QUESTPIE Autopilot

Features

Complete feature list for QUESTPIE Autopilot. Unified tools, living dashboard, zero-infra architecture, hybrid search, multiple providers, and trust model.

Everything QUESTPIE Autopilot can do today. Organized by what matters most.

Category-Defining Features

Tools-First Agent Model

Agents don't generate text. They call 12 unified tools that produce visible, auditable effects: task, message, pin, search_index, fetch, web_search, and more. Agent thinking is private. Only effects are visible. This is the core architectural differentiator. See Tools.

Living Dashboard

A React app in the company filesystem that agents edit in real time. Custom widgets (widget.tsx), custom pages (page.tsx), theme overrides (theme.css), layout configs (layout.yaml). Changes appear instantly via HMR. Your agents build your internal tools -- no Retool, no deployment, no drag-and-drop. See Living Dashboard.

Zero-Infra Architecture

One Bun process. One SQLite file. No Docker, no Postgres, no Redis, no vector DB, no Kubernetes. Install with bun add -g @questpie/autopilot, run with autopilot start. Run an AI company for ~$0.30/day in API costs. See Architecture.

Trust & Safety

Human Gates

Explicit approval points for merge, deploy, spend, and publish -- defined as type: human_gate steps in workflow YAML. Not bolted-on permissions. The workflow engine enforces them at the right moments in the process.

Deny Patterns

Hardcoded filesystem deny patterns: .auth/, .master-key, .data/, .git/ are always blocked for agents. Per-agent filesystem scoping on top via fs_scope in agent definitions. Agents cannot touch what they shouldn't touch.

Session Attach / Replay

autopilot attach max streams the agent's live session -- like kubectl logs -f for AI agents. Watch tool calls, file reads, file writes, and tool invocations in real time. Ctrl+C to detach; the agent keeps working. autopilot replay <session> for post-hoc review.

Git Audit Trail

Every agent action that modifies the filesystem is automatically committed to Git with a 5-second batch window. The entire company history is a git log. git diff to see what changed. git revert to undo. Fork a company with git clone.

Search & Knowledge

Unified Search (FTS5 + Vector)

Search across tasks, knowledge docs, channels, and agent memory with a single query. Two search engines run in parallel:

EngineTechnologyBest For
Full-textSQLite FTS5Exact keyword matches, task IDs, agent names, file paths
SemanticlibSQL native vectors (DiskANN)Conceptual similarity, natural language queries, related content

Results from both engines are merged and ranked using Reciprocal Rank Fusion (RRF, k=60).

Embedding Service

Vector embeddings power semantic search and knowledge retrieval. Automatic via OpenRouter — no extra keys needed:

ModelNotes
nvidia/llama-nemotron-embed-vl-1b-v2:freeDefault. Multimodal (text + vision). FREE.

Configurable via settings.embeddings.model in company.yaml. Uses the same OPENROUTER_API_KEY.

Agent System

Multiple Agent Providers

ProviderConfig ValueFeatures
TanStack AI + OpenRoutertanstack-aiPrimary. 300+ models (Claude, GPT, Gemini, Llama). File tools, MCP, sandboxing, hooks, sub-agents.

Persistent Memory

Per-agent memory: facts, decisions, mistakes, patterns. Extracted automatically after every session. Private -- no agent reads another agent's memory. Cross-agent sharing only through channels and message.

6-Layer Context Assembly

Before every session: identity (role, rules) + company state (role-scoped snapshot) + memory (facts, decisions) + task context (spec, plan, code, history) + skills discovery + tool list. Role-scoped company snapshots ensure agents see only what they need.

Skills System

Agents learn new capabilities via SKILL.md files in the filesystem. Same kernel, different skills, different company. Want the DevOps agent to know your k3s cluster? Write a skill. Want the marketer to know your brand voice? Write a skill. See Skills.

Communication & Transports

Transport Registry

Pluggable notification transports with per-priority routing and quiet hours:

TransportStatusNotes
TelegramAvailableBuilt-in adapter + webhook handler. Two-way messaging.
DashboardAvailableSSE-based realtime notifications.
SlackComing SoonPlanned.
EmailComing SoonPlanned.

SSE Realtime

All realtime communication uses Server-Sent Events (SSE). Session streaming, dashboard updates, and activity feed -- single HTTP connection per subscriber, no polling.

Language Configuration

Multi-language support via language and languages fields in company.yaml. Agents respond in the configured language. Semantic search works across languages via embedding providers.

Artifact System

Agents create live previews -- React apps, HTML pages, API docs. The artifact router spins up a Vite dev server with cold-start, port pool management, and idle timeout. Get a live URL to what your agent built before it ships. See Artifacts.

Workflow Engine

YAML-defined state machines with conditional transitions, review steps, and human gates. Three built-in workflows (development, marketing, incident). CEO agent owns workflow definitions; any agent can propose changes with evidence. See Workflows.

Integrations

Any external service follows the same 3-part pattern: add a secret, add a knowledge doc, agent calls the API. No integration modules, no plugins, no code changes. See Integrations.

On this page