DPAI / Docs / Getting started

Getting started

Install the binary, give it a gateway key, and start working. Three minutes.

Last updated 2026

Install

The script detects your platform, verifies a SHA-256 checksum against the release, installs to ~/.dpai/bin, and adds that directory to your shell profile.

curl -fsSL https://install.dpai.sh | bash

Supported platforms are macOS on Apple Silicon, Linux x64, and Windows x64 through Git Bash or WSL. Pin a version or skip the profile edit if you prefer:

# pin a version
curl -fsSL https://install.dpai.sh | bash -s -- --version 1.0.0

# leave shell config alone
curl -fsSL https://install.dpai.sh | bash -s -- --no-modify-path

If you would rather read before you run — and you should — the script is here.

With Bun already on PATH, the package works too:

bun add -g dpai-cli

API key

DPAI needs a key for the Dual Path Intelligence gateway. Get one from the dashboard. Pick whichever of these fits how you work:

# environment — good for CI
export DPAI_API_KEY=your_key_here

# once, on the command line
dpai -k your_key_here

Or save it in ~/.dpai/user-settings.json as { "apiKey": "your_key_here" }. A project .env is read too.

Optional: DPAI_BASE_URL, DPAI_MODEL, DPAI_MAX_TOKENS. A brand that locks its gateway ignores DPAI_BASE_URL.

Running it

dpai                                    # interactive TUI
dpai -d /path/to/repo                   # choose the project directory
dpai fix the flaky test in foo.test.ts  # open with a message

Headless runs take one prompt and exit — for scripts, CI, and schedules:

dpai -p "run the test suite and summarize failures"
dpai -p "summarize the repo state" --format json   # NDJSON event stream
dpai -p "review the repo overnight" --batch-api    # cheaper, delayed
dpai -p "refactor X" --max-tool-rounds 30

--format json emits step-level records — step_start, text, tool_use, step_finish, error — rather than prose.

Sessions persist, in both modes:

dpai --session latest
dpai -s 

Other things worth knowing early: dpai models lists what your key can actually reach with pricing hints, and --verify builds your app, boots it, and runs browser smoke checks in a sandbox with screenshots and video as evidence.

Terminals

The interactive UI is OpenTUI, and it expects a modern terminal. These are the ones tested and recommended:

  • WezTerm — cross-platform
  • Alacritty — cross-platform
  • Ghostty — macOS and Linux
  • Kitty — macOS and Linux

Others may work. Headless --prompt mode has no terminal requirements at all.

macOS. The built-in computer sub-agent drives your desktop, which needs Accessibility permission for your terminal app under System Settings → Privacy & Security. Nothing else requires it.

Settings

FileScopeHolds
~/.dpai/user-settings.json You API key, custom sub-agents, hooks, Telegram, sandbox defaults
.dpai/settings.json Project Model for this project, MCP servers
AGENTS.md Directory Instructions, merged from the git root down. AGENTS.override.md wins per directory

Custom sub-agents go under subAgents in user settings, each needing a name, model, and instruction. The names general, explore, vision, verify, and computer are reserved for the built-ins.

Hooks run shell commands on agent lifecycle events and can block an action outright by exiting 2. Skills live in .agents/skills//SKILL.md. MCP servers are configured from /mcps in the TUI.

Updating and removing

To move to a newer version, re-run the install command:

curl -fsSL https://install.dpai.sh | bash
Note. dpai update exists in the CLI but does nothing on this distribution — it expects a GitHub release feed, and DPAI publishes to object storage instead. Re-running the installer is the supported path. dpai uninstall works normally, and takes --dry-run and --keep-config.