monash

An agent with a single tool.

monash is an AI agent reduced to one verb: a small Lisp. There's no bash, no file editor, no search command — its only tool evaluates Lisp. Reading files, running commands, searching the project are all built-in primitives, so monash works by composing them into expressions instead of firing one-shot tool calls.

Install

npm i -g @guanyilun/monash

Use

monash                 # start the agent
monash -c              # continue the last session in this directory

monash is an interactive terminal chat: type a request, read the reply, and watch each expression evaluate as it works.

Authenticate

monash needs an API key for your model provider. Either store one:

monash auth login      # paste a key; saved privately under ~/.monash

monash auth list shows what's configured; monash auth logout <provider> removes a key.

Configure providers

monash works with any OpenAI-compatible API. You can configure providers in ~/.monash/settings.json, pass CLI flags, or set environment variables — they compose, with CLI flags taking highest precedence.

Provider profiles in settings.json

Define named providers in ~/.monash/settings.json so you don't pass flags every time. The apiKey field supports $ENV_VAR and ${ENV_VAR} syntax — secrets are expanded at runtime and never stored in plaintext in the file:

{
  "defaultProvider": "openai",
  "providers": {
    "openai": {
      "apiKey": "$OPENAI_API_KEY",
      "defaultModel": "gpt-4o",
      "models": ["gpt-4o", "gpt-4o-mini"]
    },
    "ollama": {
      "apiKey": "not-needed",
      "baseURL": "http://localhost:11434/v1",
      "defaultModel": "llama3",
      "models": ["llama3", "mistral", "codellama"]
    },
    "openrouter": {
      "apiKey": "$OPENROUTER_KEY",
      "baseURL": "https://openrouter.ai/api/v1",
      "defaultModel": "anthropic/claude-sonnet-4-20250514"
    }
  }
}

Then launch with:

monash                          # uses defaultProvider
monash --provider ollama        # pick a specific provider
monash --provider openai --model gpt-4-turbo  # override model

Quick start

OpenAI

export OPENAI_API_KEY="sk-..."
monash --model gpt-4o

OpenRouter

monash --api-key "$OPENROUTER_KEY" \
  --base-url https://openrouter.ai/api/v1 \
  --model anthropic/claude-sonnet-4-20250514

Ollama (local, no auth)

monash --api-key dummy --base-url http://localhost:11434/v1 --model llama3

LM Studio / vLLM (local)

monash --api-key dummy --base-url http://localhost:1234/v1 --model local-model

CLI flags

| Flag | Environment variable | Description | |---|---|---| | --provider <name> | — | Provider profile from settings.json | | --model <name> | — | Model name (overrides provider default) | | --api-key <key> | OPENAI_API_KEY | API key | | --base-url <url> | OPENAI_BASE_URL | Base URL for API endpoint |

Precedence (highest to lowest): CLI flags → environment variables → provider profile in settings.json → built-in defaults.

Switching models at runtime

Inside a session, /model shows the current model. /model <name> switches to a different model (possibly crossing providers) without losing conversation state.

Reference

For the full settings schema — per-model contextWindow, reasoning flags, defaultBackend, extensions, and more — see the agent-sh usage guide. monash shares agent-sh's configuration model; the only difference is that monash stores its settings in ~/.monash/settings.json instead of ~/.agent-sh/settings.json.

How it works

Everything monash does is Lisp. Rather than juggling separate tools for the shell, files, and search, it has one — a Lisp interpreter whose primitives cover all three. Ask for something and it writes an expression, evaluates it, reads the result, and continues; you see the source and its value at every step.

Permissions

monash runs every primitive freely by default. To require confirmation before certain ones, list them in ~/.monash/settings.json:

{ "monash": { "gate": ["bash", "write-file", "edit-file"] } }

Any evaluation that calls a listed primitive prompts once before it runs — allow once, allow for the session, or deny. The list can name built-ins or primitives added by extensions. For finer policies — gating only dangerous commands, restricting write paths, auditing calls — register a guard from an extension; see EXTENDING.md.

Extending

monash has one tool, so you add capability by registering a Scheme primitive — a plain JS function the model can call and compose — not a tool. See EXTENDING.md.

Development

monash is built on agent-sh. The monash-specific pieces live here and ship on their own cadence:

• lib/scheme.ts — the interpreter: the Scheme engine, its primitives, and the tool description.
• lib/monash.ts — the agent integration: registers the one tool, the system prompt, and the renderer.
• lib/scheme-api.ts — the extension API (@guanyilun/monash/scheme): createScheme(ctx).
• bin.mjs — the launcher.

npm install
npm link        # expose the `monash` command locally

License

MIT