poietic-mcp

An MCP server that lets an agent read, edit, validate and simulate Poietic Stock and Flow designs, alongside a human working in Poietic Playground.

The design file is the integration point. The agent edits a .poietic file through this server by driving the poietic CLI; the human opens or reloads the same file in the playground. Agent edits, human views and direct-manipulates; both sides re-read.

# 1. Build and install the poietic CLI (Swift 6.x, macOS 15+ or Linux)
git clone https://github.com/OpenPoiesis/poietic-tool
cd poietic-tool && git checkout v0.9 && ./install  # installs to ~/.swiftpm/bin/poietic

# 2. Build this server
git clone https://github.com/soobrosa/poietic-mcp
cd poietic-mcp && npm install && npm run build

The server works with any MCP client. The command to register is always:

node <absolute-path-to>/poietic-mcp/dist/index.js

with optional environment variables (see below). Environment:

• POIETIC_MCP_WORKSPACE - allowlist root. Design paths and export outputs must live inside it. Defaults to the server's working directory. Set it to a dedicated designs folder (e.g. ~/Documents/poietic-designs) for a tight fence, or your home directory to work with designs anywhere.
• POIETIC_BIN - path to the poietic binary. Defaults to ~/.swiftpm/bin/poietic.
• POIETIC_DESIGN - optional default design path, used when no current design is set.

claude mcp add --scope user poietic \
  --env POIETIC_MCP_WORKSPACE=$HOME \
  --env POIETIC_BIN=$HOME/.swiftpm/bin/poietic \
  -- /opt/homebrew/bin/node /absolute/path/to/poietic-mcp/dist/index.js

Verify with claude mcp list (should show poietic: ... - ✔ Connected) or /mcp inside a session. --scope user makes it available in all projects; use --scope project to register it only for one repository (writes to that repo's .mcp.json).

Add the server to ~/Library/Application Support/Claude/claude_desktop_config.json (create the mcpServers key if absent), then fully quit and reopen the app (Cmd+Q; MCP servers only start at launch):

{
  "mcpServers": {
    "poietic": {
      "command": "/opt/homebrew/bin/node",
      "args": ["/absolute/path/to/poietic-mcp/dist/index.js"],
      "env": {
        "POIETIC_MCP_WORKSPACE": "/Users/you",
        "POIETIC_BIN": "/Users/you/.swiftpm/bin/poietic"
      }
    }
  }
}

The poietic tools appear under the tools icon in a chat.

Register the same command in the client's MCP settings (e.g. .mcp.json in the workspace for Droid, project config for other clients):

{
  "mcpServers": {
    "poietic": {
      "command": "node",
      "args": ["/absolute/path/to/poietic-mcp/dist/index.js"],
      "env": {
        "POIETIC_MCP_WORKSPACE": "/Users/you/Documents/poietic-designs",
        "POIETIC_BIN": "/Users/you/.swiftpm/bin/poietic"
      }
    }
  }
}

From a terminal, drive the server directly like an MCP client would:

node src/test/smoke.mjs   # registers, builds a tiny model, simulates

You do not invoke tools by name; you talk, and the agent picks the poietic_* tools as the conversation turns to building models.

A typical build request:

Create a new Poietic design called fishing.poietic in ~/Documents/poietic-designs and model the fishing economy from Thinking in Systems: a fleet (capital) that grows 5% per year, fish that regenerate as a function of fish density, harvest yield falling as fish get scarce, price rising as fish get scarce. Validate and simulate 200 years.

A typical inspect-and-modify request:

Open Capital.poietic in ~/Downloads/_Converted/ThinkingInSystems, show me the structure, then double the initial resource and re-run the simulation.

Working alongside Poietic Playground: open the same design file in the playground to see and edit the agent's work by hand. After saving in the playground, the agent's next write is refused with a fingerprint conflict; tell it to re-read the design (poietic_get_design) and continue. See PLAYGROUND.md for the handoff protocol.

The guided recipe: ask the agent to use the build_stock_and_flow_model prompt for a description of a system. It encodes node roles (Stock vs FlowRate vs Auxiliary), Flow vs Parameter edge semantics, formula syntax, and the validate-then-run ordering - including two lessons learned the hard way:

• Node names can contain spaces. In formulas, wrap those names in curly braces, for example {account balance} * {interest rate}. Name matching is case-insensitive, and spaces and underscores normalize to the same name.
• poietic_auto_parameters does not wire Parameter edges into or out of GraphicalFunction nodes. After auto-parameters, validate; for each issue, connect the missing edge explicitly and re-validate.
• A stated growth-rate goal is usually a net rate. An investment goal of g% of the stock with d% depreciation yields net growth of (g - d)%; setting g = d freezes the stock. Gross goal = depreciation + desired growth.

Tips:

• Mention the design's folder so the agent stays inside the workspace fence.
• Ask for poietic_validate before simulating; the agent will usually do it unprompted, but a broken model fails in poietic_run with the same issues.
• Ask for results as summaries, not raw dumps - poietic_run already returns a per-variable summary with capped sample rows, and can write the full CSV to a file if you give it an outputPath.

Session: poietic_use_design binds the current design (and snapshots a content-hash fingerprint as the session baseline); every other tool takes an optional design override.

Resources: poietic://metamodel, poietic://builtins (formula operators, functions and variables; poietic-tool v0.8+), poietic://design/current, poietic://design/current/diagram.svg.

For a v0.9 Simulation settings object, use start_time, time_step, and final_time (replacing initial_time, time_delta, and end_time). solver_type accepts euler or the fixed rk4 solver. flow_scaling accepts outflow_first or inflow_first; set it when competing flows may exceed available stock. The steps run option is deprecated, so prefer finalTime.

Prompt: build_stock_and_flow_model - encodes node roles, Flow vs Parameter edge semantics, formula syntax, and the validate-then-run ordering.

• Argv only. The CLI is spawned with an argv array; formulas containing *, spaces or parentheses are never interpreted by a shell.
• Per-design mutex. Every CLI invocation is a whole-file read-modify-write; concurrent calls are serialized per design.
• Fingerprint guard. After the first bind, any external change to the design file (e.g. the human saving in the playground) blocks writes with a clear error until the agent re-reads the design. Conflicts are detected, never merged.
• Path allowlist. Designs and outputs must be inside POIETIC_MCP_WORKSPACE.
• Verbatim errors. Non-zero CLI exits surface the CLI's stderr unchanged.
• No silent nulls. A create whose output cannot be parsed for the new object ID is a hard error.

npm run build        # tsc
npm test             # parser fixtures + integration (needs the poietic binary)
node src/test/smoke.mjs   # drives the built server over stdio
node src/test/from-description.mjs   # builds a model from prose only

fixtures/raw/ holds recorded stdout/stderr from the real CLI (see fixtures/FINDINGS.md for observed behavior and deviations from the docs); fixtures/capture.sh reproduces them. src/test/capital.test.mjs rebuilds the Thinking in Systems Capital model through the MCP tools and asserts the rebuilt simulation reproduces the committed original variable by variable.

• poietic binary not found - build and install poietic-tool, or point POIETIC_BIN at the binary.
• Path is outside the allowed workspace - the design path is not under POIETIC_MCP_WORKSPACE; move the file, or update the server's env in the client config (restart the client afterwards).
• The design changed outside this session - the file was saved externally (playground, another agent). Re-read it with poietic_get_design, then retry the write.
• Validation issues about graphical functions - run poietic_auto_parameters, then connect the missing GraphicalFunction edges explicitly (see the from-prose lessons under "Using it").
• Claude Desktop shows no poietic tools - the app loads MCP servers only at launch; quit fully (Cmd+Q) and reopen. Logs: ~/Library/Logs/Claude/mcp*.log.

• No live playground drive. The human reopens the file; the playground has no reload-on-change.
• Concurrent editing is detected, not merged. The fingerprint guard refuses the write; re-read and re-apply.
• Multi-step agent edits are not atomic. Each tool call is one transaction; one undo reverses one call. (poietic_set_attributes with several attributes is a single transaction on poietic-tool v0.8+; older builds create one entry per attribute.)
• Duplicate object names resolve arbitrarily upstream, so name-based references are only safe when uniqueness is known. Prefer numeric IDs.