closed-loop-cli

A self-developing, multi-agent CLI coding assistant. It runs an agentic read → act → verify loop over your codebase using file, search, and shell tools, and works with any Anthropic-compatible endpoint — Anthropic itself, or a gateway/proxy serving another model (OpenRouter, LiteLLM, MiMo Token Plan, a local proxy, …).

Status: early/alpha. The core loop, tooling, permission system, and multi-model support work; expect rough edges.

Features

• Agentic loop with file read/write/edit, directory listing, grep, a local AST code index, web search, and sandboxed shell execution.
• Read-before-edit safety — a file must be read before it can be edited, and edits are rejected if the file changed on disk since it was read.
• Permission system — default / acceptEdits / plan / yolo modes with per-session "always allow" choices, on top of hard safety rules (RuleZ).
• Parallel tools — independent read-only tool calls run concurrently.
• Subagents — delegate self-contained sub-tasks to a fresh context.
• Automatic context compaction — long histories are summarized to stay in budget.
• Adaptive output budget and retry with backoff on transient API errors.
• Model-agnostic — works with any Anthropic-compatible model/endpoint via standard Claude Code environment variables; not tied to any single provider.
• Polished interactive REPL — clean banner, Markdown-rendered answers, concise progress (no wall of logs), multi-line paste, in-line cursor editing (←/→/Home/End/Delete), and ↑/↓ command history.
• Runs in your project folder — like claude, it operates in the directory you launch it from and creates files there.

Install

npm install -g closed-loop-cli

Or run without installing:

npx closed-loop-cli

Quick start

Set credentials (the CLI also prompts for and saves a key to ~/.closed-loop.json if none is found):

export ANTHROPIC_API_KEY="sk-ant-..."
closed-loop

With no model configured it defaults to claude-haiku-4-5, so a standard Anthropic key works out of the box. The CLI operates in your current working directory — cd into a project (or an empty folder) and run closed-loop there. Type your task at the prompt; paste multi-line prompts freely (they submit only on Enter); ↑/↓ recall history; type help, clear, or exit.

Configuration (environment variables)

Model selection follows the same conventions as Claude Code, so an existing Claude Code shell setup works unchanged.

| Variable | Purpose | | --- | --- | | ANTHROPIC_API_KEY / ANTHROPIC_AUTH_TOKEN | Credentials | | ANTHROPIC_BASE_URL | Point at any Anthropic-compatible endpoint | | ANTHROPIC_MODEL | Main model (also falls back to ANTHROPIC_DEFAULT_OPUS_MODEL, then ANTHROPIC_DEFAULT_SONNET_MODEL) | | CLAUDE_CODE_SUBAGENT_MODEL | Subagent model (also falls back to ANTHROPIC_DEFAULT_HAIKU_MODEL) | | ANTHROPIC_MAX_TOKENS | Override output token budget | | CLAUDE_CODE_EFFORT_LEVEL | max / ultracode raises step + token budget | | CLOSED_LOOP_PERMISSION_MODE | default / acceptEdits / plan / yolo | | CONTEXT_COMPACT_THRESHOLD_TOKENS | When to summarize history (default 120000) | | CONTEXT_KEEP_RECENT_MESSAGES | Recent messages kept verbatim on compaction |

Works with any model

The CLI talks to whatever ANTHROPIC_BASE_URL + model you configure — Anthropic directly, or any Anthropic-compatible gateway (OpenRouter, LiteLLM, a local proxy, MiMo Token Plan, …). To use Anthropic, just set ANTHROPIC_API_KEY. To use any other provider, point ANTHROPIC_BASE_URL at its compatible endpoint and set the model id. No code changes required.

Example: a non-Claude provider (MiMo Token Plan)

# Anthropic-compatibility endpoint + token
$env:ANTHROPIC_BASE_URL   = "https://token-plan-sgp.xiaomimimo.com/anthropic"
$env:ANTHROPIC_AUTH_TOKEN = "tp-...your-token..."

# Main model. mimo-v2.5-pro has a 1M context window natively.
$env:ANTHROPIC_MODEL              = "mimo-v2.5-pro"
$env:ANTHROPIC_DEFAULT_OPUS_MODEL = "mimo-v2.5-pro"

# Cheaper model for subagents (mimo-v2.5 is a deep-thinking, lower-cost model)
$env:CLAUDE_CODE_SUBAGENT_MODEL    = "mimo-v2.5"
$env:ANTHROPIC_DEFAULT_HAIKU_MODEL = "mimo-v2.5"

$env:CLAUDE_CODE_EFFORT_LEVEL = "max"

closed-loop

Note on the [1m] suffix: Claude Code uses a model[1m] notation to opt into a 1M-context beta. That suffix is not part of the model id a raw API accepts, so the CLI strips any trailing [...] automatically — a pasted Claude Code config like mimo-v2.5-pro[1m] works, and resolves to mimo-v2.5-pro.

CLI flags

| Flag | Effect | | --- | --- | | --plan | Read-only; no edits or commands | | --accept-edits | Auto-approve file edits (commands still prompt) | | --yolo | Auto-approve everything (RuleZ hard rules still apply) | | -e, --effort <standard\|ultracode\|max> | Reasoning/step budget | | --codeact / --openhands | CodeAct-style coder profile | | --dashboard | Launch the monitoring dashboard server | | --telegram | Run the Telegram ChatOps bot |

Equivalent npm scripts when running from source: npm start, npm run dashboard, npm run telegram, npm run dgm.

Permission modes

• default — read-only tools run automatically; file edits and shell commands ask for confirmation (with an "always allow" option for the session).
• acceptEdits — file edits auto-run; shell commands still ask.
• plan — read-only only; mutating tools are denied.
• yolo — everything auto-runs. Use with care.

In a non-interactive session, actions that would prompt are denied unless the mode auto-approves them (acceptEdits / yolo).

Build from source

git clone <repo>
cd closed-loop
npm install
npm run build
npm start

Requires Node.js 18+ and a TypeScript toolchain (installed via devDependencies).

License

See repository.