#Rein

Turn intent into CLI commands. Rein learns your tools, plans commands, checks their risk, and iterates on the output.

Use the open-source CLI on your machine, or connect to Rein Enterprise for organization policies, approval workflows, and audit visibility through Rein Control.

rein in git "what changed in the last two commits, and who wrote them?"
rein in kubectl "which pods are crashlooping in staging?"
rein in gh "how many open PRs are assigned to me?"

Rein discovers commands and flags from the installed tool, so it can work with internal CLIs a model has never seen. You see each proposed command and its output; commands classified as mutating or destructive ask for approval by default.

#Open source and Enterprise

The CLI works without a Rein Control account. Enterprise connects that execution layer to shared controls for teams and their agents.

	Open-source CLI	Enterprise through Rein Control
Tool discovery	Local capability maps and caching	Uses the same CLI foundation
Models	Agent CLIs, hosted APIs, and local endpoints	Uses the same backend choices
Execution safeguards	Risk classification and local approval prompts	Organization policies for MCP operations
Agent integration	MCP server	Named caller registration and team profiles
Approvals	Terminal prompts and MCP approval ceilings	Central approval requests for registered MCP callers
Run visibility	Local transcripts and archives	Central audit events for registered MCP callers

This repository contains the CLI and its Cloud integration. The enterprise control plane is a separate service. See Rein Enterprise and the enterprise setup guide for the connection flow and current integration limits.

#Quickstart

You need Go 1.27+, the tool you want to run on your PATH, and a model backend. The default backend uses an installed, signed-in Claude Code CLI.

go install github.com/jordandalton/rein/cmd/rein@latest
export PATH="$HOME/go/bin:$PATH"
rein in git "what changed in the last commit?"

The install directory is $(go env GOPATH)/bin, usually ~/go/bin. Add it to your shell profile if needed. Run the Git example inside a Git repository.

Already use another supported agent CLI? Select it explicitly:

rein in --backend codex-cli git "what changed in the last commit?"
rein in --backend grok-cli git "what changed in the last commit?"

For a local model, start Ollama with your chosen model available:

rein in --backend ollama --model qwen2.5 git "summarize recent work"

Hosted APIs and custom endpoints are supported too. See model backends and configuration. Model-provider charges may apply; a Rein Control account is not required for local use.

#Everyday use

rein git "summarize recent work"                    # `in` is optional
rein in --dry-run git "create a branch called demo" # preview the first command
rein spec kubectl                                  # discover commands ahead of time
rein spec kubectl --show                           # inspect the capability map
rein list                                         # list learned tools
rein in --refresh gh "list my open pull requests"   # rediscover after an upgrade

Discovery runs automatically on first use and is cached by tool version. Subsequent runs reuse the map; planning and execution still take time. See the CLI usage guide for flag placement and approval options.

#Use with an agent

Rein exposes tool discovery and execution over MCP. For a registered Control agent, run guided setup to install the persistent Gateway bridge:

rein configure codex
# or: rein configure claude-code

For a standalone local server, configure your MCP client to launch rein mcp:

{ "mcpServers": { "rein": { "command": "rein", "args": ["mcp"] } } }

The default server permits only commands classified as read-only. Operations requiring higher approval return a structured outcome to the calling agent. See MCP setup and approval levels.

#Connect your organization

For an existing Rein Control organization, enroll your installation and download its policy bundle:

rein login
rein status
rein sync

On macOS, cloud credentials use Keychain. On Linux and other Unix hosts, Rein stores them in ~/.rein/credentials.json with mode 0600; protect that account and home directory accordingly. For named MCP callers and central approvals, follow the enterprise setup guide.

For guided project setup, run rein configure claude-code or rein configure codex in an interactive terminal. It summarizes changes and asks for confirmation, guides login if needed, registers the caller, starts the persistent local Rein Gateway, installs persistent MCP/settings with backups, and saves the project launch profile (the same operation as --apply), and verifies the MCP handshake without running tools or making model requests. If a previously installed file is missing, guided setup offers to restore it from its receipt after confirmation. Existing edited files are never overwritten, and the original undo backups are preserved. Both normal harness launches and rein configure HOST --launch are supported. To undo both, run rein undo HOST. A conflicting saved launch profile is not replaced automatically; the error points to that same command before changing its settings. If verification fails, setup reports failure and leaves installed settings recoverable with rein undo claude-code. The installed MCP bridge uses rein gateway connect --agent HOST; agent sessions share the gateway daemon instead of launching a separate Rein MCP runtime. Manage it with rein gateway status, rein gateway stop, and rein gateway start. Then start claude, trust the project settings/MCP, inspect /mcp and /hooks, and verify native-tool blocking. For Codex, start codex instead; setup checks hook support and installs .codex/config.toml. Codex coverage remains partial: trust the project configuration and hooks, and verify restrictions in-session. Connectivity is not proof of runtime enforcement.

To require cached specs for your intended tools during guided setup, use rein configure codex --require-spec cat,git (also supported for claude-code). Add --persistent --check for a noninteractive configuration/spec check. Missing specs fail with an explicit trusted-operator rein spec TOOL instruction; setup never discovers tools automatically. Governed MCP records policy/spec and authorization blocks as non-executed Activity events. If audit delivery fails, the operation stays blocked and the error reports that Activity could not be updated.

Alternatively, preview and save a gateway-backed restricted launch profile without rewriting existing Claude Code or Codex settings:

rein gateway start
rein agent register claude-code
rein configure claude-code --gateway --dry-run
rein configure claude-code --gateway --apply
rein configure claude-code --launch

See harness configuration, coverage limits, and undo. See also the runtime validation results and test command. Restrictions apply only to this launcher; Codex coverage is explicitly partial.

For persistent project settings on macOS/Linux, use rein configure claude-code --gateway --persistent --dry-run, then --gateway --persistent --apply after starting the Gateway. Replace claude-code with codex for Codex. This installs a default-deny tool hook plus Rein MCP registration for normal project launches, with backups, --persistent --check, and --persistent --undo. Review/trust the hooks in a fresh session and run the guide's bypass tests before rollout. Codex coverage remains partial; editable project hooks are not an administrator-enforced boundary.

#GitHub Actions (beta)

Use the GitHub Actions guide for OIDC workload authentication and policy-governed commands through rein ci check and rein ci run -- TOOL ARGS. Includes an authentication action, pinned-version workflow template, approvals, and audit reporting. This is not runner isolation.

#Safety and data handling

Rein executes argument arrays directly and checks the target binary on every step. A model assessment and a static classifier determine the approval level; the higher risk assessment wins.

Local approval mode	Classified read-only	Mutating	Destructive
Default	Runs	Asks	Asks
`--yes`	Runs	Runs	Asks
`--auto`	Runs	Runs	Runs

Execution happens on your machine. Intent, tool capabilities, and redacted command observations go to your selected model backend. Connected, registered MCP callers also send audit and approval metadata to Rein Control.

Risk classification and redaction are heuristic. Rein is not an OS sandbox, and controls apply to operations routed through it. Interactive tools requiring a PTY are not supported. Read safety, data flow, and limitations before configuring unattended execution.

#Documentation

#Contributing and license

Bug reports, CLI compatibility examples, and pull requests are welcome. See Contributing for local build and validation commands.

A license file has not yet been added to this repository. Explicit open-source license terms still need to be published.