rlm-cli

CLI wrapper for rlm with directory-as-context, JSON-first output, and self-documenting commands.

Upstream RLM: https://github.com/alexzhang13/rlm

Install

One-liner (recommended)

curl -sSL https://raw.githubusercontent.com/rawwerks/rlm-cli/master/install.sh | bash

This clones the repo to ~/.local/share/rlm-cli and symlinks rlm to ~/.local/bin/.

To uninstall:

curl -sSL https://raw.githubusercontent.com/rawwerks/rlm-cli/master/uninstall.sh | bash

uvx (no install)

Run directly without installing:

uvx --from git+https://github.com/rawwerks/rlm-cli.git rlm --help

pipx

pipx install git+https://github.com/rawwerks/rlm-cli.git

Development install

git clone --recurse-submodules https://github.com/rawwerks/rlm-cli.git
cd rlm-cli
uv venv
uv pip install -e .

Claude Code Plugin

This repo includes a Claude Code plugin with an rlm skill. The skill teaches Claude how to use the rlm CLI for code analysis, diff reviews, and codebase exploration.

Installation

Claude Code (Interactive)

/plugin marketplace add rawwerks/rlm-cli
/plugin install rlm@rlm-cli

Claude CLI

claude plugin marketplace add rawwerks/rlm-cli
claude plugin install rlm@rlm-cli

What the skill provides

The /rlm skill gives Claude knowledge of:

• All rlm commands (ask, complete, search, index, doctor)
• Input types (files, directories, URLs, stdin, literal text)
• Common workflows (diff review, codebase analysis, search + analyze)
• Configuration and environment variables
• Exit codes for error handling

Once installed, Claude can use rlm to analyze code, review diffs, and explore codebases when you ask it to.

Authentication

Authentication depends on the backend you choose:

• openrouter: OPENROUTER_API_KEY
• openai: OPENAI_API_KEY
• anthropic: ANTHROPIC_API_KEY

Export the appropriate key in your shell environment, for example:

export OPENROUTER_API_KEY=sk-or-...

Usage

Ask about a repo

rlm ask . -q "Summarize this repo" --json

Ask about a URL (auto-Markdown)

rlm ask https://www.anthropic.com/constitution -q "Summarize this page" --json

Same with uvx and OpenRouter:

uvx --from git+https://github.com/rawwerks/rlm-cli.git rlm ask https://www.anthropic.com/constitution -q "Summarize Claude's constitution" --backend openrouter --model google/gemini-3-flash-preview --json

Ask about a file

rlm ask src/rlm_cli/cli.py -q "Explain the CLI flow" --json

Use stdin as context

git diff | rlm ask - -q "Review this diff" --json

No context, just a completion

rlm complete "Write a commit message" --json

OpenRouter quickstart

rlm complete "Say hello" --backend openrouter --model z-ai/glm-4.7:turbo --json

Options

• --json outputs JSON only on stdout.
• --output-format text|json sets output format.
• --backend, --model, --environment control the RLM backend.
• --max-iterations N sets max REPL iterations (default: 30).
• --max-depth N enables recursive RLM calls (default: 1, no recursion).
• --max-budget N.NN limits spending in USD (requires cost-tracking backend like OpenRouter).
• --backend-arg/--env-arg/--rlm-arg KEY=VALUE pass extra kwargs.
• --backend-json/--env-json/--rlm-json @file.json merge JSON kwargs.
• --literal treats inputs as literal text; --path forces filesystem paths.
• --markitdown/--no-markitdown toggles URL and non-text conversion to Markdown.
• --verbose or --debug enables verbose backend logging.
• --inject-file FILE executes Python code between iterations (update variables mid-run).

Early Exit and Cancellation

Ctrl+C (Reply Now)

Pressing Ctrl+C during execution returns the best partial answer as success (exit code 0) instead of raising an error. This is useful when you want to stop waiting but keep what the LLM has produced so far.

rlm ask . -q "Analyze in detail" --max-iterations 20
# Press Ctrl+C after a few iterations
# Output: partial answer with exit_code=0, early_exit=true

In JSON mode, the result includes early_exit and early_exit_reason fields:

{"ok": true, "result": {"response": "...", "early_exit": true, "early_exit_reason": "user_cancelled"}}

SIGUSR1 (Programmatic Early Exit)

Send SIGUSR1 to request graceful early exit without using Ctrl+C:

# In another terminal
kill -SIGUSR1 <rlm_pid>

This is useful for programmatic control over long-running RLM tasks.

--inject-file (Update Variables Mid-Run)

The --inject-file option executes Python code between iterations, allowing you to update REPL variables while the RLM is running.

# Create inject file
echo 'focus = "authentication"' > inject.py

# Start RLM with inject file
rlm ask . -q "Analyze based on the 'focus' variable" --inject-file inject.py

# In another terminal, upd

...