Stop explaining your project to every new chat.
Nibvine keeps one structured brain per project — the goal, the decisions, the reasons behind them — and hands it straight to Claude Code, Cursor, Codex, or anything else that speaks MCP.
Free during open beta. Export everything as JSON whenever you like.
-
Feedback loop for solo founders
-
Positioning
- Who it is not for
-
Decisions
- Postgres over SQLite decided
- No team seats in v1
-
Open questions
- What anchors the price?
-
Positioning
// the agent asks once, at the start of the session { "project_id": 42, "profile": "default" } { "goal": "Weekly feedback loop a solo founder keeps", "decisions": [ "Postgres over SQLite — needs concurrent writes", "No team seats in v1" ], "open_questions": ["What anchors the price?"], "constraints": ["One developer, ships on Fridays"], "tokens": 1840 }
Three moves, and the context stops evaporating
Capture, structure, hand off. The tree does the remembering.
-
Capture without deciding where it goes
Drop a thought in from the web, or send it to the Telegram bot. Nibvine reads it, proposes the branch it belongs under, and waits for you to confirm before anything lands.
-
Let depth carry the meaning
Notes nest inside notes. The goal sits at the top, decisions underneath it, the reasoning under those. Tasks, ideas and comments attach to the branch they belong to, so a branch answers a question completely instead of pointing at four other places.
-
Hand the whole thing over in one call
A context pack assembles the project into a single payload sized for the model's window, summarising branches that are too big and caching the result. Your agent starts every session already knowing the project.
Plug Nibvine into the tools people already code with
One remote MCP endpoint, one scoped token. The connect page writes the config for you; here is what it produces.
- Claude Code — remote MCP setup for agentic coding and review loops
- Cursor — repo-native MCP config plus a one-click install link
- Codex — an exact codex mcp add command, or the TOML snippet
- VS Code — workspace mcp.json for Copilot agent mode
- Any MCP client — generic JSON, a smoke test, and the discovery manifest
{
"mcpServers": {
"nibvine": {
"type": "http",
"url": "https://nibvine.com/mcp",
"headers": {
"Authorization": "Bearer ${NIBVINE_TOKEN}"
}
}
}
}
An agent cannot change your project in one move
Every write from an external tool is staged first. You see exactly what would change, the plan expires if nobody acts on it, and committing applies it in a single transaction that is recorded in the project history.
- Tokens carry explicit scopes and can be pinned to one project.
- Tokens are stored as hashes, expire on a date you choose, and revoke instantly.
- Every change records whether a person, an agent or the system made it.
› preview_update_note { "note_id": 318, "content": "Postgres — needs concurrent writes" } "action_id": "wa_7f21" "status": "preview" "expires": "in 30 minutes" "diff": "content: 1 line replaced" › commit_write_action { "action_id": "wa_7f21" } "status": "committed" "logged": "source=llm · note 318"
Two interfaces, one model underneath
Whatever you connect reads the same tree and obeys the same scopes. Both references are generated from the running server, so they cannot drift.
Put one real project in and see.
Ten minutes of notes is usually enough to notice the difference in the next agent session.