baze

Run Claude Code, Codex, OpenCode and Hermes Agent through the Baze API gateway with one API key; generate images, video, speech and music; ask your organization's studio chatbots; and read models, credits and usage from a shell.

Gateway documentation: docs.mindlogic.ai (AWS) · docs.factchat.kr (NCP).

Install

npm i -g @mindlogic-ai/baze
baze --version

For a one-off call without installing:

npx @mindlogic-ai/baze models

The scoped name is required. Unscoped baze on npm is an unrelated 2017 package.

Node 20.10 or newer. No runtime dependencies.

Log in

baze login

Paste an API key issued in your Baze console (how to get one). Echo is off. Baze runs as two separate deployments, AWS (factchat-cloud.mindlogic.ai) and NCP (api.baze.co), and the key names the one it belongs to, so everyone runs the same command; it is checked against the server before it is saved to ~/.baze/credentials.json (mode 0600).

For CI, containers, or an older key that does not name its deployment, pipe the key and say which Baze it belongs to:

printf %s "$KEY" | baze login --with-key --host aws
printf %s "$KEY" | baze login --with-key --host ncp

--host also takes a full URL for an on-premise server. BAZE_API_KEY in the environment beats the stored key, so a container needs no login at all.

Run a coding agent

baze claude
baze codex
baze opencode
baze hermes

Each command starts the tool you already have installed, pointed at the gateway with your key, and exits with the tool's exit code. Nothing is written to disk: the configuration travels as a temporary settings file (Claude Code), -c overrides (Codex), OPENCODE_CONFIG_CONTENT (OpenCode) or environment variables plus --provider custom (Hermes), and your own config files stay untouched.

Every argument other than --model, --dry-run and --api-base (Codex: also --reasoning-effort) goes to the tool unchanged, --help and --json included. baze claude --help is Claude Code's help; baze's own is baze help claude.

baze claude --model claude-opus-5 -p "fix the failing test"
baze codex --model gpt-5.6-terra --reasoning-effort high exec --json "summarize this repo"
baze opencode --model gemini-3.1-pro-preview run "explain src/"
baze hermes --model gpt-6-sol -z "summarize this repo"
baze claude --dry-run          # print the plan, start nothing

Which harness reaches which models

Harness	Protocol the tool speaks	Models it can use through baze
Claude Code (baze claude)	Anthropic Messages	Every chat model; non-Claude models are converted to chat completions by the gateway
Codex (baze codex)	OpenAI Responses	GPT models only
OpenCode (baze opencode)	Anthropic, Responses, Gemini and OpenAI-compatible providers	Every chat model this key can call, each on its native lane
Hermes Agent (baze hermes)	OpenAI Chat Completions	Every chat model this key can call, all over Chat Completions

baze models shows the lane of every model. --model takes the full id as baze models prints it.

Hermes sends a key from the environment only to the host it was issued for, and derives the variable's name from the domain, so baze hermes passes your key as MINDLOGIC_API_KEY (AWS) or BAZE_API_KEY (NCP) together with CUSTOM_BASE_URL. A ~/.hermes/.env that sets either of those wins over the launch; baze warns when it does. There is no baze setup hermes yet.

Generate images, video, speech and music

baze models --type image                       # what this key can call
baze gen image "a poster for a jazz night" --model gpt-image-1 --set quality=high
baze gen image "make it blue" --model gpt-image-1 --image poster.png -o blue.png
baze gen video "the camera slowly circles" --model bytedance/seedance-1-pro --image blue.png --set duration=5
baze gen speech "안녕하세요" -o hello.wav
baze gen music "short upbeat synth jingle" --model lyria-3-clip-preview
baze gen transcribe meeting.m4a --lang ko,en > meeting.txt
baze gen video --model bytedance/seedance-1-pro --params    # what --set takes

Results land in the current directory and their paths are printed, so the next command (or your coding agent) can use them directly. --set is checked against the model before anything is sent. A video is waited for; if you stop it, baze gen wait <id> picks it up again, and baze history lists every call with its files.

Talk to a studio chatbot

baze bot ask 3840 "연차 이월 규정 알려줘"
baze bot ask 3840 "그럼 신입은?" --continue
baze bot chat 3840

The chatbot answers from its own knowledge and workflow, as it does in the studio. Its owner has to turn on API access.

Make it permanent

baze setup claude
baze setup codex
baze setup opencode

setup rewrites the tool's user-level configuration so a plain claude, codex or opencode uses the gateway. The previous file is backed up first, and the command prints the backup path. --dry-run prints the file without writing it, with the key masked.

Where the key ends up differs by tool:

• Claude Code: ~/.claude/settings.json holds the key in plain text, because Claude Code reads it only from there or the environment. It is the user-level file, not a project one, so it is not committed, but on a shared machine prefer baze claude, which leaves nothing behind.
• Codex and OpenCode: the config names the BAZE_API_KEY variable and holds no key. Put eval "$(baze env)" in your shell's startup file, or use baze codex / baze opencode, which need no variable.

Commands

Command	What it does
baze login [--with-key] [--host aws|ncp|<url>]	store and validate an API key
baze logout [--revoke]	delete the stored key; --revoke disables it on the server first
baze auth	the masked key, where it came from, and the host in effect
baze env	eval "$(baze env)" exports BAZE_API_KEY
baze models [query] [--type T]	models this key can call, with lane, context and list price
baze embed [text...]	embedding vectors, billed per token
baze rerank <query> [doc...]	order documents by relevance to a query
baze guard [text...]	find or mask personal data in text
baze mcp [list] | tools <suite> | call <suite> <tool> ['{json}']	the Baze first-party MCP suites
baze gen image|video|speech|music|transcribe	generate media; results are saved as files in the current directory
baze gen wait <id>	collect a job started with --no-wait or interrupted
baze bot ask|chat <chatbot id>	talk to a studio chatbot, with its knowledge, workflow and tools
baze history	every generation and bot call from this machine
baze credits	monthly, purchased and org-granted balances
baze usage [--from] [--to] [--by model|key|day] [--scope key|member]	gateway spend, Korea time, at most 92 days
baze claude | codex | opencode | hermes [--model ID] [--dry-run]	run that tool through the gateway
baze setup <tool> [--model ID] [--dry-run]	write the configuration permanently, with a backup
baze doctor	check key, host, Node, tool versions and conflicting config
baze update	npm install -g @mindlogic-ai/baze@latest
baze docs	print the agent manual
baze help <command>	help for one command

Global flags

Flag	Effect
--json	machine-readable stdout
--api-base <url>	same as BAZE_API_BASE, for this call
--no-input	never prompt; fail instead
-q, --quiet	drop hints and progress notes

Launchers do not take the global flags. On baze claude, baze codex, baze opencode and baze hermes, --json, -q and --help belong to the tool.

Environment

Variable	Effect
BAZE_API_KEY	beats the stored key (CI, containers)
BAZE_API_BASE	beats the stored host
BAZE_CONFIG_DIR	config directory (default ~/.baze)

If a turn fails with "Request blocked by content policy"

A Baze admin can screen API requests for personal data and restricted words. With blocking on, a turn whose text carries a phone number, an ID, or a listed word is refused with that 400, and coding agents send code, logs and git history, so it can happen. Nothing in baze lifts it: ask the admin to switch the API filter to masking, or to exempt the API.

Output and exit codes

stdout is data and stderr is prose, so baze models --json | jq and eval "$(baze env)" are safe. Exit codes: 0 ok, 1 error, 2 out of credit, 127 the tool is not installed. A launcher exits with the tool's own code.

For coding agents

skills/baze/SKILL.md is an agent skill for baze gen, embed, rerank, guard, mcp and bot, released with the CLI so it never describes a flag this version does not have. baze skill install puts it into Claude Code and Codex (baze skill status shows whether it is current); for another agent, copy the folder into its skills directory. The full machine contract is baze docs.

Contributing

npm ci
npm test            # compiles to .build/ and runs node --test
npm run typecheck
npm run build       # dist/index.js with the version stamped in

Releases are a v* tag on main; .github/workflows/release.yml publishes to npm with provenance through Trusted Publishing. Adding a model to the gateway never needs a release: the CLI reads the catalog at run time.

License

Apache-2.0. See LICENSE. The license covers this code, not the Baze name or marks.