@averyyy/pi-client

Client CLI for connecting Pi to a pi-server instance.

Install

npm i -g --ignore-scripts --legacy-peer-deps @averyyy/pi-client

--legacy-peer-deps avoids npm peer override warnings when upstream Pi is already installed globally.

Use

Connect to your server:

PI_SERVER_URL=https://pi-server.example.com pi-client

Send one prompt and exit:

PI_SERVER_URL=https://pi-server.example.com pi-client -p "Say exactly: ok"

Send a file or folder to the server:

PI_SERVER_URL=https://pi-server.example.com pi-client send /path/to/file-or-folder

The server saves it under its configured upload directory, which defaults to ~/.pi/upload_files.

Start the browser UI:

pi-client install npm:@averyyy/pi-tau-codex
# or: pi install npm:@averyyy/pi-tau-codex
PI_SERVER_URL=https://pi-server.example.com pi-client web

The web command starts pi-client in Tau mirror mode. Install the standalone @averyyy/pi-tau-codex extension into the shared ~/.pi/agent settings first. Tau listens on http://127.0.0.1:1838 by default.

Server Auth

If your server uses an auth token, set it on the client:

PI_SERVER_AUTH_TOKEN=your-token PI_SERVER_URL=http://127.0.0.1:4217 pi-client

Use matching client and server releases. Provider credentials configured on the server are sufficient for chat, image generation, classifiers, and compaction. Client-side credentials, environment overrides, and request headers are forwarded when supplied.

The authenticated model catalog includes chat, image, and classifier models with availability tracked separately for each model type. Extensions and codemode scripts use the normal model registry APIs; inference runs on the server. MCP connections, tools, and provider login remain in the client process.

PI_CLIENT_MAX_REQUEST_KB=64 limits each client HTTP request body to 65,536 UTF-8 bytes. Larger prompts, tool definitions, images, operation payloads, and session trees use the shared chunk transport.

Prompt Cache Warming

Cache warming replays eligible requests through pi-server with the original prompt, tools, model, and cache options, limiting output to one token. The default streaming mode warms during active agent runs. To also warm while waiting for input, set this in your global Pi settings:

{
  "cacheWarming": "idle"
}

The server confirms the cache retention used by each request, and the client uses that policy for refresh timing and cost decisions. The model must declare a prompt-cache lifetime, and the expected savings must reach $0.05. /session shows the warming decision. Set cacheWarming to "off" to disable it. Refresh usage counts toward session totals without adding model messages; cancellation and changed model/context stop warming. Requests use the same 64 KiB transport limit.

Devin Pro

Update both @averyyy/pi-client and @averyyy/pi-server to a version with native Devin support. Start pi-client with your server connection configured, then:

/login devin
/model devin/swe-2-medium

Login opens Devin's browser OAuth flow and saves the session token in Pi's normal credential store. The account catalog loads after login; pi-client --list-models swe-2 lists the discovered SWE-2 variants. Choose the exact variant (for example swe-2-high or swe-2-max) to change effort. Known configured SWE-2 IDs remain selectable before account discovery finishes; successful discovery supplies the account's models and limits, and discovery failures are reported.

Authentication and model discovery may contact Devin from the client. Model inference, including compaction, goes through pi-client → pi-server → Devin. Direct Devin inference rejects PI_SERVER_MODE=true. Pi continues to own tools, agent loops, session history, and subagents. No Devin CLI installation or provider extension is needed. Remove conflicting Devin extensions before enabling the native provider.

The integration exposes concrete models, not Devin's Fusion/router harness. A revoked session requires /login devin again; no refresh endpoint or token lifetime is invented.

Protocol references: pi-devin, pi-devin-provider, and the Devin protocol descriptors.

Update

pi-client update installs the latest client and server packages without stopping active client sessions. Existing sessions block new prompts until you run /reload; /reload restarts that session on the new runtime and resumes its persisted history.

Related Package

Install the server separately:

npm i -g --ignore-scripts @averyyy/pi-server