Minimax CLI

minimax-cli

This is a manual, not a tool. It points you at the official MiniMax CLI (mmx). The CLI's own --help is the source of truth for syntax; the live docs are the source of truth for models, quotas, regions, and evolving flags. This file covers the LingTai-specific glue: install, credential discovery, region matching, and when to route through this CLI instead of another capability.

1. Scope and entry points

Use this skill when a task needs MiniMax-backed media generation or one-shot shell vision:

This skill is the canonical MiniMax CLI reference shipped with the TUI. The top-level minimax-cli skill is only a discoverability pointer into this nested reference, so keep MiniMax command recipes and credential guidance here rather than duplicating them there.

2. Install the CLI

The official CLI is mmx-cli on npm (source: MiniMax-AI/cli). Check first, install if missing:

command -v mmx >/dev/null || npm install -g mmx-cli

Requires node + npm on PATH. If neither is installed, ask the user to install Node; do not bootstrap a Node runtime yourself inside the project.

After install, verify the binary and let the CLI describe its current surface:

mmx --help
mmx doctor --help || true

3. Discover credentials without leaking them

Never print, commit, or paste MiniMax keys. The TUI stores keys in ~/.lingtai-tui/.env; presets declare which env slot they use through manifest.llm.api_key_env. Modern saved presets usually live under ~/.lingtai-tui/presets/saved/, so scan recursively instead of only checking the top-level presets directory.

Resolution order:

1. Find MiniMax presets and their declared key slot. LingTai has no minimax provider any more: a MiniMax preset is an anthropic (or openai) preset whose base_url points at a MiniMax host (api.minimaxi.com for CN, api.minimax.io for INTL). Older saved presets may still say provider: minimax.
2. Match the slot to ~/.lingtai-tui/.env without printing the key value.
3. Export the selected value as MINIMAX_API_KEY for mmx.
4. Match the region/base URL to the preset; do not guess a host.

List candidate presets (prints slot names and base URLs only, never secret values):

python3 - <<'PY'
import glob, json, os

root = os.path.expanduser("~/.lingtai-tui/presets")
paths = sorted(set(
    glob.glob(os.path.join(root, "**", "*.json"), recursive=True)
    + glob.glob(os.path.join(root, "**", "*.jsonc"), recursive=True)
))

def strip_jsonc(text: str) -> str:
    # Remove // and /* */ comments outside strings. A regex comment stripper
    # would corrupt normal preset URLs such as "https://api.minimaxi.com/anthropic".
    out = []
    i = 0
    in_string = False
    escape = False
    while i < len(text):
        ch = text[i]
        nxt = text[i + 1] if i + 1 < len(text) else ""
        if in_string:
            out.append(ch)
            if escape:
                escape = False
            elif ch == "\\":
                escape = True
            elif ch == '"':
                in_string = False
            i += 1
            continue
        if ch == '"':
            in_string = True
            out.append(ch)
            i += 1
            continue
        if ch == "/" and nxt == "/":
            i += 2
            while i < len(text) and text[i] not in "\r\n":
                i += 1
            continue
        if ch == "/" and nxt == "*":
            i += 2
            while i + 1 < len(text) and not (text[i] == "*" and text[i + 1] == "/"):
                i += 1
            i += 2 if i + 1 < len(text) else 0
            continue
        out.append(ch)
        i += 1
    return "".join(out)

for path in paths:
    try:
        doc = json.loads(strip_jsonc(open(path, encoding="utf-8").read()))
    except Exception:
        continue
    llm = doc.get("manifest", {}).get("llm", {}) or {}
    base = llm.get("base_url") or ""
    legacy = (llm.get("provider") or "").lower() == "minimax"
    if not legacy and "minimaxi.com" not in base and "minimax.io" not in base:
        continue
    slot = llm.get("api_key_env") or "MINIMAX_API_KEY"  # legacy fallback
    base = base or "(default CLI region)"
    region = "CN" if "minimaxi.com" in base else "INTL" if "minimax.io" in base else "unknown"
    rel = os.path.relpath(path, os.path.expanduser("~/.lingtai-tui"))
    print(f"{rel:55s} slot={slot:32s} region={region:7s} base_url={base}")
PY

Export the chosen slot without echoing the key. For ordinary API-key lines a grep extraction is enough; use Python only if a value is quoted or whitespace-padded (strip surrounding quotes/space, error if the slot is absent):

SLOT=MINIMAX_CN_1_API_KEY  # replace with the slot printed above
export MINIMAX_API_KEY="$(grep -E "^${SLOT}=" ~/.lingtai-tui/.env | head -1 | cut -d= -f2- | sed -E 's/^["'"'"' ]+|["'"'"' ]+$//g')"
[ -n "$MINIMAX_API_KEY" ] || echo "missing $SLOT in ~/.lingtai-tui/.env"

If multiple MiniMax presets exist and the user did not specify which account/region to use, ask. If no MiniMax preset exists, ask the user to save one through the TUI preset library — the anthropic template with base_url set to MiniMax's Anthropic-compatible endpoint (for example https://api.minimaxi.com/anthropic) and their MiniMax key pasted; the TUI will populate the appropriate slot in ~/.lingtai-tui/.env.

Token-plan keys often have prefix sk-cp-...; pay-as-you-go keys (sk-... without cp) may also work but are billed per call. Treat both as secrets.

4. Match the region/base URL

MiniMax has mainland China and international ecosystems. The credential, API host, and quota plan must match. The preset's base_url is the best local hint:

• minimaxi.com -> mainland China account.
• minimax.io -> international account.
• Empty/unknown -> use the CLI default unless the user or live docs say otherwise.

Do not hardcode an unverified host. If the CLI supports a region flag or MINIMAX_BASE_URL, set it from the preset's base_url or the current official docs, not from memory:

mmx doctor --help || true
mmx --help | sed -n '1,120p'
# If the CLI documents MINIMAX_BASE_URL and your selected preset has a base_url:
# export MINIMAX_BASE_URL="<base_url from the selected preset>"

A region/key mismatch commonly appears as 2049 invalid api key.

5. Discover subcommands at call time

Do not memorize command syntax. Ask the installed CLI immediately before generating:

mmx --help
mmx image --help
mmx video --help
mmx music --help
mmx speech --help
mmx vision --help
mmx doctor

As of this writing the CLI commonly exposes text, image, video, music, speech, and vision subcommands. Outputs usually land in ./minimax-output/, and most subcommands have an output flag such as --out; verify the current flag with --help before every scripted use.

Live docs when --help is not enough:

• International: https://platform.minimax.io/docs/token-plan/minimax-cli
• Mainland China: https://platform.minimaxi.com/docs/token-plan/minimax-cli

6. Generation workflow

1. Confirm the requested modality and cost. Media calls can take minutes and may consume paid quota. For ambiguous prompts, ask before calling.
2. Create an artifact directory inside the project or requested destination, e.g. artifacts/minimax/<task-slug>/.
3. Run mmx <subcommand> --help and build the command from the live flags.
4. Send a progress update for human-facing long calls (>5 seconds) before waiting.
5. Save outputs explicitly with an output flag/path when available; otherwise move files out of ./minimax-output/ after the call.
6. Report only paths and a concise summary. Do not include keys, raw auth headers, or huge provider payloads.

Skeleton:

mkdir -p artifacts/minimax/example
mmx music --help
# Build the real command from the help output; example flags may change:
# mmx music generate --prompt "..." --out artifacts/minimax/example/

7. Vision-specific routing

MiniMax vision can appear through multiple LingTai paths:

• Kernel vision tool: best when present in the agent's tool list; follow the tool's current manual.
• MiniMax CLI (mmx vision ...): best for one-shot shell/OCR/image-description jobs.
• MCP vision tools: only if an MCP is already registered; MCP setup belongs to mcp-manual.

If a user asks for image understanding and you already have the built-in vision tool, use it first. Load this skill for CLI fallback or batch shell work.

8. Failure modes

The older MCP-server route (minimax-mcp, minimax-coding-plan-mcp) still exists for MCP workflows and is owned by mcp-manual. Prefer the official CLI here when a shell command is sufficient: one binary covers media, speech, text, and ad-hoc vision.

Found a bug or issue? If you encounter any problems with this skill, load the lingtai-issue-report skill and follow its instructions to report it.