scritty

Product Introduction

1. Definition: Scritty is a local-first, agent-agnostic terminal emulator and memory substrate for AI CLI agents. Technically, it is a PTY-based terminal application that functions as a cross-agent conversation capture, indexing, and retrieval system.
2. Core Value Proposition: It exists to solve AI context fragmentation by automatically capturing, indexing, and making searchable every interaction with any AI command-line interface (CLI) agent—such as Claude Code, GitHub Copilot CLI, Codex CLI, Antigravity, Aider, and Ollama—into a single, unified, and user-controlled corpus. Its core value is persistent, shareable memory across all AI tools.

Main Features

1. Agent-Agnostic Capture & Process-Level Detection: Scritty captures all terminal I/O, detecting the specific AI agent (e.g., Claude, Copilot) via the running process, not prompt sniffing. It automatically tags each exchange with the provider. This works without requiring plugins, SDKs, or vendor cooperation, as it operates at the OS terminal boundary.
2. Unified Hybrid Search Corpus: All captured conversations are indexed into a single local corpus. It employs a hybrid search system combining vector embeddings (using dual offline ONNX models tuned for code and prose) and keyword search, fused with Reciprocal Rank Fusion (RRF). Users query this corpus via Ctrl+Shift+M directly in the terminal.
3. Memory Substrate via MCP & CLI: The indexed corpus is exposed as a MemoryService accessible three ways: 1) In-terminal panel search, 2) A full Model Context Protocol (MCP) server (over stdio or Streamable HTTP with bearer-token auth), allowing agents themselves to query past conversations, and 3) A scriptable CLI (scritty memory ...) for shell pipelines.
4. Multi-Device Session Sync & PWA: The desktop terminal embeds a token-secured web server. The live PTY session, including scrollback and memory panel, can be opened in any browser or as a Progressive Web App (PWA) on desktop or mobile, maintaining byte-for-byte sync. A QR code pairing system enables secure phone access over LAN or Tailscale Funnel.
5. Unified Rule Application & Control Plane: Scritty assembles and injects user-defined rules (from prompt.toml), enforcement constraints, goal states, and vendor-specific rule files (e.g., .claude/CLAUDE.md) into every message sent to any agent, ensuring consistent instruction application. For teams, a central control plane offers per-tenant isolation, federated search with role-based access, and admin features.

Problems Solved

1. Pain Point: AI Context Silos and Loss. Conversations with different AI CLI agents are trapped in per-vendor, per-session logs with no cross-reference, leading to repeated explanations, lost decisions, and inefficient workflows.
2. Target Audience: Software Developers and Engineering Teams who regularly use multiple AI coding assistants (Claude Code, Copilot CLI, etc.) across different phases of a project. DevOps Engineers and Technical Leads who need audit trails and cross-team knowledge sharing of AI-assisted work.
3. Use Cases: Project Continuity: A developer uses Claude for architecture, Codex for tests, and Copilot for boilerplate over a week; Scritty allows any agent (or the developer) to search "auth decisions" across all sessions. Remote Collaboration & Monitoring: A lead can view a team member's live AI-assisted terminal session via browser for guidance. Offline Development: Using local Ollama models with full access to the previously cloud-agent-generated project memory.

Unique Advantages

1. Differentiation: Unlike Warp (captures only its own agent), Cursor/Continue (editor-based, indexes code not conversations), or memory-as-a-service libraries (mem0, Zep—require you to build the agent), Scritty captures any CLI agent at the terminal level and provides a ready-to-use, searchable memory system. It occupies the unique intersection of terminal emulator and agent-agnostic memory layer.
2. Key Innovation: The Terminal as the Universal Capture Point. Its fundamental innovation is leveraging the terminal (PTY) as the single point through which all CLI agents must communicate, enabling automatic, vendor-agnostic capture without integration overhead. This is combined with a pluggable vector backend (embedded, Qdrant, pgVector) and exposing the corpus as a standard MCP server, transforming captured data into a live substrate for both humans and AI agents.

Frequently Asked Questions (FAQ)

1. Where is my AI conversation data stored when using Scritty? Scritty is a local-first application; by default, all captured conversations, embeddings, and indexes are stored solely on your local machine. No data is sent to the cloud unless you opt into a team plan with central federation.
2. How does Scritty work with offline AI models like Ollama? Scritty captures conversations with local Ollama instances the same way it captures cloud agents. The hybrid search index is built and queried locally using ONNX runtime, enabling full search and memory recall for offline AI work within the same unified corpus.
3. Can I use Scritty's memory with AI agents running in my IDE or browser? No, directly, it is designed for CLI (Command Line Interface) agents. However, agents that support the Model Context Protocol (MCP) can connect to Scritty's MCP server (over stdio or HTTP) to query the memory corpus as a resource.
4. What is the difference between Scritty's Personal and Team/Pro plans? The Personal plan is for solo developers, runs fully locally, and keeps all data on your machine. Team plans (Pro, Pro Plus, Enterprise) introduce a central control plane, federated search across team members with tenant isolation, SSO/SAML, audit logs, and the option for managed or self-hosted deployment.
5. How does Scritty's pricing work? Scritty uses a consumption-based model metered by "exchange-pairs" (one user prompt + one AI response), not per-seat. Overage rates apply after the included monthly quota. This aligns cost with actual usage volume across a team.