IMI

IMI is the AI product manager for AI agents.

You open Claude Code in the morning and your agent has no idea what happened yesterday. You re-explain the codebase, the decisions, the blockers, and what matters now. Then you do it again in the next session. IMI fixes that by keeping the thinking layer inside your codebase so every session starts with context instead of drift.

Website: https://useimi.com

IMI stores the parts agents forget:

• what you're building
• why you're building it
• what decisions were made
• what got blocked
• what changed between sessions

That gives you a simple loop:

1. You steer in natural language.
2. IMI stores the intent, decisions, and direction.
3. Your next agent session picks up from there without a re-brief.

IMI is not another task board. It is the PM layer that sits between you and your coding agents.

bunx imi-agent@latest

That downloads the binary and runs imi init in your project.

Or via curl:

curl -fsSL https://useimi.com/install | bash

Make sure ~/.local/bin is in your $PATH:

echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc && source ~/.zshrc

Install the IMI plugin into your agent CLI so the session gets IMI context automatically.

Claude Code

/plugin marketplace add ProjectAI00/imi-agent
/plugin install imi

GitHub Copilot CLI

/plugin marketplace add ProjectAI00/imi-agent
/plugin install imi

Manual skill install

npx skills add ProjectAI00/imi-agent@imi

These are the commands that matter most when you want to understand the project and steer it:

imi context   # what we're building, active work, key decisions
imi think     # are we still building the right thing?
imi plan      # full goals and task list
imi check     # verification state

To capture human thinking directly:

imi decide "what" "why"   # firm decision with reasoning
imi log "note"            # direction, observation, or concern

In practice, the human should mostly talk to their agent in natural language. The agent reads IMI, writes back to IMI, and keeps the project aligned without making the human manage command syntax all day.

IMI remembers your goals, decisions, blockers, and project direction so your agents can resume work across sessions, teammates, and tools without starting from zero.

Ask your agent things like:

• what are we building?
• how is it going?
• are we still aligned?
• what changed since yesterday?

Under the hood, the agent can use IMI to read context, reason over direction, and persist new decisions or notes back into the repo-local state.

IMI is the state and alignment layer. Execution tools can plug in underneath it.

• Hankweave for task execution and checkpointing
• Entire for session audit and rewind
• Claude Code, GitHub Copilot CLI, Cursor, Codex as the session layer on top

The point is not which execution tool you use. The point is that the project context survives.

• Rust — single binary
• SQLite — local persistent state in .imi/state.db
• Works with Claude Code, GitHub Copilot CLI, Cursor, Codex, and terminal-based agents

The repo includes prompts for the main IMI modes: