Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@become-human where did we leave off on the login screen project?"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Your AI agent, working with you. Not just for you.
Install · What changes · Your files · Make it yours
AI agents are fast, but they forget you every session, decide things you wanted to decide, and leave work you cannot follow. BecomeHuman fixes that in any agent that speaks MCP: Claude Code, Codex, Cursor, Gemini CLI and others.
Related MCP server: memory-bank-mcp
Run it once per computer. From then on your agent starts it by itself, every session.
Claude Code, as a plugin (inside a session):
/plugin marketplace add 201Tech/become-human
/plugin install become-human@become-humanOr as an MCP server, which also lets you pass settings:
claude mcp add become-human --scope user -- npx -y become-humancodex mcp add become-human -- npx -y become-human
gemini mcp add -s user become-human npx -y become-humanFor Cursor, Windsurf, Claude Desktop and any other agent, add this to its MCP config:
{
"mcpServers": {
"become-human": {
"command": "npx",
"args": ["-y", "become-human"]
}
}
}You only need Node.js 20 or newer.
You: let's continue Agent: Last time we left off at #4 "add the login screen". There is also an open question: email or phone? Which do you prefer? You: email, and stop using semicolons Agent: Done. Learned: you prefer code without semicolons.
It does the work and keeps you in the loop: what it will do, why, and what changed.
You lead: it asks before any choice that shapes the work, and your word is final.
It makes your life easier: it spots what slows you down and offers to fix it, only with your consent.
It teaches: plain explanations, adapted to how much you already know.
It learns how you work: your preferences and corrections, so it never asks twice.
It remembers each project: code, writing, research or planning: ideas, decisions, questions, next steps.
It is a real partner: matches your tone, notices when you are tired, admits mistakes, disagrees with respect.
Read the full rules in rules/human-first.md.
Everything stays on your computer, in a folder you can open: ~/BecomeHuman.
File | What is inside |
| What the agent learned about you, grouped by topic |
| Your own rules for every agent |
| One checklist per project: ideas, decisions, questions, next steps |
They are plain text. Read them, fix them, add notes: the agent follows what is there and never erases what you wrote. Not sure where they are? Ask your agent: "where is my data?"
Your rules: write them in ~/BecomeHuman/my-rules.md, one per line. They win over the defaults.
Your language: set BECOME_HUMAN_LANG=pt to get your files in Portuguese (en is the default).
Another folder: set BECOME_HUMAN_HOME to keep the files elsewhere, like a synced folder.
Profile size: the agent keeps at most 50 facts and merges similar ones. Change it with BECOME_HUMAN_PROFILE_LIMIT.
To pass a setting to Claude Code, add -e when installing:
claude mcp add become-human --scope user -e BECOME_HUMAN_LANG=pt -- npx -y become-humanSee CONTRIBUTING.md. License: MIT.
Remove a learned fact. Pass words from it; it must match exactly one fact.
Inputs: fact (required).
Full idea map of a project, plus what is still open.
Inputs: project (required).
Everything learned about how your human works.
Inputs: no named parameters.
Reread the rules for working with your human.
Inputs: no named parameters.
Record an idea, decision, question or next step. parentId links it to a node.
Inputs: kind (required), text (required), project (required), parentId.
Mark a question answered or a next step done.
Inputs: id (required), project (required).
Save one preference, habit or correction about your human. Tell them in one line.
Inputs: fact (required), category (required).
Where the human's BecomeHuman files are and what each holds.
Inputs: no named parameters.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 8 tools
The two domains (human memory vs. project idea map) are clearly separated, and remember/forget/map_add/map_resolve each target a distinct operation. The only mild overlap is between get_rules, get_profile and where_is_my_data, which all return loosely related read-only context and could be confused by an agent looking for stored knowledge.
There is a readable grouping (get_* for reads, map_* for map operations), but conventions are mixed: bare verbs (remember, forget), verb_noun (get_profile, get_map), noun_verb (map_add, map_resolve), and a full sentence (where_is_my_data). An agent can follow it, but it is not a predictable pattern.
Eight tools is well within the sweet spot and each one maps to a concrete operation: read rules/profile/map, write/remove a fact, add/resolve a map node, and locate data. Nothing feels redundant or padded.
Fact memory covers create, read and delete, but there is no update/edit of an existing fact, and the map surface lacks any node removal, rename or edit operation — only add and resolve. An agent that mis-records a node or needs to correct a preference will hit a dead end.
An MCP memory server. One memory your agents share — across models, devices and apps.
One MCP server for the services your AI needs: memory, files, Microsoft 365, kanban and more.
Hosted MCP server and Agent Skills for specs, engineering threads, and project memory.
Cloud-hosted MCP server for durable AI memory
