RuleDrift
Local test suite that checks whether AI agents keep following rules in long chats
- Pricing
- Open source
- Platforms
- CLI, self-hosted, API
- Runs
- Locally
- Version
- 0.1.1
- Listing updated
- today
Similar MCP servers
About RuleDrift
RuleDrift is an open-source MCP server for agent benchmarks. It runs locally on your machine and is started by your AI client.
The README says RuleDrift replays conversations to find the exact turn where an agent stops honoring an important instruction, which tends to happen once a session grows long or gets distracted. It is meant to run like a regression test whenever a prompt, model, memory layer or agent workflow changes.
It runs on your own machine and uses a local Ollama model by default, so no paid model API, account or telemetry is needed. It can also be pointed at your own agent, and it sits beside production traffic rather than proxying it. Reports come out in terminal, JSON and JUnit formats, with stable exit codes for CI pipelines.
An optional MCP mode lets AI coding clients run the same checks from inside their editor or chat session.
Install
Run in a terminal
claude mcp add rule-drift -- npx -y rule-driftAdd to claude_desktop_config.json
{
"mcpServers": {
"rule-drift": {
"command": "npx",
"args": [
"-y",
"rule-drift"
]
}
}
}Add to ~/.cursor/mcp.json
{
"mcpServers": {
"rule-drift": {
"command": "npx",
"args": [
"-y",
"rule-drift"
]
}
}
}Add to .vscode/mcp.json
{
"servers": {
"rule-drift": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"rule-drift"
]
}
}
}Add to ~/.codeium/windsurf/mcp_config.json
{
"mcpServers": {
"rule-drift": {
"command": "npx",
"args": [
"-y",
"rule-drift"
]
}
}
}Add to your client's MCP settings (most use this shape)
{
"mcpServers": {
"rule-drift": {
"command": "npx",
"args": [
"-y",
"rule-drift"
]
}
}
}Using several servers? Build one config file with the MCP config generator.
What it does
- Drift point detection
- Pinpoints the first conversation turn at which a stated rule is broken.
- Local Ollama runs
- Uses a locally hosted Ollama model, so no paid API key is required.
- Bring-your-own agent
- Can test an agent you already run instead of only the bundled setup.
- Multiple report formats
- Outputs terminal, JSON and JUnit reports.
- CI-friendly exit codes
- Returns stable exit codes so test runs can gate a pipeline.
- Optional MCP tools
- Exposes its checks to AI coding clients over the Model Context Protocol.
Good for
- Regression-testing an agent after changing its system prompt
- Checking whether a new model still respects core instructions over long sessions
- Verifying a memory or context-management change does not cause rules to be forgotten
- Adding instruction-retention tests to a CI pipeline
Pricing
Open source
Free and MIT-licensed; runs locally.
RuleDrift FAQ
What do I need to run it?
The README says it requires Node.js 22 or newer and Ollama, with a local model such as qwen3:1.7b pulled.
Does it need a paid API or account?
The README states that no paid model API, account, backend or telemetry is required.
Does it sit in front of production requests?
No. The README says it runs locally and does not sit in front of production requests.
How is it installed as an MCP server?
The README gives the command npx -y rule-drift mcp --root with a path to your project, and points to a separate MCP setup guide for client configuration.
Is RuleDrift free?
Yes. RuleDrift is open source, so you can use and self-host it for free.
How do I add RuleDrift to Claude Code?
Run this command in your terminal: claude mcp add rule-drift -- npx -y rule-drift
What are the best alternatives to RuleDrift?
The closest alternatives to RuleDrift are Agent Reliability, First Dollar Test, ScoreIA Open Chamber, mizan-ledger and ICONIC — MCP Selection Test for MCP publishers.
Where this comes from
- Description: written from the project's README and its MCP Registry entry.
- Install steps: from the official MCP Registry entry.

