MCP Server and Agent Skills¶

The same validated operations the command line offers are available to an AI agent over the Model Context Protocol, and five agent skills describe the common workflows.

The MCP server¶

impact-model mcp --root ./impact-models

The server speaks MCP over standard input and output. Configure an MCP client to run that command; for example, in a client that takes a JSON configuration:

{
  "mcpServers": {
    "impact-model": {
      "command": "impact-model",
      "args": ["mcp", "--root", "./impact-models"]
    }
  }
}

It makes no network calls and no model calls.

Tools¶

Tool	What it does
`impact_schema`	List the embedded JSON Schemas or return one
`impact_validate`	Validate an impact model
`impact_lint`	Lint an impact model with a profile
`impact_graph`	Query the causal graph
`impact_status`	Compute a status snapshot from observations
`impact_report`	Render a definition or status report
`impact_deck`	Generate a deck as Marp Markdown
`impact_export`	Export for the PRISM tools as JSON: `okr`, `journey`, `v2mom`, or `dmaic`. With observations, the export carries the latest values and status. A model that lacks what the format needs is refused, naming what to add
`impact_diff`	Compare two versions of an impact model
`impact_judge_bundle`	Build the judge bundle
`impact_patch`	Apply a JSON Patch, atomically

A tool takes an impact model inline (the document argument) or by path. impact_patch rejects any patch that adds validation errors, exactly as the patch command does, so an agent cannot leave an impact model worse than it found it.

Where an agent can read and write¶

A path must lie inside --root, which defaults to the current directory. These are refused with the same answer on every platform: a path with a .. element, a rooted or drive path outside the root, and a symbolic link that leads outside the root. An inline document is limited in size.

Agent skills¶

A skill is a Markdown file that tells an agent how to carry out one task with the commands above. They live in the repository's skills/ directory.

Skill	Use it to
`impact-draft-from-sources`	Draft an impact model for an organization from its public pages and reports. Every source is recorded as evidence, and every element is labelled stated, inferred, or hypothesis
`impact-author`	Add, change, or remove outcomes, links, assumptions, indicators, actors, and other elements with validated edits
`impact-status-refresh`	Load a new period's numbers from a spreadsheet, then report where the impact model stands and what changed
`impact-learning-review`	Prepare and record a learning review for a period
`impact-audience-story`	Draft a short update for one audience from the impact model's numbers and consented stories

Each skill is tested: the commands and flags it names must exist, and its sequence of steps is run against the examples to check it works and repeats.

Next¶

Examples · CLI reference