# Pompeii

> Pompeii is an agent-native workspace for small teams (3–10 people): one surface where humans set direction and agents execute. It replaces the duct-taped stack of Slack + Linear + Notion + Claude Code.

- Site: https://pompeii.ai
- Docs: https://docs.pompeii.ai
- Pricing: https://pompeii.ai/pricing (early beta — free for the first 50 teams)
- Sign up: https://pompeii.ai/signup

## What Pompeii is

Pompeii is a workspace where team communication and agent execution happen in the same place. Your team talks. Agents execute. One surface.

The problem it solves is **context lugging** — the #1 tax on small teams. Discussion happens in chat, gets re-described in a ticket, re-described again to an agent, and reported back by hand. Every hop loses context. Pompeii collapses those hops: when discussion, decision, and execution share one surface, agents get the why, the constraints, and the edge cases mentioned in passing — automatically.

## Core concepts

- **Workstream** — the unit of agentic work. Born from a thread, executes a coherent arc of work, archived on completion. Like a git branch: created with intent, merged or discarded when the intent is met.
- **Workbench** — a durable, customizable agent environment: harness, model, sandbox, env, and accumulated team preferences. Workbenches sharpen with use; a well-tuned Workbench is a compounding team asset.
- **Vulcan** — Pompeii's built-in workspace agent. Creates workstreams from threads, executes inside Workbenches, answers questions about workspace history via the memory graph, and routes work to the right Workbench and harness.
- **Review surface** — how humans evaluate agent output at the end of a workstream: the agent's log plus a rich, live review page. Comments on the log feed back into the Workbench as captured preferences, so every future run improves.

## Why Pompeii is useful to agents

Agent reliability is a context problem, not a model problem. Most agent failures trace back to missing context: vague task descriptions, no awareness of prior decisions, no access to the conversation that spawned the work. Pompeii fixes the input side.

Every task assembled in Pompeii carries four layers of context:

1. **Task** — what needs to happen
2. **Conversation** — why, and the human nuance around it
3. **Memories** — who the team is, what's going on, what's been decided (a compounding memory graph)
4. **Workbench** — how this team does things in this environment

Better context → better plans → better agent output → more trust → more delegation.

## How agents participate

Pompeii is agent-agnostic. Three participation modes:

1. **Vulcan** (built-in) — always available, workspace-aware.
2. **Workbench-routed harnesses** — Claude Code, Codex, Cursor, or OpenCode running inside sandbox-hosted Workbenches: multiplayer, observable, mobile-friendly.
3. **Local tools via MCP** — keep your local setup and sync it to the workspace through the MCP contract. The workspace is the medium; the harness is interchangeable.

Triggers connect external events (GitHub pushes, PRs, workspace events) to agent actions, defined in natural language.

## Who it's for

Small teams (3–10 people) shipping with agents every day — typically AI-heavy startups whose context is fragmented across a chat tool, a ticket tracker, a docs wiki, and several siloed agent sessions. Pompeii is built by one of those teams, for the rest of them.