Extend a running agent

Browser automation backed by vercel-labs/agent-browser: tools (incl. page→PDF) + skill + workflows + a live, drivable browser panel. Ships in-tree, off by default; needs the agent-browser CLI on PATH.

Render HTML/SVG/Mermaid/React into a sandboxed iframe (show_artifact). Ships built-in with protoAgent.

Turns 'implement X, here are the tests' into a TEST-VERIFIED solution via a difficulty-gated search ladder — gated on the tests passing, not an LLM judge. Off by default; needs a declared delegate.

Office-document skills for the Cowork archetype — Word, Excel, PowerPoint, and PDF as real files — plus a /daily-brief, drop-folder watches over your fenced work folders, schedule, memory-consolidation, and writing-voice habits (ADR 0083). Ships in-tree, on by default.

Engineering rituals you invoke by typing them — /grill, /standup, /code-review, /due-diligence — plus an ADR-authoring skill and a skill_writer subagent. Prompt-only: no tools, routes, network, or filesystem access.

A hot-swappable registry of the agents and endpoints your agent can talk to — fleet A2A agents, OpenAI-compatible endpoints, ACP coding agents — managed from the console.

Answer questions about protoAgent from its own documentation — a keyword index over the bundled docs, docs_search / docs_read tools, and a console reader view. On by default.

The skill pack behind the Engineer archetype — onboard a repo (its README, manifest and a code-linked architecture diagram, then a guided tour), then debug one checkpoint per turn: you write the fix, the agent reviews. Prompt-only; off by default.

A Python interpreter the agent runs code in — headline use is programmatic tool-calling: call several tools, compose the results in code, return only what matters. Ships disabled: a scrubbed-env subprocess is isolation, not a true sandbox.

The agent records its own friction — missing or awkward tools, confusing errors, wrong paths — alongside auto-captured tool errors, so the improvement backlog writes itself from real use. Open friction is projected into the agent's working state, so it acts on its own backlog instead of re-reporting it. On by default.

A minimal starting point — one tool, one bundled skill, and a console view. Copy it to scaffold your own plugin by hand.

A single shared markdown note the agent and the operator both read and write, with a live-preview editor the plugin serves itself. The flagship reference plugin — on by default.

Live diagram of the agent fleet — every delegate a node, a directed edge wherever one agent can delegate to another (dashed where the hub supervises a member). Draws all three types: A2A peers crawled outward, plus ACP coding agents and OpenAI-compatible model endpoints as leaves badged with their owner's probe result. Crawls the effective roster with cached, concurrent probes; the view paints instantly from a stale-while-revalidate snapshot. Off by default.

Build your own plugins — the authoring kit + reference plugin. A scaffold tool, an architect subagent, the building-plugins skill, and a console guide, in one bundle.

Run your agent as a Telegram bot — inbound DMs and groups, replies back out. A tiny reference communication plugin.

Declarative multi-step subagent workflows — chain subagent steps, some running in parallel, into repeatable recipes defined once as YAML. Off by default (lean core).

The deterministic half of the PR-review QA tier — budgeted protoPatch structural reviews as a fifth review-panel member (ADR 0078).

Board-driven coding orchestration — a beads board + ACP spawn loop + planning layer + console view.

protoAgent archetype: Brand & Launch — a draft-only launch manager for any app: campaign plan, scripted recordings, GIFs, cards, copy drafts, operator approval

Read AND write GitHub over the gh CLI — PRs, issues, diffs, CI, repo files; plus gated create/edit/merge/close, comments, labels (per-agent write gating). Ships a read-only Issues/PRs board, a util-bar New-issue widget, and the user-only /issue command.

A full terminal (xterm.js + a real PTY over WebSocket) as a console view, themed from the design system.

protoAgent archetype: Engineer — a hands-on pair-programming navigator: onboards any repo, points at the evidence, you write the fix and it reviews

Your career coach & job-hunt copilot — strategy, mock interviews, honest CV feedback, and offer/negotiation coaching, plus an autonomous evaluate→tailor→apply pipeline. A full-surface plugin showcase.

Multi-team orchestration — one PM across many project boards over A2A (ADR 0055).

An adaptive learning wiki — an LLM-maintained, interlinked concept wiki plus a learner ledger: per-concept strength, misconceptions, and FSRS-scheduled review.

Drive native GUI apps on this machine — snapshot a window's accessibility tree, then click/type/scroll by element without stealing focus. Wraps the cua-driver binary.

Trading research as a full bundle — market data, backtesting, factor IC, a behavioral journal, a gated paper broker, the research desk, and a Quant Desk dashboard.

protoAgent archetype: Portfolio Manager — manager-of-teams for the protoLabs.studio autonomous fleet

protoAgent archetype: Project Manager — board-driven single-repo PM: reads deeply, never holds the keyboard; every change ships through a verdict-gated board of coding agents

An autonomous SpaceTraders fleet commander — crew subagents, navigation/mining/trading tools, workflows, and a Fleet dashboard. The reference full-bundle plugin.

Run your agent as a Discord bot — inbound DMs + @-mentions, outbound send/read/react tools.

Explore Claude Code's on-disk state — sessions, memory, scratchpads — and translate its skills, subagents, and MCP config over to protoAgent.

protoAgent archetype: Cowork — document skills, Google connectors, schedules, memory, and Claude Code state import (ADR 0083)

A subreddit feed view plus tools to read and interact with Reddit as you.

Google Workspace for your agent, built to grow — Gmail (read + draft-only, never sends, plus mailbox hygiene) and Calendar (read + free/busy + search).

Native image generation and editing via the protoBanana gateway — text-to-image, instruction edits, and identity-preserving edits over self-hosted ComfyUI workflows.

But can it run DOOM? Play shareware DOOM in a console view via WebAssembly (prboom).

PM toolkit — 65 PM skills, a provenance-enforced PM Brain, specialist subagents, and a brain dashboard.

Give your agent access to Monarch Money via its MCP server.

A live window into your design system — browse tokens and components in the console (gallery backed by your own published Storybook), play with a component's props, and give the agent the real vocabulary to design against.

Organic social marketing that never posts — a brand kit every draft follows, nine platforms' hard limits and dated norms, a content queue, a linter, and a copy-ready export pack, with a writer + editor + researcher crew. The Brand & Launch archetype installs it.

A real 26×15 Blood Bowl pitch in the console — drag positionals onto squares, hover for stats, and set scenarios up by hand or by asking the agent. Ships roster data for 30 teams.

Let the lead agent spawn a CLI coding agent — protoCLI, Claude Code, Codex, or Gemini CLI — and hand it a real coding job over the Agent Client Protocol.

A versioned prompt library and a streaming playground view, with agent tools to manage both.