overview

The toolbox around a coding agent.

Agent Utils is a Pi package and a handful of standalone tools that give a coding agent a voice, eyes, hands and a fleet. Speech goes in and out; images and screens get shared both ways; questions reach whichever device you are holding; sessions compact, resume, fork and update themselves; and one command shows every machine at once.

41Pi extensions
5standalone tools: ag, pi-session, skill-search, web-search-mcp, linear-extra-mcp
1install: pi install git:github.com/harryaskham/agent-utils@v1

The catalogue

Every capability, by family. Each family has its own page.

speech

/readRead aloud what you send, as you send it.
/tts · /narrateSpeak replies verbatim, or one-line summaries around tool batches.
/stt · /pttVoice input with local voice detection or push-to-talk.
/rtA Realtime voice agent as a normal Pi provider.
/tts-queueOne machine-wide playback queue for every session.
/audio · audio_shareNamed audio providers; an audio reel for clips agents share.

Go to the Speech page

screens

kitty_image_preview_*A persistent image gallery in the terminal, using kitty graphics.
shared imagesDurable, checksummed archives of every image an agent shares.
image 413 recoveryHalve an oversized image and retry, once, automatically.
/gfx · /eink · chipsEditor chrome, themes and an e-ink mode for tablets.
tendril_* · macOSDesktop screenshots and window placement, shared into context.
android_* · xvfb_* · firecracker_vm_*Emulators, virtual displays and microVMs for agents to drive.

Go to the Screens page

sessions

interactive_choiceSpoken, keyboard, voice and remote answers to one question.
async_choiceAsk without blocking; answers arrive as steering messages.
/compact-next · /endless · /afterCompact at the next safe point, keep going, or schedule a prompt.
/update · /restartUpdate packages, reload tools, or re-exec the session in place.
/m · /effortSwitch models and effort, with guarded self-selection for agents.
/session-fork · /nix devshellFork a session to another machine; enter a flake shell on demand.

Go to the Sessions page

fleet & work

agSpeech, images, audio, questions and health across every machine.
ag dashAll of it as one live local web page.
search_webLive web search with citations, through Copilot's Responses API.
app_automation_* · slack_*Snapshots of Slack, Outlook, Teams and calendars; native Slack tools.
skill-searchOne meta-tool to discover skills and MCP servers on demand.
linear-extraAn MCP wrapper with drafts for Linear.

Go to the Fleet & work page

How it is built

Opt in, fail safe: Speech, auto-updates, agent queue tools and display helpers are off until asked for, and guarded when on: a missing provider or display is a clear message, never a hang.
Local first, fleet aware: Everything works in one Pi session on one machine. Durable archives on each machine then let ag give a fleet-wide view without a central server.
Independent of the host: Every extension works on its own. When the Paratenic session host is present, the same features are also offered to its remote clients through the Agent Host Protocol, using one state machine for local and remote input.
Agents and operators are different users: Operator commands are unrestricted; agent-callable tools such as model self-selection, speech mute and question answering are narrower or need confirmation.