# impact-preview [Health: Active]

**Category:** 💻 Developer Tools  
**Repository:** https://github.com/agent-polis/impact-preview  
**GitHub Stars:** 1  
**Views:** 3  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/impact-preview

## Description
Impact preview for AI agents - see what changes before any action executes.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "impact-preview": {
    "command": "uvx",
    "args": ["impact-preview"],
    "env": {
      "SECRET_KEY": "",
      "DATABASE_URL": "",
      "REDIS_URL": ""
    }
  }
}
```

**Requires environment variables:** `SECRET_KEY`, `DATABASE_URL`, `REDIS_URL` — the values above are empty placeholders; fill in real credentials before running (see the repository for what each one is for).

## Documentation

## What impact-preview MCP server does

The impact-preview MCP server adds a review step between an AI agent's proposed operation and its execution. It is intended for workflows where an agent can modify files, run commands, or otherwise perform actions that need human oversight. A proposed action is analyzed, a preview is generated, and a person can approve, reject, or modify it before the operation proceeds.

The MCP quick-start server focuses on four tools: previewing file writes, previewing file deletions, checking shell commands for dangerous behavior, and checking the risk associated with a path. File previews can show the expected diff, while risk analysis identifies concerns such as production data or system-file access.

## How it works

An agent submits or proposes an action. The service evaluates its likely impact and assigns a risk level, ranging from low to critical. The result can include a file diff, warnings, and machine-readable governance information. A human then reviews the result and chooses whether the action should continue.

The wider workflow follows the same pattern through the HTTP API: register an agent, submit an action, retrieve its preview or diff, and approve or reject it. Approved operations can then be executed. Events for proposed and executed actions are retained in an event-sourced audit trail, and preview events can include policy decisions, matched rule identifiers, scanner reason identifiers, and maximum severity.

## Setup and configuration

Install the Python package with `pip install impact-preview`, then start the MCP service with `impact-preview-mcp`. The README shows a local endpoint at `http://localhost:8000/mcp` for Claude Desktop configuration. Cursor is also listed as a supported way to try the MCP server.

The full service can be started with Docker Compose or with the `impact-preview` command. Its documented configuration includes `SECRET_KEY`, `DATABASE_URL`, and `REDIS_URL`; optional settings include `FREE_TIER_ACTIONS_PER_MONTH` and `LOG_LEVEL`. The dashboard requires the additional UI installation and is launched with Streamlit.

## Tools and capabilities

The impact-preview MCP server exposes these documented preview capabilities:

- Preview file content changes before a write.
- Show what would be lost before deleting a file.
- Flag potentially dangerous shell commands.
- Check the risk of a path.

Beyond MCP, the package includes an `AgentPolisClient` SDK and a `require_approval` decorator. The decorator submits a wrapped operation, blocks while approval is pending, and executes it only after approval. The API supports file, database, HTTP, shell, and custom action types, although the roadmap identifies database and API-call preview as planned work rather than current MCP functionality.

## Limitations and notes

The project is currently marked at version 0.2.0, with file-operation preview listed as the current focus. Database-operation preview, API-call preview, IDE integrations, and production readiness appear in the roadmap as future milestones. The README does not document integrations beyond Claude Desktop and Cursor, and it does not establish that every listed API action type is already implemented in the MCP server.

The service is designed around human approval, so it is not a replacement for permissions, backups, sandboxing, or other execution controls. Its risk assessment can identify documented high-risk categories, but the material does not state that it guarantees detection of every unsafe action.

_Full upstream README: https://allmcps.com/mcp/impact-preview/readme_