convez.ai
AgenteAutomaçãoAvançado
daanielcruz

Otherside CLI

daanielcruz/otherside-cli

Agent terminal multi-provider com controle total, privacidade e suporte a orquestração de LLMs

Ver no GitHub
BomPluginScore · de 0 a 100

Agent de codificação em terminal com suporte multi-provider (Anthropic, Codex, Gemini, DeepSeek, modelos self-hosted). Executa como binário Bun único com interface interativa, workflows automáticos e orquestração de múltiplos LLMs na mesma sessão, sem telemetria.

Instalação

$npx convez-ai daanielcruz/otherside-cli

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
14
Forks
0
Acompanhando
2
Saúde
57%
Último commit
há 4 semanas
Versão
v1.2.1
Issues abertas
0
Contribuidores
1

Casos de uso

  • Automatizar tarefas de codificação em terminal com múltiplos modelos de IA
  • Executar workflows de desenvolvimento rooteados localmente com modelos self-hosted
  • Orquestrar LLMs por tier de capacidade (emperor/shogun/daimyo/samurai) em projetos complexos

Veredito

Ferramenta sofisticada para desenvolvedores que buscam máximo controle sobre orquestração de LLMs em terminal. Excelente para fluxos privados e multi-provider, porém requer experiência técnica para explorar todo potencial.

README

A tradução para o português deste README ainda está na fila.

status typescript license

a shell for the reversed world.

Terminal coding agent for Anthropic, Anthropic Cloud, Codex, Antigravity, xAI, Kimi, DeepSeek, MiniMax, GLM (Z.AI), and self-hosted LLMs.

Privacy. Control. Provider choice.

Overview · Coming next · Multiprovider orchestration · Design · Features · Install · Build · Operation · Providers · Contributing

Overview

Otherside is an open-source terminal coding agent that runs entirely on your terms. Built as a single-executable Bun binary, it supports rich interactive workflows, terminal UI pairing, and automated script runs. Its string-view terminal engine renders rows directly, keeping live output separate from settled terminal scrollback.

With native support for top-tier AI providers and local models, Otherside allows you to orchestrate multiple LLMs within the same session. It serves as a multiprovider alternative to Claude Code, capable of running any LLM, either individually or orchestrated together, giving you full control over your data with zero telemetry.

Coming next

  • Non-verbose mode: A quieter transcript that keeps tool noise out of the way.
  • Memory usage improvements: Lower footprint on long interactive sessions.
  • Customizable agents panel: Managing agent definitions from a dedicated panel.
  • Multiprovider polish and new modes: Refinements to orchestration plus additional routing modes.
  • UI polish: Continued refinement of the terminal interface and its interactive panels.

Multiprovider orchestration

Multiprovider is disabled by default and can be enabled through /config or /multiprovider, with two modes: default (concrete provider/model routing from the full catalog) and feudalism (tier routing below).

Most agent CLIs delegate every subtask to the same model family. Otherside can route delegated Agent and Workflow work across eligible providers, while keeping the concrete provider and model visible.

Feudalism routes each delegated task through four fixed capability tiers:

  • emperor: deep reasoning for planning, synthesis, hard calls, and final review;
  • shogun: complex execution with judgment — tactical planning, multi-step implementation, review passes;
  • daimyo: fast capable execution for code edits, debugging, and iterative tool work;
  • samurai: the cheapest, fastest ranks, used only on explicit selection for mechanical fan-out.

Tier rosters, rank-ordered (rank 1 is the resolver's first choice):

TierModels
emperorclaude-opus-5-5 · gpt-6-astra
shogungrok-4.7 · k3 · gpt-5.6-terra · claude-sonnet-5-5 · glm-5.2
daimyogpt-5.6-luna · gemini-3.8-flash · gemini-3.1-pro-high · grok-composer-2.5-fast · deepseek-v4-pro · kimi-for-coding
samuraigemini-3.8-flash-low · gemini-3.8-flash-medium · glm-5-turbo · claude-haiku-5-5 · deepseek-flash · minimax-m3 · kimi-for-coding-highspeed

Routing rules:

  • Candidates are checked against credentials, quota, cooldowns, and runtime availability on every allocation.
  • An unusable model cascades to the next rank; an empty tier cascades down (emperor → shogun → daimyo). samurai is never a fallback — it only runs when explicitly selected.
  • Chain of command (default on, /config): a nested agent cannot launch above its own tier — higher requests clamp to the caller's tier.
  • A named agent without an explicit tier is inferred from intent (inference never selects samurai); explicit tier selection takes precedence over inference.
  • Workflow diversify: true can spread independent calls across distinct eligible providers in one tier.
  • /parallel, /fork, and /workflows provide direct control over concurrent and deterministic delegated work.
  • Rosters ship with releases and are never fetched at runtime.

Design

Otherside ships a design studio: run /design in the TUI and pair the session with design.othersidecli.com in your browser. The design agent authors interactive HTML artifacts (components, slide decks, prototypes, and mockups) rendered in a live preview as it works.

  • Live authoring: the agent writes and edits design files; the preview updates in real time, streamed from your local session.
  • Tweak controls: designs can expose typed controls (colors, spacing, copy, options) you can adjust instantly in the browser with no model round-trip; reset or save your changes as new defaults.
  • Visual verification: after substantive changes the agent verifies its own output with programmatic checks and screenshots before handing back.
  • Local-first: designs are created and stored on your machine; the browser is a paired viewport, and the coding agent can read saved designs back with the ReadDesign tool.

Design uses the same provider you are signed into, with no separate subscription or key.

Features

Native toolsetBuilt-in shell, search, web, read, edit, write, numbered task list, image-generation, and image-parsing tools for codebases.
Provider choiceAnthropic, Anthropic Cloud, Codex, Antigravity, xAI, Kimi, DeepSeek, MiniMax, GLM (Z.AI), and OpenAI-compatible endpoints, including local models.
Accessible terminal UIKeyboard-first interface designed for screen readers, low-bandwidth links, and high-latency connections.
Mobile companionPair an iOS or Android app over an end-to-end encrypted channel to follow sessions, read logs, and respond remotely.
Goal-driven executionUse /goal to set success criteria and continue work across compaction until the result is met and approved.
Deterministic workflowsCoordinate subagents with JavaScript workflows, including parallel, pipeline, and nested runs; manage them with /workflows.
Planning and side questions/ultraplan coordinates multi-agent planning; /btw keeps secondary questions out of the main thread.
Background workManage long-running tasks with /tasks, shell sessions with /bashes, scheduled prompts, and recurring loops with /loop.
Interactive permissionsPer-tool permission prompts with file previews and diffs; approvals are saved as rules you manage with /permissions.
Sessions and checkpointsResume saved sessions and restore code or conversation history with /rewind.
MCP and pluginsConnect external tools through MCP (/mcp) and manage extensions with /plugins.
ImagesGenerate images through Codex, Grok, or Gemini; ParseImage and Read can route images through a configurable vision provider.
Project memory and usage/dream stores project memory; /usage shows provider usage and available account information.
Data controlNo product analytics or background tracking. Your code goes from your machine to the provider you choose.
Single binaryInstall and run one executable; no Bun, Node, or source checkout is needed for released binaries.

Installation

Install uses the published release binary. You do not need Bun, Node, or a source checkout.

macOS

curl -fsSL https://othersidecli.com/install.sh | bash
  • Supports Apple Silicon and Intel Macs.
  • Installs otherside to ~/.local/bin by default.
  • If the command is not found after install, add this to ~/.zshrc or ~/.bashrc:
export PATH="$HOME/.local/bin:$PATH"

Linux

curl -fsSL https://othersidecli.com/install.sh | bash
  • Supports x64 and arm64 Linux.
  • Requires curl, uname, mkdir, and install.
  • Verifies the release checksum when sha256sum or shasum is available.
  • If ~/.local/bin is not on your PATH, add it to your shell rc:
export PATH="$HOME/.local/bin:$PATH"

Windows

Run in PowerShell:

irm https://othersidecli.com/install.ps1 | iex
  • Supports 64-bit Windows x64.
  • Installs otherside.exe to %LOCALAPPDATA%\Programs\otherside.
  • Adds the install directory to the user Path; restart PowerShell, Windows Terminal, or your IDE terminal after install.
  • Git for Windows is recommended so shell commands can use Git Bash semantics when needed.

Install options

Pin a release:

OTHERSIDE_VERSION=v1.2.1 bash -c "$(curl -fsSL https://othersidecli.com/install.sh)"
$env:OTHERSIDE_VERSION = "v1.2.1"
irm https://othersidecli.com/install.ps1 | iex

Install somewhere else:

OTHERSIDE_INSTALL_DIR=/usr/local/bin bash -c "$(curl -fsSL https://othersidecli.com/install.sh)"
$env:OTHERSIDE_INSTALL_DIR = "$env:USERPROFILE\bin"
irm https://othersidecli.com/install.ps1 | iex

Verify and launch

otherside --version
otherside

First run walks you through provider login.

Build from source

Build from source is for contributors or local changes. It requires Bun 1.4 canary (bun upgrade --canary), Rust/Cargo, and the platform's native build tools. The commands below target macOS arm64 and require Xcode Command Line Tools.

git clone https://github.com/daanielcruz/otherside-cli.git
cd otherside-cli
bun install --frozen-lockfile
bun run build
./dist/otherside --version

bun run build writes ./dist/otherside for the current configured target. The repository default targets macOS arm64 because local development happens there; releases are cross-built by CI for macOS, Linux, and Windows.

To install your locally built binary on macOS or Linux:

mkdir -p "$HOME/.local/bin"
cp ./dist/otherside "$HOME/.local/bin/otherside"
chmod +x "$HOME/.local/bin/otherside"

Or use the helper:

bun run build:install

On Windows, install Visual Studio Build Tools and Rust's MSVC toolchain, then build the microphone addon and executable together:

rustup target add x86_64-pc-windows-msvc
bun run build:win
.\dist\otherside.exe --version
$InstallDir = "$env:LOCALAPPDATA\Programs\otherside"
New-Item -ItemType Directory -Force -Path $InstallDir | Out-Null
Copy-Item .\dist\otherside.exe "$InstallDir\otherside.exe" -Force

Contributor checks:

bun run typecheck
bun run lint
bun run test

Operation

Three modes share the same harness.

# Interactive TUI (default)
otherside

# Resume a saved session for the current project
otherside --resume <session-id>
otherside -c   # most recent session in this cwd

# One-shot run (scripts, CI, cron)
otherside -p "refactor the session loader to use the shared path helper"

# Pipe stdin
cat logs.txt | otherside -p "find the stack trace that caused the segfault"

# Skip every permission prompt
otherside --yolo

Providers

Each provider has its own authentication, tool adaptation, and model catalog behind one interface.

otherside --provider deepseek
otherside --provider codex --model gpt-6-astra
otherside --provider openai --model <your-model>
ProviderSlugAuthDefault modelNotes
AnthropicanthropicOAuthclaude-opus-5-5Sonnet 5.5 and Haiku 5.5; Fable 5.1 is also available for explicit selection; subscription usage and web search.
Anthropic Cloudanthropic-cloudOAuthclaude-opus-5-5Runs Opus 5.5 in managed Cloud sessions while tools and permissions execute locally.
CodexcodexOAuthgpt-6-astraAstra uses an 872k-token context window; GPT-6.1 Sol, GPT Daybreak Blue, and GPT-5.6 Terra/Luna are also available; hosted image generation.
AntigravityantigravityOAuthgemini-3.8-flashGoogle account path; Gemini 4 Argon (Ultra accounts), Gemini 3.x Flash/Pro, Claude, and GPT-OSS models.
xAIxaiOAuthgrok-4.7Grok models through SuperGrok OAuth.
KimikimiAPI keyk3K3 and K2.7 Code models; usage limits in /usage.
DeepSeekdeepseekAPI keydeepseek-v4-proV4 Pro, V4.1 Flash, and V4 Flash; configurable vision side-channel.
MiniMaxminimaxAPI keyminimax-m2.7M2.7 and M3 models.
GLM (Z.AI)glmOAuthglm-5.2GLM-5.2 and GLM-5-Turbo; native web search.
OpenAI CustomopenaiAPI key(set your own)Any OpenAI-compatible endpoint, including local runtimes.

From the interactive TUI, sign in or out with:

/login [provider]
/logout [provider]

openai works with any OpenAI-compatible endpoint: LM Studio, llama.cpp, vLLM, Hugging Face TGI, or any self-hosted runtime.

Important: Antigravity

Using antigravity through a third-party client may violate Google's Terms of Service.

Contributing

TypeScript + Bun. PRs welcome, and so are issues and feedback.

See Build from source for setup and local build commands. Requires Bun 1.4 canary.

For security details and reporting, see our Security Policy.

Ver o original no GitHub

Bom para

Relacionados