Invincat

Invincat is a terminal-native AI coding assistant for local repositories. It can inspect and edit files, run shell commands with approval, use web/MCP tools, keep long-term memory, plan before execution, run scheduled tasks, and bridge remote WeCom messages into a project session.

Features

• Work directly from your project directory in a terminal UI.
• Read, edit, and create files with approval-gated tool execution.
• Run shell commands under configurable safety controls.
• Use /plan to review and approve an execution plan before implementation.
• Keep user and project memory across sessions.
• Create recurring or one-shot scheduled tasks in natural language.
• Extend capabilities through MCP tools, skills, and WeCom bot daemon integration.

Quick Start

pip install invincat-cli
cd /path/to/your/project
invincat-cli

On first launch, run /model to configure a provider and model.

Installation

Requires Python 3.11+.

pip install invincat-cli

Install from source:

git clone https://github.com/dog-qiuqiu/invincat.git
cd invincat
pip install -e .

Start

Run Invincat from your project directory:

cd /path/to/your/project
invincat-cli

Model Configuration

After the first launch, run /model to open the model manager.

• Press Ctrl+N to register a model.
• Fill in the provider, model name, API key, and optional base URL.
• Select a model and press Enter to activate it.

You can also provide credentials through environment variables such as:

export OPENAI_API_KEY="..."
export ANTHROPIC_API_KEY="..."
export GOOGLE_API_KEY="..."
export DEEPSEEK_API_KEY="..."
export OPENROUTER_API_KEY="..."

Example for DeepSeek:

export DEEPSEEK_API_KEY="sk-..."

Then register a model in /model with:

Invincat supports a primary model for normal work and an optional memory model for post-turn memory extraction. If no memory model is configured, memory extraction uses the current primary model.

Basic Commands

Goal Mode

Use /goal <objective> when you want Invincat to keep working toward one long-running objective across multiple turns. Goal mode keeps the active objective in context, asks before drifting away from it, and only exits when you complete, cancel, or clear the goal.

Common commands:

Goal state is scoped to the current thread and stored under .invincat/goals/, so an active goal can be restored when the same thread is resumed.

Memory Management

Invincat keeps durable memory in two scopes:

After completed non-trivial turns, a background memory agent extracts useful updates and refreshes the memory context for later sessions. Explicit requests such as "remember this" or "save this" are treated as stronger memory signals. If no dedicated memory model is configured, extraction uses the current primary model; for a session-level memory model, use /model 2 <provider:model>.

Use /memory to inspect and manage memory entries. The viewer supports switching user/project scope, refreshing, sorting, showing archived items, and deleting selected entries with confirmation. Prefer this UI over manual JSON edits because the memory files are managed by the memory subsystem.

Plan Mode

Use /plan when a task needs review before changes are made. In this mode, Invincat first produces an execution checklist instead of immediately editing files or running implementation commands. After you approve the checklist, the main agent executes the approved steps.

Plan mode is useful for risky, multi-file, ambiguous, or architecture-level work where you want to inspect the approach first. The planner is limited to read and planning tools such as file reads, search, web lookup, todos, user questions, and plan approval; implementation tools are reserved for the post-approval execution phase.

Built-in Subagents

Invincat includes focused subagents that the main agent can call through the task tool for isolated, context-heavy work.

These roles are separate so the main agent can keep orchestration and execution focused while delegated agents handle specialized work in isolated contexts and return concise findings.

Built-in Skills

Skills are lightweight capability packs loaded into the agent when a task matches their description. They provide focused instructions, workflows, and supporting scripts without becoming separate agents.

Skill Installation and Configuration

Built-in skills are enabled by default. Custom skills can be added at user or project scope:

Project skills override user skills with the same name, and custom skills override built-in skills. Each skill lives in its own directory with a SKILL.md file that defines its name, description, and instructions.

Some built-in skills rely on document-processing libraries that are not needed for normal coding sessions. Install the matching extras when you want those workflows available:

pip install "invincat-cli[pdf]"
pip install "invincat-cli[office]"
pip install "invincat-cli[all-skills]"

Office conversion and spreadsheet recalculation workflows may also need system tools such as LibreOffice, Poppler, Tesseract, or Node.js packages named by the skill instructions.

For local release checks, install:

pip install "invincat-cli[release]"

WeCom Bot Daemon

Configure a robot on the enterprise WeChat side to obtain the "Bot ID" and "Secret":

https://developer.work.weixin.qq.com/document/path/101463

Invincat can run a foreground WeCom bot daemon for project-scoped remote turns and scheduled-task delivery.

Configure the bot credentials first:

export WECOM_BOT_ID="your_bot_id"
export WECOM_BOT_SECRET="your_bot_secret"

WECOM_WS_URL is optional and defaults to wss://openws.work.weixin.qq.com. Set it only when you need to override the WeCom websocket endpoint.

Start the daemon from the project directory: