JARVISLABSCO SUMMARY
Archived, and still worth reading.
The patterns it shipped in June 2024 — todos as an agent dispatch queue, tools gated by phase, a planning stage before execution — later turned up in mainstream tooling. It runs locally the same as it always did; it just is not moving any more.
What it is
A Node MCP server for workflow orchestration. It runs an explicit INIT → QUERY → ENHANCE → KNOWLEDGE → PLAN → EXECUTE → VERIFY → DONE loop, restricts which tools are reachable in each phase, and treats todo items as dispatches for specialised agents. The author has archived the project.
What you get
JARVIS— the eight-phase workflow controllerAPITaskAgent— API discovery and fetching with SSRF protectionPythonComputationalTool— Python execution for data analysisIronManusStateGraph— session state managementSlideGenerator— HTML slide generationHealthCheck— runtime diagnostics- Phase-gated tool access, so writing files during verification is structurally impossible rather than discouraged
- Nine roles with distinct thinking methods:
planner,coder,critic,researcher,analyzer,synthesizer,ui_architect,ui_implementer,ui_refiner - File-based handoff between agents that cannot share context directly, under
./iron-manus-sessions/{session_id}/
Requirements
Node.js 20+, built from source with npm install && npm run build, or run from the Docker image. Outbound calls are gated: ALLOWED_HOSTS is the allowlist, ENABLE_SSRF_PROTECTION turns the check on, and KNOWLEDGE_MAX_CONCURRENCY and KNOWLEDGE_TIMEOUT_MS bound them. Private IP ranges are blocked and only HTTP and HTTPS are allowed. MIT. The repository is archived — nothing new is landing in it.
Setup effort
Build from source — git clone https://github.com/dnnyngyen/iron-manus-mcp
USE CASES
01
Forcing a plan before any code is written
The planning phase can only reach the todo tool, so the model cannot skip ahead to editing files.
02
Keeping verification honest
During verification only the read tools are available, so a failing check cannot be quietly rewritten away.
03
Splitting a task across specialists
A todo carries the role, the context and the expected output, and becomes a dispatch to an agent primed for that role.
04
Letting agents that cannot see each other share findings
Research and output land as files in the session workspace, which the next agent reads.
05
Fetching from an API without opening the whole network
Calls go through the allowlist and the SSRF check, with concurrency and timeouts capped.
06
Turning the result into something presentable
The findings go straight into generated HTML slides rather than a separate write-up pass.
6 tools
Listed in project docs. Your client may show a different list.
APITaskAgentPythonComputationalToolIronManusStateGraphSlideGeneratorHealthCheckBEFORE YOU INSTALL
What to know first
Tools you get
6 tools named in the project's documentation.
Requires
node on PATH
INSTALLATION
Connect your client
Server setup
TERMINAL
git clone https://github.com/dnnyngyen/iron-manus-mcp
cd iron-manus-mcp
npm install
npm run buildMCP Clients
Claude Code
CLAUDE CODE · TERMINAL
claude mcp add iron-manus-mcp node dist/index.jsRun this from inside the iron-manus-mcp folder you built: dist/index.js is a relative path.
Any MCP client
ANY MCP CLIENT · JSON
{
"mcpServers": {
"iron-manus-mcp": {
"command": "node",
"args": [
"<YOUR_PATH_TO_IRON_MANUS_MCP>/dist/index.js"
]
}
}
}<YOUR_PATH_TO_IRON_MANUS_MCP> is the absolute path of the folder you cloned and built.
