coding-project-init

Coding project init — connect your project to tapps-brain

This scaffold wires a new project to a deployed tapps-brain hub in two dimensions:

1. Design/build-time — your IDE's coding agent (Claude Code, Cursor, VS Code Copilot) talks to tapps-brain via MCP to save/recall memories as you code.
2. Runtime — your shipped app embeds AgentBrain to give its own agent loop persistent memory.

These are independent. You can wire either, both, or neither. The scaffold gives you a minimal working example of each so you can delete what you don't need.

---

Prerequisites

A deployed tapps-brain hub reachable over MCP stdio (via Docker) or HTTP. If you haven't deployed one yet:

# from a tapps-brain checkout
docker compose -f docker/docker-compose.hive.yaml up -d --build
# → tapps-brain-http on :8080, visual dashboard on :8088, Postgres hive on :5433

---

Files in this scaffold

File	Purpose	Edit?
.mcp.json.template	MCP server entry for Claude Code / Cursor — spawns the deployed image.	Rename to .mcp.json, substitute {{PROJECT_ID}}.
brain_init.py	Runtime AgentBrain factory for your app's agent loop.	Import from your app code. Edit to match your agent_id/profile.
.env.example	Every env var the scaffold honors, with defaults.	Copy to .env, fill in.
profile.yaml	Per-project memory profile (layers, decay, ranking).	Edit for your domain; register once against the hub.

---

1. Design-time — Claude Code / Cursor

Copy the MCP config template and substitute your project id:

cp .mcp.json.template .mcp.json
sed -i "s/{{PROJECT_ID}}/$(basename $(pwd))/g" .mcp.json

Open the project in Claude Code / Cursor. On session start, the mcp__tapps-brain__* tools appear. Ask the agent to save or recall — writes land in the deployed hive.

One-time project registration (deployed / multi-tenant)

A deployed hub serves many projects, so it needs to know yours before it will accept writes (ADR-010 / EPIC-069 — tracking the wire-protocol and CLI for this; until shipped, use in-process AgentBrain with project_dir for isolation):

# Once EPIC-069 lands:
tapps-brain project register <your-project-id> --profile ./profile.yaml

---

2. Runtime — embed AgentBrain in your app

brain_init.py exposes a get_brain() factory. Use it from your agent loop:

from brain_init import get_brain

def run_task(user_input: str) -> str:
    with get_brain() as brain:
        context = brain.recall(user_input)
        # ... call your LLM with `context.memory_section` injected into the prompt ...
        brain.learn_from_success(f"Handled: {user_input[:60]}")
        return answer

The factory reads config from env vars (see .env.example). It's intentionally twenty lines — copy-paste it and edit to taste. Nothing here is a framework; it's the canonical example of the three calls you were going to write anyway.

---

What this scaffold deliberately does not do

• No framework adapter (LangChain / LlamaIndex / …). Upstream APIs churn; embed AgentBrain directly in your integration.
• No auto-save hooks. Agents call save() / learn_from_success() explicitly. Automatic background saves hide the contract and make debugging harder.
• No docker-compose for your project. You already have one. Link your hub via env vars.

---

Further reading

• Agent integration guide — full AgentBrain API surface.
• MCP guide — all MCP tools, project identity, auth.
• Profiles — how profile layers, decay, and ranking tune recall.
• ADR-010 — why project_id is on the wire instead of a filesystem profile.yaml.