@maya-os/code — maya code (mayac)

The coding agent of Maya OS — a local-first AI operating system for developers. mayac is a provider-agnostic coding-agent harness built fresh on the official Anthropic and OpenAI SDKs, with a full-screen Ink TUI and a CLI. One internal message+tool format; adding a third provider is a single new file. It runs on your own subscription or API key, and inside Maya OS it's a selectable agent runtime alongside Claude Code and Codex.

Install

npm install -g @maya-os/code      # then run:  mayac
# or, no install:
npx @maya-os/code                 # interactive TUI
npx @maya-os/code -p "add a /health route and run the tests"

Highlights

• Two providers, one interface. Provider.stream(req) → AsyncIterable<StreamEvent>. Anthropic Messages API and OpenAI Responses API both translate to the same internal ModelMessage / tool format.
• API key and subscription auth, per provider, auto-detected.
  • Anthropic: ANTHROPIC_API_KEY, or Claude Code OAuth (~/.claude/.credentials.json / macOS keychain Claude Code-credentials).
  • OpenAI: OPENAI_API_KEY, or ChatGPT/Codex login (~/.codex/auth.json) which targets the ChatGPT/Codex backend.
• Real agent loop. Streaming, tool-use turn loop, context auto-compaction near the token limit, retry/backoff, and token+cost accounting.
• Tools with a permission gate. read, write, edit (exact-string replace), bash, glob, grep, ls. Reads auto-allow; writes/bash are gated unless --yolo.
• Full-screen Ink TUI (Claude-Code-like). A fixed layout — status header, a managed scroll viewport, and a live input box:
  • Scrollback viewport. PageUp/PageDown + auto-stick to the bottom; a ▼ N new below hint when you've scrolled up.
  • Collapsible tool calls. The most recent call shows full detail; older ones collapse to one line. Ctrl+P/Ctrl+N focus a call, →/Enter expand, ← collapse.
  • Autocomplete. Slash commands, @/path files (from the cwd), and command history — Tab accepts, Esc dismisses, ↑/↓ moves.
  • Send while working. Type mid-turn to queue the next message (shown as a chip); Esc interrupts the running turn (aborts the stream + any running tool).
  • Dynamic /model list fetched from the provider's /v1/models (cached, falls back to a built-in list offline).
  • Background tasks. /bg <cmd> runs a detached shell command; /tasks lists/kills them; a ⚙ N bg chip shows in the header.
  • Sessions. Every session persists to ~/.maya-harness; /resume (or --resume/--continue) reloads a past transcript. ↑/↓ walks input history.

Build from source

pnpm -C packages/harness build
pnpm -C packages/harness test

Usage

# Interactive TUI (default provider: anthropic)
mayac
mayac --provider openai --auth subscription

# Headless one-shot
mayac -p "add a /health route to server.ts and run the tests"

# Pick provider / model / auth explicitly
mayac --provider anthropic --model claude-opus-4-8 --auth subscription
mayac --provider openai   --model gpt-5.5         --auth subscription

In the TUI

• Type a message + Enter to chat; streamed replies, tool calls + results render in the viewport.
• Header shows provider/model/auth, a live token ↑↓ and $ cost meter, and ⚙ N bg when tasks run.
• Scroll: PageUp/PageDown (auto-sticks to the newest output; ▼ N new below when scrolled up); Esc jumps to the latest.
• Tool calls: Ctrl+P/Ctrl+N focus a call, →/Enter expand, ← collapse, Esc leaves focus.
• Autocomplete: start / (commands), @/a path (files), or type a prefix of a past command (history) — Tab accepts, Esc dismisses, ↑/↓ moves.
• While the agent works: the input stays live — type + Enter to queue the next message; Esc interrupts the current turn.
• Line editing: ←/→ move, ctrl+a/ctrl+e home/end, ctrl+w delete word, ctrl+u clear, ↑/↓ history.
• ctrl+c twice to exit.

Slash commands:

| Command | Effect | | --- | --- | | /model [id] | switch model (no arg = picker; list fetched live) | | /provider [id] | switch provider (no arg = picker) | | /auth [api\|subscription] | switch auth mode | | /resume [id] | reload a past session (no arg = picker) | | /bg <cmd> | run a shell command as a background task | | /tasks | view / kill background tasks | | /yolo | toggle auto-approve for writes/bash | | /clear | reset the conversation | | /help | list commands + keybindings | | /quit | exit |

Writes and bash prompt for permission (y = allow once, a = always/yolo, n = deny); reads auto-allow.

Flags

| Flag | Meaning | | --- | --- | | --provider <anthropic\|openai> | Provider (default: anthropic) | | --model <id> | Model override (default: provider's default) | | --auth <api\|subscription> | Force auth mode (default: auto-detect, api-env first) | | --cwd <dir> | Working directory (default: process.cwd()) | | -p, --print "<prompt>" | Headless: run one prompt, stream to stdout, exit | | --resume [id] | Reload a past session (no id = picker in the TUI) | | -c, --continue | Continue the most recently updated session | | --yolo | Auto-allow gated actions (writes/bash) — no prompts | | -h, --help | Usage |

Architecture

src/
  types.ts        # THE contract: Message/Tool/Provider/Auth/Session interfaces
  providers/      # Provider registry + factory; anthropic.ts, openai.ts adapters
  auth/           # AuthResolver per provider → Credential (api key | oauth token)
  loop/           # createSession: turn loop, tools, compaction, cost accounting
  tools/          # read/write/edit/bash/glob/grep/ls, each schema + permission gate
  models/         # listModels/fetchModels: live /v1/models discovery (cached)
  tasks/          # background bash task runner (start/list/output/kill)
  session/        # session persistence + command history (~/.maya-harness)
  tui/            # full-screen Ink app: App + ScrollView + AutocompleteMenu +
                  #   TasksView; lines/viewport/autocomplete/edit/transcript (pure)
  cli.ts          # arg parsing → auth → provider → session → TUI or headless
  prompt.ts       # default system prompt

Data flow: cli resolves a Credential via resolveAuth, builds a Provider via createProvider, and constructs a Session (createSession). The TUI (or headless loop) calls session.run(userText) and renders the streamed LoopEvents. The loop streams from the provider, runs any tool calls through the permission gate, feeds results back, and repeats until the model ends its turn.

Adding a provider

1. Add a src/providers/<name>.ts exporting create<Name>Provider(cred): Provider.
2. Add a catalog entry (models + pricing) and a case in createProvider.
3. Add an AuthResolver in src/auth/.

License

MIT