Hablas CLI

Single-agent terminal engineering runtime — v2.9.2

One agent. Plans, edits, tests, commits — on the machine in front of you.

hablas --setup: 1 NVIDIA · 2 Mistral · 3 Custom. NVIDIA finishes in one step with hosted keys (not baked into the package).

Install

npm install -g hablas-ai
hablas --setup
hablas

Node 20+. One provider:

| Option | What you bring | |---|---| | NVIDIA | Nothing. Hosted NIM, one tap. | | Mistral | Nothing if hosted; otherwise your key. | | Ollama | Nothing. Local and private. | | Custom | Your URL + key (Groq, OpenRouter, DeepSeek, …) |

Config ladder

flags  >  exported env  >  ~/.hablas/.env  >  ~/.hablas/config.json  >  defaults

# copy the example
cp flowagent/.env.example ~/.hablas/.env
# or just:
export HABLAS_PROVIDER=nvidia
export HABLAS_API_URL=https://integrate.api.nvidia.com/v1
export HABLAS_MODEL=stepfun-ai/step-3.7-flash
export HABLAS_API_KEY=nvapi-...
hablas env    # keys redacted
hablas

Commands

| Command | What it does | |---|---| | hablas | Interactive session | | hablas --setup | Provider wizard | | hablas run "<prompt>" | One shot | | hablas env | Active config, secrets hidden | | hablas doctor | Connectivity check | | hablas --model <id> | Override model | | hablas --provider ollama\|nvidia\|custom\|mistral | Override provider | | hablas --auto | Skip confirm-level tools. Dangerous tools stay blocked. | | hablas mcp-serve | Expose tools as an MCP server | | hablas --version | Print version |

Safety that actually exists

• File tools go through resolveInside(). /etc/passwd is an error.
• run_command and execute_code are dangerous — skipped in --auto.
• execute_code runs with a stripped env. It cannot print HABLAS_API_KEY.

See docs/SECURITY.md.

License

MIT — LICENSE