Pond: Agente IA para Fish Shell
bndlfm/pond.fish
Agente IA autônomo para o Fish shell com auditoria e controle total
Ver no GitHubAgente autônomo de IA integrado ao Fish shell que lê arquivos, lista diretórios, busca na web e executa comandos para alcançar objetivos complexos. Oferece controle fino sobre permissões, auditoria em tempo real e suporte a skills customizáveis via Markdown.
Instalação
npx convez-ai bndlfm/pond.fish1 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
Casos de uso
- Exploração e navegação de codebases complexos
- Automação de tarefas multi-step no terminal
- Geração e correção de comandos shell com contexto
- Pesquisa web integrada para troubleshooting
- Extensão de capacidades do agente via skills customizáveis
Veredito

Pond é um agente genuinamente útil para quem vive no terminal Fish. A arquitetura de permissões em 4 níveis e auditoria visual são destaques reais. Recomendado para desenvolvedores que querem automação inteligente sem abandonar a segurança.
README
A tradução para o português deste README ainda está na fila.
🐟 pond: AI for Fish shell
Agentic fork of the original fish-ai by Bastian Fredriksson.
pond is a minimalist yet capable autonomous AI agent living directly in your Fish shell. It prioritizes auditability, security, and seamless shell integration.
🚀 Key Features
- Autonomous AI Agent (
Ctrl+A): A multi-turn expert that can read files, list directories, search the web, and execute shell commands to achieve complex goals. - Unified
pondCommand: A master utility for piping data to an LLM, managing the agent, or asking quick questions. - SKILL.md Support: Fully compatible with the
skills.sh/agentskills.iostandard. "Teach" the agent new expertise by dropping Markdown folders into~/.config/fish-ai/skills/. - Question / Explain (
Ctrl+Q): Instantly turn natural language into shell commands or get clear explanations of what a command does. Query / Question. - Autocomplete / Fix (
Ctrl+Space): Intelligent, context-aware command completions and instant fixes for your last failed command. - Brave Search Integration: Real-time web access for troubleshooting, documentation, and research.
- Advanced Audit UI: Color-coded streaming of agent thoughts, tool calls, and truncated results directly in your terminal.
- Surgical Permissions: A 4-tier permission system (
[y/t/a/n]) that puts you in full control of every system-modifying action. - Session Persistence: Maintains conversation state between loops, allowing for long-running, multi-step collaborations.
📦 Installation
Install using fisher:
fisher install bndlfm/pond.fish
🤖 Usage Guide
🦾 The AI Agent
Type a goal in plain text or as a comment and press Ctrl + A:
# find all python files and search for TODOs
(press Ctrl+A)
The agent will work turn-by-turn. When it needs to execute a command, it will prompt you:
[y]Allow once: Permit only this specific command.[t]Allow for this task: Grant temporary autonomy until the current goal is met or the agent asks a question.[a]Always allow: Grant full autonomy for the rest of the shell session.[n]Deny: Prevent the command from running and let the agent rethink.
Manage Session State:
pond -a <goal>: Trigger the autonomous agent.pond forget: Wipes the agent's memory to start a fresh session.pond compress: Manually trigger a summarization of long histories.pond status: View current session statistics.pond skill list: List all available specialized skills.
🐚 Unified pond Command (Piping & Quick/Query/Question)
The pond command provides a master interface for AI tasks. To ask a quick, stateless Question, or run a Query, use the -q flag:
# Pipe context in
cat README.md | pond -q "summarize this"
# Ask a direct question
pond -q "what is the capital of Spain?"
# Output raw JSON
pond -q "find python entrypoints" --json
Implicit queries (unflagged strings) are disabled for safety. It does not include shell history or previous agent state, making it ideal for scripting and data processing.
🛡️ Command Whitelist
pond can run safe commands automatically. Customize this list in your configuration:
[fish-ai]
whitelist = ls, grep, find, cat, pwd, date, eza, fd, rg
Security Guard: Any command containing redirections (>), pipes (|), or chaining (;, &&) is never whitelisted and will always require your manual approval.
🌐 Web Search
To enable web research, add your Brave Search API key to your config:
[fish-ai]
brave_search_api_key = <your_key>
🔌 Specialized Skills
pond supports the agentskills.io open standard. To add an expert, use the add command to pull directly from a GitHub repository:
pond skill add anthropics/skills/skills/pdf
Or manually drop a skill folder into your skills directory:
# Location: ~/.config/fish-ai/skills/
pond skill list
📝 Quick / Query / Question
Press Ctrl + Q to swap between natural language and shell commands:
list files larger than 1gb→find . -size +1Gtar -xvzf archive.tar.gz→ Explains the command and flags.
🪄 Autocomplete & Fix
- Press Ctrl + Space while typing for intelligent completions.
- Press Ctrl + Space after a command fails to receive an immediate fix based on the error output.
🤸 Configuration
Edit ~/.config/fish-ai/config.ini or use environment variables (ideal for Nix/Home Manager):
[fish-ai]
configuration = my-provider
whitelist = ls, rg, fd, cat
[my_provider]
provider = google
api_key = <your_key>
model = gemini-3.1-pro-preview
Runtime Keybindings (Nix/Home Manager)
You can customize keybindings via environment variables in your shell config:
FISH_AI_KEYMAP_1: Defaults toctrl-q(Question)FISH_AI_KEYMAP_2: Defaults toctrl-space(Autocomplete)FISH_AI_KEYMAP_3: Defaults toctrl-a(Agent)
🛠️ Development
pond includes a Nix Flake for a reproducible development environment.
Run nix develop to enter a shell with all dependencies configured.
See DEVELOPMENT.md for more details.