local stdio MCP server `flowx-mcp`: register the virtualenv binary by absolute path (MCP hosts may not inherit your shell `PATH`) and set `FLOWX_LLM_API_KEY` and `FLOWX_DEFAULT_WORKSPACE` in the server entry. FlowX also ships a Python SDK and a non-streaming A2A endpoint, and an optional DeepSeek Harness plugin that adds a read-only workflow viewer, installed separately in each profile.
What exactly does a build produce?
A workflow graph, node documentation, audited and tested Python nodes and a FastAPI backend, all under `workspace/workflows/<workflow_id>`; `run_service` then returns a connectable `base_url`, and `run_step` / `reset_session` let you feed single steps and cancel a paused run.
Do failing nodes get published?
No. Each node goes through a generation → audit → test → bounded repair chain and failing nodes are not published. A failed node test pauses the build by default: answer with the user guidance, `test: <guidance>` to amend the test, `skip` to skip the amendment, or abort with `abort_background_build`.
Why does the MCP build interface run in the background?
Because builds can pause for input. Status and paused builds live in the MCP server process, only one build can be active across workflows, and they cannot be resumed after that process exits — so the host polls `get_build_status`, resumes with `submit_background_build_input`, and uses `list_pending_builds` to recover pending questions while the server is still alive.
What do I need before installing?
Python 3.10 or later, Git (for FlowX and its `ag_ui_workflow` runtime dependency), an OpenAI-compatible LLM provider — DeepSeek is the default — and an MCP-capable agent client. Install with `python -m pip install -e .` inside a virtualenv and keep the provider key in `.env` (or the MCP server env block) rather than in a repository.
07Related DSH workflows
huashu-design
by alchaincyf
Creates selectable design directions and delivers interactive prototypes, PPTX, animations, and visual assets.
dsh-tui
by ccch1mneyyy
Claude Code 风格全屏终端 TUI：像素鲸鱼顶栏、双流光大字、实时工作状态行、思考流式展开、双击 Esc 时间回溯、上下文进度条 + TPS 仪表；接入 Agent preset（standard/code/minimal/cordis）、Skills、MCP、Goals、Todos、子代理，/ 命令注册表与随附技能（audit/bug/review/practice/pr_comments/release-notes/vuln-check）；纯插件挂载零核心改动。
dsh-better-sidebar
by omdsh-dev
Full sidebar workbench with file rendering and editing, terminal, Git, and subagents; third-party plugins can register new tabs.
dsh-media-desk
by woniuniuniu
Capture ideas and generate scripts, titles, covers, images, storyboards, schedules, and review data in a sidebar media workspace.
08Data and sources
FlowX lets AI agents **build, run, and evolve workflows** through MCP, a Python SDK, or A2A.
the optional [DeepSeek Harness plugin](plugins/deepseek-harness-flowx/README.md) provides a read-only workflow viewer fo…
**Skill and plugin update.** The DeepSeek Harness plugin settings now accept `FLOWX_LLM_API_KEY` and no longer ask for a…
This page is generated from the project’s public documentation, repository metadata and a structured parse of DSH Plugins; last verified on 2026-11-13. Found an error? Submit a correction.
Best DeepSeek Harness Plugins
Twelve plugins worth installing first — picked from the whole catalog, across every category.