Building Blocks

emailEmail for your AI engineering assistant.

An MCP server that gives Claude the ability to send and receive email. Verifies PGP signatures on inbound mail against a four-level trust model — trusted, verified, untrusted, unverified — so the agent knows how much to trust what it reads. Supports Proton Bridge over STARTTLS, external SMTP hosts, and the Resend API. Written in Go.

• MCP tools covering list, read, send, move/archive, batch move, download attachments, verify signatures, inspect MIME, classify trust, list folders, address book, whoami, switch identity, and inbox polling
• Four-level PGP trust model — trusted (Proton E2E), verified (valid PGP), untrusted (bad PGP), unverified (no signature)
• Multi-identity via ethos — identity resolved per-request from ethos sidecar
• Two-dimensional trust — transport trust + identity permissions (rwx per contact)
• Proton Bridge native — IMAP STARTTLS for reading, SMTP for sending, Resend API fallback
• Credential isolation — secrets resolved at runtime from OS keychain, never stored in config
• Slash commands — /inbox, /mail, /send, /contacts, /beadle — plugin-native UX beyond raw MCP tools
• Repo-scoped mailbox — an X-Beadle-Repo header keeps each project's mail separate, or read across all with --all-repos

coordinationTeam communication for engineers who never leave the terminal.

Resurrects the BSD Unix communication vocabulary as MCP-native slash commands over a NATS relay. Humans and AI agents show up side by side — no separate app, no browser tab, no context switch. Self-hostable via a relay Docker image when you don't want to share the public demo relay.

• BSD Unix vocabulary: /who, /finger, /write, /read, /wall, /talk, /plan, /mesg, /last, /tty
• Interactive REPL — biff launches a terminal client with readline, real-time notifications, and modal talk
• Humans and autonomous agents appear side by side in /who
• Status bar — live unread count, wall broadcasts, talk messages wrapping your existing status line
• /talk for real-time conversations with ≤2s notification latency
• Cross-machine messaging via NATS relay

visualsA visual output surface for AI agents.

A Hub process and a separate Display renderer, connected by Unix socket IPC — agents send JSON element trees via MCP tools, the Hub relays them to the Display at 60fps. Dozens of element kinds including interactive controls (sliders, checkboxes, combos, color pickers), data visualization (tables, plots, markdown), and layout nesting (windows, tabs, groups). Incremental updates patch elements by ID without replacing the scene. The visual counterpart to Vox.

• Dozens of element kinds — text, buttons, images, sliders, tables, plots, markdown, draw canvases, and more
• Persistent dismissable tabs — each show() opens a new tab, dismiss individually with close button
• Layout nesting — windows contain tab bars contain groups contain any element, arbitrarily deep
• High-level tools — show_table for filterable data explorers, show_dashboard for metrics and charts
• Interaction events — clicks, slider changes, menu selections queue as events the agent reads via recv
• Auto-spawn — display server starts on first connection if it isn't running
• Lux applet for Beads — filterable issue board with detail panel via /beads command
• Named ImGui themes, and a pub/sub topic bus separate from UI interaction events
• Homebrew install (macOS arm64) alongside the curl installer and PyPI package

isolationOne daemon, many sessions — process isolation for MCP servers.

Lightweight Go proxy bridging MCP stdio transport to a shared daemon process via WebSocket. Claude Code spawns a fresh MCP server per session — if the server loads an ML model or holds a database, each session duplicates all of it. MCP Proxy puts a process boundary between them: a ~6MB binary with <10ms startup forwards messages to a single shared daemon. The integration layer that fronts Quarry, Biff, Vox, and Lux — one shared daemon behind every session.

• Process isolation — daemon crashes don't take down Claude Code sessions
• Shared state — one embedding model, one audio device, one connection pool across sessions
• Hook speed — relays JSON-RPC to daemon in ~15ms, within Claude Code's hook budget
• Auto-reconnect — WebSocket keepalive with 5s ping, 2s pong timeout
• Works with any MCP server — never inspects message content
• Config profiles and bearer-token auth for remote daemons, plus a --health flag for launchd, CI, and quarry doctor
• Verified against a Z spec — ProB model checking found no invariant violations across the explored state space

standardsInternal scaffolding and the shared standards our tools are built against.

Mostly internal plumbing: the standards for code quality, CI, and project structure that Punt Labs projects share, plus scaffolding to apply them. Its one broadly interesting idea is the engine-and-clients model — library, CLI, MCP server, and REST from a single codebase — which Punt Kit defines and `punt init` scaffolds.

• Defines the engine-and-clients model: library, CLI, MCP, and REST from one codebase
• Python standards: ruff, mypy, pyright, pytest conventions
• punt init scaffolds new projects from templates; punt seed renders and merges managed sections into existing ones
• punt audit checks compliance against all standards; punt doctor checks install health
• punt release runs a multi-phase release workflow; punt pii scans staged files for personal data
• A pattern library covering common architecture decisions
• Beads issue tracking for per-project and org-wide work

speech & musicVoice and music for your AI engineering assistant.

Text-to-speech and music engine with five providers — ElevenLabs, OpenAI, AWS Polly, and two offline fallbacks (macOS say, Linux espeak-ng) that need no API key. Delivers spoken notifications when tasks finish, chimes when Claude needs input, synthesizes arbitrary text, holds real-time voice conversations, and generates background music with vibe-driven instrumental tracks. Opt-in only: no audio until you enable it.

• Mic API — unmute/record/vibe/who MCP tools with uniform segment input
• Five providers — ElevenLabs (recommended), OpenAI, AWS Polly, macOS say, Linux espeak-ng
• Voice or chime — /mute switches to audio tones with no TTS API calls
• Session vibe — /vibe sets mood, auto-mode reads session signals
• CLI product commands: unmute, record, vibe, on/off, mute, version, status
• Background music — pools of vibe-driven tracks per program, zero-credit rotation and replay, saved as tagged albums
• Conversation mode — live voice calls with mic input, speech-to-text, and streamed Claude replies
• Lux integration — a visual music player and a Vox control panel appear in the Lux display when it's running

recallUnlock the knowledge trapped on your hard drive.

Local semantic search across PDFs, images, spreadsheets, source code, and dozens of other formats. Finds what you mean, not just what you typed. Runs fully local by default — no API keys required — or point it at a remote GPU host when you need one.

• Semantic search across dozens of file formats, including scanned documents via OCR
• Fully local — embedding model downloads once, everything stays on your machine
• Remote mode — point at a GPU host over TLS with token login, or run entirely local
• REST API for scripting outside the CLI or MCP surface
• Named databases keep work and personal content isolated
• Session capture — ambient hooks record web research and preserve the free-form conversation transcript before compaction; an opt-in private repo extension keeps a capture history for agent memory