Gato

An AI sales agent harness built on Pi.

🐒 Overview

Gato turns Pi into a conversational sales representative for a messaging channel: it keeps context across sessions, personalizes each conversation from your contact data, and acts on your business systems through MCP tools. Everything that shapes the agent — its prompt, tools, and behavior — is plain files you can read and edit.

🍌 Features

• 💬 Memory across sessions — carries a rolling, language-aware summary of past interactions into every conversation, so returning customers are met with full context.
• 🎯 Focused context — keeps only recent conversation in view (configurable, default 60 days) and strips stale tool chatter from earlier turns, so replies stay fast and on-topic even in long-running relationships. Each part can be toggled independently.
• 🔌 Connects to your systems — exposes your business tools over MCP, so the agent can look things up and take action — orders, inventory, records — as part of the conversation.
• 🧩 Drop-in skills — add a skill (a folder of instructions) to your skills directory and the agent discovers it automatically, pulling it in whenever a conversation calls for it — no wiring or code.
• 💾 Persistent conversations — every conversation is saved automatically and can be resumed, so each customer's thread is preserved and you pick up exactly where you left off. Closed sessions are gzipped to keep disk usage low; /resume-compressed reopens one.

📦 Requirements

• Pi installed
• Credentials for your model provider (Gato defaults to the hebo provider; change it in .pi/settings.json)

🚀 Quick Start

git clone https://github.com/8monkey-ai/gato
cd gato
cp .env.template .env   # set your values, then `source .env` (or use direnv)
pi -c                   # start — continues the most recent conversation

Tip

Run pi -c from the repo so Pi picks up the project .pi/ — the system prompt, settings, packages, and prompts that make up Gato. Pi installs the declared packages automatically on first startup.

🗂️ Your Content

Gato ships the agent; you provide the business context.

What	Where
Business guidelines	AGENTS.md in the repo root
Skills	.pi/skills/
Tool servers	.pi/mcp.json
Sessions	~/.pi/agent/sessions/ (auto-managed by Pi)
Summary	~/.pi/agent/compact.md (auto-generated, don't edit)

mcp.json format:

{
  "mcpServers": {
    "crm": {
      "url": "https://your-mcp-server/endpoint",
      "lifecycle": "eager",
      "directTools": ["check_inventory", "create_order"]
    }
  }
}

Note

Set directTools to true to register everything the server exposes. See the pi-mcp-adapter docs for all options.

⚙️ Configuration

Set via environment variables (Pi does not auto-load .env — source it or use direnv):

Variable	Default	Effect
PI_HISTORY_DAYS	60	Days of chat history kept in context
PI_COMPACT_STALENESS_DAYS	3	Age before the conversation summary is regenerated

Context management comes from the @8monkey/pi-context-history package. Each feature is on by default and can be switched off by setting its flag to false or 0:

Flag	Default	Feature
PI_TRIM_HISTORY	on	Drop messages older than PI_HISTORY_DAYS
PI_STRIP_TOOL_HISTORY	on	Strip prior turns' tool calls and results
PI_COMPACT	on	Keep a rolling summary and fold it into the system prompt
PI_APPEND_MESSAGE	on	/add-user-message and /add-assistant-message commands

Model provider, thinking level, and other defaults live in .pi/settings.json.

🧪 Customizing

Gato's behavior comes from three Pi packages declared in .pi/settings.json — @8monkey/pi-context-history for cross-session memory and context trimming, pi-mcp-adapter for MCP tools, and @8monkey/pi-session-gzip for compressing closed sessions — plus the summary prompt in .pi/prompts/compact.md. Pi installs missing packages automatically on startup; edit the prompt, swap packages, or add your own extensions under .pi/extensions/.