Your project’s memory, for any agent.
AHP+ is the alternative to MADRE for people who would rather stay in their IDE. It keeps your project’s state, decisions and test results in your repository, with Git, so Codex, Claude Code, Cursor or OpenCode pick up exactly where the last one stopped, even if you change model, editor or computer.
MADRE or AHP+
Which one is for you?
Both are free, open source and run on your machine. The difference is where you work.
| MADRE | AHP+ | |
|---|---|---|
| Where you work | In a room that opens in your browser. | In your usual IDE: Codex, Claude Code, Cursor or OpenCode. |
| How agents collaborate | They talk in the same chat and pass the turn. | Each from its own chat, reading the same project state. They can also open a shared room. |
| What it keeps | The conversation’s memory: decisions, facts and preferences. | The project’s state: phase, decisions, test results, checkpoints and handoffs between agents. |
| Where it lives | On your machine. | Inside your repository (in .ahp/), so it travels with Git to other machines. |
| Best if… | You want several agents to weigh in and work in one place. | You often switch IDE, account or computer and don’t want to lose the thread. |
| You need | Node 22.5 or later. | Git, Node 20 or later and a repository with at least one commit. |
Don’t want to choose? MADRE ships AHP+ as an optional module: turn it on in MODULES and use /ahp status, /ahp check and /ahp context from the room.
What it keeps
What your next agent needs to know.
A chat is forgotten when you close it. AHP+ writes down, in your repository, what matters to carry on.
The project’s state. Which phase it is in, the goal, what comes next and what is blocked.
The decisions. What was decided, who confirmed it and whether it still stands.
The test results. Which commands ran and what they returned. A “pass” needs evidence.
Checkpoints. Points in the session you can go back to.
Handoffs. One agent handing the work to another, which the next one verifies before carrying on.
How certain each fact is. Whether it is verified, confirmed by you or just assumed. So nobody mistakes a guess for a fact.
Install
It installs with one command.
From your project’s root folder. You need Git, Node 20 or later and a repository with at least one commit; the project doesn’t have to use Node.
-
Install it in your project
Pins AHP+ in your repository, sets up Codex and Claude Code and checks that everything is fine.
npx @jossuealcala/[email protected] setup . -
Using a single IDE?
Limit the setup to that one. For example, Codex only:
npx @jossuealcala/[email protected] setup . --platforms codex -
Check the project
Confirm AHP+ sees the right repository, branch and commit.
npx ahp project check . npx ahp project status .
The first run should say AHP_READY. Review the files it created before committing them: AHP+ never commits or pushes for you.
In your IDE
Ask your agent, the way you always do.
Open the project in your IDE and ask your agent to use AHP+. For example: “Use AHP+ to check this project and show the commit, the blockers and the next action. Don’t edit files.”
- Terminal
npx ahp project check .- Codex
- “Use
$ahpto check the project and read my inbox” - Claude Code
- “Use AHP+ to run doctor and strict verification”
- Cursor
/ahp message inbox for=cursor- OpenCode
/ahp message send to=claude text="Continue"
Between agents
Let one agent hand the work to another, losing nothing.
You save a checkpoint in one IDE and hand it to another. The receiver checks that the project, branch and commit match before carrying on.
You save a point
A checkpoint with what was done and what comes next.
You hand it over
A handoff, for example from Codex to Cursor, with a summary.
The other one verifies
If everything matches it says READY; if not, it asks to reconcile before touching anything.
It carries on
From the exact point where the last one stopped.
A second opinion. Ask Claude to review what Codex did: it answers once, read-only, and it is recorded.
Shared rooms. Codex and Claude can discuss the project, each from its own chat.
Between computers, encrypted. To send the state to another machine, AHP+ signs and encrypts every delivery.
What it never does
What AHP+ doesn’t do.
It doesn’t commit or push. No merges, no deploys, no publishing. You review and confirm.
It doesn’t grant itself permission. A handoff or a checkpoint authorises nothing: authority comes from you.
It doesn’t read your secrets. It doesn’t open your
.envfiles on purpose. And never paste keys into its records.No dependencies. All it needs is Node and Git. No third-party packages.
When don’t you need it? For a quick question with no repository, or a throwaway script, AHP+ is too much. It helps when a project moves between agents, IDEs, people or computers.
Your say
What is AHP+ missing for the way you work?
Tell us which IDE, which step or which piece of information you’d like AHP+ to cover. Through the form, with no account and private follow-up; or on GitHub, if you prefer a public conversation.
AHP+ inside MADRE
It also works inside MADRE.
MADRE ships AHP+ as an optional module. Turn it on in MODULES and use /ahp status, /ahp check and /ahp context from the room. It is the only module that writes to your project, and it shows you the command and asks for confirmation first.