mise-ai-cli-overview

name	mise-ai-cli-overview
description	The four main AI CLIs compared — Claude Code, OpenAI Codex, Google Gemini, and aichat — with provider lock-in, model-agnostic fallback strategy, and why installing all four via mise is a lightweight dev-environment win. Use when picking AI CLIs for a project.

AI CLIs for dev — claude, codex, gemini, aichat

The AI-assisted coding landscape has consolidated around a few CLIs. All four are installable via mise (with one caveat for gemini). Running more than one is normal — they have different strengths and you'll want a model-agnostic fallback.

The four CLIs

CLI	Provider	Model backbone	Primary strength	Registry backend
Claude Code (claude)	Anthropic	Claude (Sonnet/Opus)	Agentic coding, long-context refactors, tool use, plugins	aqua:anthropics/claude-code
Codex (codex)	OpenAI	GPT-5 / o-series	Fast completions, strong at one-shot tasks	aqua:openai/codex
Gemini CLI (gemini)	Google	Gemini 2.x / 3.x	Huge context window, multimodal, Google Workspace integration	npm:@google/gemini-cli (not in registry yet)
aichat	Any (OpenAI/Anthropic/Gemini/Ollama/...)	Multi-provider adapter	Model-agnostic fallback, local models via Ollama	aqua:sigoden/aichat

Why install more than one

1. No single provider is best at everything. Claude excels at agentic workflows; Codex is fastest for simple completions; Gemini has the biggest context window; aichat lets you use local models.
2. Rate limits and outages. When one provider is down (happens more than you'd think), having another pre-installed means you're not blocked.
3. Cross-model verification. For critical code, ask two models the same question and compare.
4. Bring-your-own-key flexibility. Different projects may have different provider budgets.

The cost of installing all four is negligible — a few MB of binaries + a few env vars. The upside is you're never stuck when one is unavailable.

aichat as the fallback layer

aichat is special: it's not tied to a specific provider. It speaks to OpenAI, Anthropic, Gemini, Ollama, Cohere, and more via config. One CLI, many backends.

Use it when:

• You want to query a local model (Ollama, llama.cpp) for privacy-sensitive code.
• You want to A/B test across providers without switching CLIs.
• Your provider-specific CLI is broken and you need a quick fallback.
• You're scripting and want a stable CLI surface regardless of which provider you're hitting this month.

Config lives at ~/.config/aichat/config.yaml:

model: claude:claude-sonnet-4-5
clients:
  - type: claude
    api_key: $ANTHROPIC_API_KEY
  - type: openai
    api_key: $OPENAI_API_KEY
  - type: gemini
    api_key: $GEMINI_API_KEY
  - type: ollama
    api_base: http://localhost:11434

Provider lock-in and why it matters

Each vendor's CLI is optimized for their models and their tool-calling conventions. Switching between them means:

• Different prompt styles (Claude likes long context; Codex prefers terse).
• Different tool-use protocols.
• Different plugin / agent ecosystems (Claude Code has plugins; Codex doesn't, yet).
• Different pricing models (per-token vs per-request vs subscription).

You won't fully escape lock-in by installing all four, but you'll have off-ramps when you need them.

The mise.toml block

[tools]
node = "24"                # required for Gemini CLI (npm)
claude = "latest"          # aqua:anthropics/claude-code
codex = "latest"           # aqua:openai/codex
aichat = "latest"          # aqua:sigoden/aichat
"npm:@google/gemini-cli" = "latest"

[env]
ANTHROPIC_API_KEY = { required = "console.anthropic.com → API Keys", redact = true }
OPENAI_API_KEY    = { required = "platform.openai.com → API keys", redact = true }
GEMINI_API_KEY    = { required = "aistudio.google.com/app/apikey", redact = true }

[redactions]
patterns = ["*_API_KEY", "*_TOKEN", "*_SECRET"]

node = "24" is needed because Gemini CLI is an npm package. Even if you don't otherwise use Node, mise will install it transparently.

Keys stay out of mise.toml

Never write the actual key values into mise.toml. The required = "…" form is a prompt — when the variable is missing, mise tells the user where to get it. The value is set via their shell rc, keychain, or secrets manager. See mise-ai-cli-keys for the threat model.

The ai-status task

A simple sanity check that all four CLIs are installed and their keys are set:

[tasks.ai-status]
description = "Verify AI CLIs are installed and authenticated"
run = [
  "claude --version",
  "codex --version",
  "aichat --version",
  "gemini --version",
]

If any CLI prints an auth error instead of a version, that tells you the key isn't loaded. Run /mise-ai-keys to fix.

Anti-patterns

• Installing all four via curl/brew/pip, bypassing mise. Version pinning goes out the window.
• Committing a .env with keys "just for this project". Never.
• Using mise set ANTHROPIC_API_KEY=… — writes plaintext to mise.local.toml. Not a secret store.
• Assuming the gemini CLI package name is stable. It's changed twice in a year. Verify with npm view @google/gemini-cli before pinning.
• Only installing one CLI "to keep it simple". The day that provider is down is the day you'll wish you had a fallback.

See also

• mise-ai-cli-setup — the full mise.toml wiring and the ai-status task.
• mise-ai-cli-keys — threat model, keychain/1Password integration, rotation.
• mise-env-directives — how required = "" and redact = true work.
• data/ai-cli-research.md — authoritative registry mappings (fetched 2026-04-07).
• /mise-ai-init — scaffold this from a project.