⌬ Zero CLI

Start from zero. Ship from anywhere.

A geometric, multi-model AI coding agent for the terminal — concentric, focused, and at home in any shell. Drives Anthropic Messages and OpenAI Chat Completions-compatible LLM APIs (Anthropic Claude, Zhipu GLM, DeepSeek, Volcengine, Alibaba Qwen, Moonshot, and more) from a single, brand-distinct CLI marked by the ⌬ Zero glyph.

Zero CLI focuses on optimizing the harness layer (context window management, token budgeting, request headers, timeouts, agent runtime) for third-party model providers. For the official Anthropic Claude models, use Anthropic's official Claude Code — that's where they're tuned best.

Install

# Recommended (faster)
bun add -g @paean-ai/zero-cli

# Or via npm
npm install -g @paean-ai/zero-cli

Guides

• Windows + DeepSeek (English, npm-only): docs/windows-install-zero-cli-deepseek.md
• 简体中文： docs/windows-install-zero-cli-deepseek.zh.md

Quick Start

Zero supports two backends, picked at the CLI level without opening the TUI:

# 1) Paean AI (default, recommended)
zero login

# 2) Third-party provider (e.g. z.ai's GLM Coding Plan)
zero provider set zai --api-key <YOUR_GLM_API_KEY>

Text-only gateways can declare their affected model IDs during setup:

zero provider set byok --api-key <KEY> --base-url <URL> \
  --model dsv4-0731 --non-multimodal-models dsv4-0731

Then start a session:

zero                         # interactive
zero -p "explain this repo"  # one-shot

Also available as tide

Zero CLI installs a second command called tide that launches the same binary under the DeepTide brand:

tide                         # same as `zero`, branded as DeepTide
tide -p "explain this repo"

DeepTide is our macOS-native AI coding app; Zero CLI is the cross-platform CLI from the same team that serves Linux, Windows, and non-native-app macOS users. The two products are developed in parallel and evolve independently — Zero CLI focuses on the harness layer for terminal-first workflows; DeepTide focuses on a tuned native macOS experience. The shared interface contract (tool names, slash commands, CLI flags, hook env vars, model aliases) lives in docs/tide-spec.md so a user's hooks and muscle memory port between the two.

On macOS we recommend the native DeepTide app for the best GUI experience. The tide CLI is the cross-platform terminal complement.

zero login — Paean AI account

zero login           # browser-based OAuth
zero auth status     # check who you're logged in as
zero auth logout

zero provider — Third-party API provider

Use this when you want Zero to talk to a non-Paean endpoint that implements Anthropic Messages or OpenAI Chat Completions. Three cloud presets ship out of the box: z.ai (GLM), Volcengine (Ark), and DeepSeek; oMLX is also available for local Anthropic-compatible serving. All commands are scriptable — no TUI required.

# Show currently active provider (if any)
zero provider
zero provider status --json

# List available provider presets
zero provider list

# Configure a preset (writes the API key to userSettings + cache)
zero provider set zai        --api-key <YOUR_GLM_API_KEY>
zero provider set volcengine --api-key <YOUR_ARK_API_KEY>
zero provider set deepseek   --api-key <YOUR_DEEPSEEK_API_KEY>

# Optional: override model IDs per tier
zero provider set zai --api-key <KEY> \
  --opus glm-5.1 --sonnet glm-5-turbo --haiku glm-4.5-air

# Switch back to Paean.ai
zero provider clear

The API key may also be supplied via the ZERO_PROVIDER_API_KEY environment variable (useful in CI). Restart zero after changing the provider for the new config to take effect.

OpenAI Chat Completions-compatible endpoints

For vLLM, SGLang, New API, and other servers that expose /v1/chat/completions, configure a BYOK provider with --protocol openai-chat:

zero provider set byok \
  --name NPU \
  --api-key "$ZERO_PROVIDER_API_KEY" \
  --base-url https://provider.example \
  --model dsv4-0731 \
  --protocol openai-chat \
  --non-multimodal-models dsv4-0731

The base URL may be the server root (https://provider.example), /v1, or a provider-specific API prefix. Zero appends the appropriate /chat/completions path and translates messages, function tools, streamed tool arguments, tool results, token usage, reasoning content, and stop reasons at the API boundary. The agent loop and JSONL transcript keep their existing internal shape.

--protocol auto is the default. It currently recognizes dsv4-0731 (also with a [1m] suffix) and automatically selects OpenAI Chat Completions. Use an explicit protocol for other model IDs:

zero provider set byok --api-key "$ZERO_PROVIDER_API_KEY" \
  --base-url https://provider.example/v1 \
  --model your-model-id --protocol openai-chat

The equivalent manual override is ZERO_CLI_API_PROTOCOL=openai-chat; set it to anthropic to force Anthropic Messages. OpenAI Responses API is not part of this adapter. See docs/openai-protocol.md for the supported surface and compatibility details.

Quick switch between providers (cached credentials)

Once a preset has been configured at least once with zero provider set, its credentials are cached in userSettings.providerCache. You can then jump between providers without retyping the API key:

# Re-activate a preset from cache (no --api-key needed)
zero provider use zai
zero provider use volcengine
zero provider use deepseek

# Inspect what's cached (API keys are masked)
zero provider cache list
zero provider cache list --json

# Evict one entry, or wipe the whole cache
zero provider cache clear deepseek
zero provider cache clear

zero provider use <preset> fails fast if no cache entry exists — you'll be told to run zero provider set first.

Preset defaults

z.ai (GLM Coding Plan)

Zero's z.ai preset is for the GLM Coding Plan key used through Z.ai's Anthropic-compatible Claude Code endpoint:

https://open.bigmodel.cn/api/anthropic

Z.ai's GLM Coding Plan docs also list https://open.bigmodel.cn/api/coding/paas/v4; that is its dedicated OpenAI-compatible Coding API prefix. Do not paste that URL into the fixed z.ai preset, which intentionally uses Anthropic Messages. Configure it through BYOK with --protocol openai-chat and the model ID supported by your plan:

zero provider set byok --name z.ai-openai \
  --api-key "$ZERO_PROVIDER_API_KEY" \
  --base-url https://open.bigmodel.cn/api/coding/paas/v4 \
  --model your-zai-model-id --protocol openai-chat

| Tier | Model | Context | |------|-------|---------| | pro (default) | glm-5.1 | 200K | | flash | glm-5-turbo | 200K | | lite | glm-4.5-air | 128K |

Get an API key at https://bigmodel.cn/usercenter/proj-mgmt/apikeys.

When the z.ai preset is active, Zero also exposes /zai-usage (aliases: /glm-usage, /coding-plan-usage) to query GLM Coding Plan model usage, tool usage, and quota limits. This is built into Zero and follows Z.ai's glm-plan-usage plugin behavior; no separate @z_ai/coding-helper installation is required.

Volcengine (火山引擎 Ark Coding Plan)

| Tier | Model | Context | |------|-------|---------| | pro (default) | ark-code-latest | 256K | | flash | ark-code-latest | 256K | | lite | ark-code-latest | 256K |

Volcengine's coding endpoint serves a single unified model (ark-code-latest) across all tiers. Pin to a specific snapshot with zero provider set volcengine --opus <snapshot-id> or via the wizard.

Get an API key at https://console.volcengine.com/ark.

DeepSeek (Official)

| Tier | Model | Context | |------|-------|---------| | pro | deepseek-v4-pro | 1M | | flash (default) | deepseek-v4-flash | 1M | | lite | deepseek-v4-flash | 1M |

Two-model lineup: deepseek-v4-pro is the flagship, deepseek-v4-flash is the daily driver. The default tier is flash, mirroring DeepSeek's own pricing-page recommendation.

Get an API key at https://platform.deepseek.com/api_keys.

Model Tiers

Zero maps three capability tiers to any supported model:

| Preferred alias | Legacy alias | Role | |---|---|---| | pro | opus | Flagship — most capable, highest cost | | flash | sonnet | Standard — best for everyday tasks | | lite | haiku | Fast / cheap — quick answers, simple tasks |

Switch tiers in-session with either name — both resolve to the same model:

/model pro      # or /model opus
/model flash    # or /model sonnet
/model lite     # or /model haiku

The pro / flash / lite naming aligns with how third-party Anthropic-protocol providers (DeepSeek, Volcengine, GLM, Gemini, Qwen) tier their own catalogs. Zero CLI's display layer prefers these names; the legacy opus / sonnet / haiku remain accepted inputs for backward compatibility.

zero provider set writes the right env vars for you; to override manually:

export ZERO_CLI_DEFAULT_OPUS_MODEL=glm-5.2        # pro tier
export ZERO_CLI_DEFAULT_SONNET_MODEL=glm-5.1      # flash tier
export ZERO_CLI_DEFAULT_HAIKU_MODEL=glm-4.5-air   # lite tier

(Env-var infixes still use the canonical internal form OPUS / SONNET / HAIKU so existing scripts and CI keep working.)

Or switch models in-session with /model.

Paean Models

| Model | Provider | Tier | |-------|----------|------| | gpt-5.6-sol | OpenAI | Flagship | | gemini-3.6-flash | Google | Standard | | glm-5.2 | Zhipu AI | Flagship | | qwen3.8-max-preview | Alibaba | Flagship | | deepseek-v4-flash | DeepSeek | Fast | | kimi-k3 | Moonshot AI | Flagship | | MiniMax-M3 | MiniMax | Flagship |

Paean's /model picker intersects this curated latest-only list with the live /v1/models response. A model is not shown until the Paean gateway reports it as available; older generations and specialized image models are not offered.

Configuration

Zero stores config in ~/.zero/. Key environment variables:

| Variable | Description | |----------|-------------| | ZERO_CLI_AUTH_TOKEN | API key for the configured provider | | ZERO_CLI_BASE_URL | API base URL (defaults to Paean AI) | | ZERO_CLI_API_PROTOCOL | Wire protocol override: anthropic or openai-chat; unset means automatic selection | | ZERO_CLI_DEFAULT_OPUS_MODEL | Model for the "opus" (flagship) tier | | ZERO_CLI_DEFAULT_SONNET_MODEL | Model for the "sonnet" (standard) tier | | ZERO_CLI_DEFAULT_HAIKU_MODEL | Model for the "haiku" (fast) tier | | ZERO_CLI_NON_MULTIMODAL_MODELS | Comma-separated model IDs that must not receive image/document blocks; use * for every model on the configured third-party provider | | ZERO_PROVIDER_API_KEY | Fallback key read by zero provider set in CI |

Legacy ANTHROPIC_* equivalents are still honored but the ZERO_CLI_* names are preferred.

For a text-only provider, declare the affected model IDs in user settings (or the process environment). Zero then blocks visual Read calls, replaces any image/document blocks before the API request, and tells the model to use text extraction instead:

{
  "env": {
    "ZERO_CLI_NON_MULTIMODAL_MODELS": "deepseek-v4-flash,dsv4-0731"
  }
}

License

MIT