Quickstart

Five minutes. One query. Any LLM provider you already use.

Research Preview — RLMX is experimental. The RLM algorithm is from a recent paper. Things will change. Tell us on Discord.

Install RLMX

npm install -g rlmx

Verify the install:

rlmx --version

RLMX requires Node.js >= 18 and Python 3.10+ (for the REPL subprocess). Make sure both are available on your PATH.

Set your API key

RLMX uses Google Gemini by default. Set your API key:

rlmx config set GEMINI_API_KEY your-api-key-here

rlmx config set GEMINI_API_KEY your-key

rlmx config set ANTHROPIC_API_KEY your-key
rlmx config set model.provider anthropic
rlmx config set model.model claude-sonnet-4-20250514

rlmx config set OPENAI_API_KEY your-key
rlmx config set model.provider openai
rlmx config set model.model gpt-4o

Settings are stored at ~/.rlmx/settings.json with 0600 permissions.

Initialize a project

Navigate to your project directory and scaffold a config:

rlmx init

This creates an rlmx.yaml with sensible defaults and inline comments explaining every option. You can skip this step — RLMX auto-scaffolds on first query if no config exists.

Run your first query

Point RLMX at some context and ask a question:

rlmx "How does authentication work?" --context ./src/

RLMX loads your source files into a Python REPL, then iterates — writing Python code to navigate the context, executing it, and refining until it calls FINAL() with the answer.

$ rlmx "How does authentication work?" --context ./src/

Authentication uses JWT tokens issued by the /auth/login endpoint.
The middleware in src/middleware/auth.ts validates tokens on every
request and attaches the decoded user to req.user. Refresh tokens
are stored in HttpOnly cookies with a 7-day expiry.

$ rlmx "How does authentication work?" --context ./src/ --output json

{
  "answer": "Authentication uses JWT tokens issued by the /auth/login endpoint...",
  "references": ["src/middleware/auth.ts", "src/routes/auth.ts", "src/utils/jwt.ts"],
  "usage": { "inputTokens": 12500, "outputTokens": 3200, "llmCalls": 5 },
  "iterations": 3,
  "model": "google/gemini-3.1-flash-lite-preview"
}

$ rlmx "How does authentication work?" --context ./src/ --verbose

[iter 1] LLM wrote 12 lines of Python... executed in 0.3s
[iter 2] LLM wrote 8 lines of Python... executed in 0.2s
[iter 3] FINAL() called — answer ready

Authentication uses JWT tokens issued by the /auth/login endpoint...

Try different input modes

RLMX handles directories, files, and piped input:

# Directory of docs (recursively loads .md files by default)
rlmx "Summarize the API" --context ./docs/

# Single file
rlmx "What are the key findings?" --context paper.md

# Piped data
cat data.csv | rlmx "Analyze this dataset"

# Custom file extensions
rlmx "Review this code" --context ./src/ --ext .ts,.js,.py

Output modes

Mode	Flag	Description
text	`--output text` (default)	Plain text answer to stdout
json	`--output json`	Structured JSON with answer, references, usage stats
stream	`--output stream`	JSONL events per iteration, then a final event

What just happened?

Under the hood, RLMX:

Loaded your files into a persistent Python subprocess as the context variable
Sent the LLM a system prompt with metadata about the context (not the content itself)
The LLM wrote Python code to search, filter, and read specific parts of the context
RLMX executed each code block and fed the results back
After a few iterations, the LLM called FINAL() with its answer

This is the RLM algorithm — the LLM navigates your code programmatically instead of having it stuffed into the prompt. It’s cheaper and scales to larger codebases.

Next steps

CLI Reference

Every command and flag documented.

Configuration

Customize model, tools, caching, and budget limits.

Batch Mode

Run hundreds of questions against cached context.

Stuck? Ask on Discord

Real humans. Real answers.

​Quickstart

​Output modes

​What just happened?

​Next steps

CLI Reference

Configuration

Batch Mode

Stuck? Ask on Discord

Quickstart

Output modes

What just happened?

Next steps