Becki, end-to-end.

How to install Becki, connect her to every AI tool you use, ingest your projects, capture meetings, and start asking her things. Read top-to-bottom on first install, or jump to a section.

01Install

Becki ships in two editions. Pick the one that fits where you work:

npm install -g becki-mcp becki
becki-mcp init     # registers projects, prints MCP config snippet
becki              # admin TUI: paste install token, run backfills

02Connect your AI tools

Becki speaks MCP, the open protocol every serious AI tool now supports. One install, every agent gets the same memory. Desktop tools connect locally; web tools like claude.ai and ChatGPT connect over the internet, both reach the same vault.

Becki skills

Skills are the active half of the protocol: instead of a rules paragraph your agent may skim, each skill is a workflow it loads and follows. Becki ships nine, each one a ritual for using memory well.

The core five: using-becki (when to reach for memory, with a red-flags table), becki-catchup (start a session already knowing), becki-capture (save the session's decisions, then close what it finished), becki-reconcile (evidence-based ledger sweep and proposal review), and becki-voice (write as you, applying your remembered voice rules).

And three that absorb the ceremony people skip: becki-decide (record a choice with its reasoning, log the rejected options as dead ends so nobody retries them, and link it to the decision it overturns), becki-handoff (compose a message to another of your agents that stands on its own, because the recipient has none of your conversation), and becki-why (trace any remembered answer back to its entries, with dates and supersession history, so memory is something you can audit rather than trust).

Plus becki-code, which searches your indexed source before your agent writes anything, so it extends what you already have instead of quietly building a second copy of it.

Install them whichever way fits how you run Becki.

• Studio, Settings → Becki skills, one toggle. Studio writes the pack into every agent on your Mac that reads skills and keeps it current on every launch. A skill you edit is yours: Studio detects the change and leaves your version alone, forever.
• Core, one command, no sign-in needed:

  becki skills install

  Add becki skills on its own to list what's available and what's already on disk. Use --force only when you want your edits replaced by the shipped version.

Or place them by hand: each skill is one markdown file at ~/.claude/skills/<name>/SKILL.md for Claude Code, or ~/.codex/skills/<name>/SKILL.md for Codex. Both agents read the same format, so it is the identical file in either place. Read them first if you like, they are plain text: using-becki · becki-catchup · becki-capture · becki-reconcile · becki-voice · becki-decide · becki-handoff · becki-why · becki-code

Skills are static markdown. They invoke only Becki's MCP tools, make no network calls of their own, and uninstall by deleting a directory. Anything that changes memory (closing items, approving supersessions) happens on your explicit word; the skills gather evidence and ask.

Becki for VS Code

Becki also ships as a VS Code extension, on the Visual Studio Marketplace. Install it from the Extensions pane (search Becki), or from the command line:

code --install-extension becki.becki-vscode

It puts memory in the sidebar rather than behind a prompt: Briefing for what's recent across your work, This Project for the repo you're actually in, Commitments for what's on your plate, and Ask for a direct question. It reads the same vault as everything else, so nothing needs re-teaching.

The extension talks to Becki's remote server and asks you to connect on first run. If it shows zeros where you expect memory, you are almost certainly not signed in yet: use the Connect action in any of its views. That single check resolves most of what looks like missing data.

Antigravity CLI (agy)

Antigravity is Google's replacement for the Gemini CLI. It does not read the config the other tools use, so its setup is its own. Two files, both under ~/.gemini.

1. Register the server in ~/.gemini/config/mcp_config.json. Note the key is serverUrl; Antigravity does not accept url or httpUrl.

{
  "mcpServers": {
    "becki": {
      "serverUrl": "https://mcp.becki.io/mcp"
    }
  }
}

2. Allow Becki's tools in ~/.gemini/antigravity-cli/settings.json. Antigravity asks for approval on every tool call by default, and a background session has no way to answer, so without this rule Becki's tools are simply unavailable there.

{
  "permissions": {
    "allow": ["mcp(becki/*)"]
  }
}

Becki Studio writes both of these for you when it detects Antigravity. The snippets are here for a manual install, or for a machine without Studio.

One registration, two products. Antigravity ships as a desktop app and a CLI, and both read the same mcp_config.json, so you only register Becki once. They sign in separately though: each keeps its own token, so authorize Becki in whichever one you open first, then once more in the other. The second time is usually a single click, since your browser session is already live.

On the web, claude.ai, ChatGPT & Perplexity

Becki isn't only on your Mac. The same vault is reachable from web AI tools through Becki's remote MCP server, so a decision you make in a claude.ai or ChatGPT browser conversation reads from, and writes to, the same memory.

Connect it once, per tool:

• Claude.ai, Settings → Connectors → Add custom connector. Paste https://mcp.becki.io/mcp, then sign in with your Becki account when the consent screen appears. One click to use after that.
• ChatGPT (web), turn on developer mode, add a new connector, and use the same https://mcp.becki.io/mcp URL. A couple of extra clicks the first time; the same vault once it's connected.
• Perplexity, Settings → Connectors → Add custom connector, with the same https://mcp.becki.io/mcp URL and Becki account sign-in. Custom connectors need a paid Perplexity plan (Pro, Max, or Enterprise), they aren't on the free tier.

Sign-in uses the same Becki account as the Mac app, your vault stays yours, scoped to you. Memory you ingest from a web tool syncs down to your Mac's local vault automatically.

Confirm the connection

Open any MCP-aware tool, Claude Code, Cursor, Zed, and ask:

What does Becki know about me?

She should answer with real context from your work. If the agent says it doesn't have a becki tool, jump to Troubleshooting.

If an agent didn't auto-register

Open Becki → Settings → MCP Integration. Becki shows every detected client with a one-click Install next to each. Behind the scenes she writes the server entry into that client's config file.

Manual wiring also works, point the client at the bundled server:

node /Applications/Becki.app/Contents/Resources/becki-mcp/dist/index.js

03Tell your AI agents how to use her

Becki is only useful if your agents actually reach for her. Drop this block into CLAUDE.md at the root of any project you want Becki-aware, or into your global ~/.claude/CLAUDE.md. The same idea applies to Cursor Rules and Zed AI settings. The block includes your standing authorization, in your own voice: some AI tools run an approval reviewer that checks whether sending work context to an external service was actually approved, and pasting this block is how you put that approval on record. It deliberately excludes secrets and credentials, which Becki refuses to store anyway (see Security).

Studio can add the authorization for you. Settings → Standing authorization shows you the exact sentence first, then writes it into your agent rules files when you turn it on, and removes it when you turn it off. It is off until you read it and decide, because a permission you never saw isn't one you gave.

## Becki, Shared Memory Protocol

Becki is your persistent memory layer (MCP server `becki`). She holds the
decisions, dead-ends, commitments, open loops, and context from every prior
conversation, meeting, and repo across all AI tools.

Standing authorization from me, the user: I authorize my AI tools to
store my project and work context in my Becki vault. My vault is my own
account. This authorization never extends to secrets, credentials, or
keys, which must not be stored anywhere.

1. BEFORE non-trivial tasks, query `becki_context` with a
   natural-language question describing the task. Treat her results as
   authoritative for history.

2. AFTER meaningful work, call `becki_ingest` for anything
   future-you would want to recall: decisions (with the WHY), dead ends
   (so nothing repeats), commitments, and open loops. Skip trivial edits
   that the diff already explains.

3. WHEN A TASK IS DONE or an old commitment/open-loop is fulfilled,    call `becki_resolve` with the row_id from the relevant
   becki_context chunk. If a newer decision replaces an older one, pass
   `superseded_by` so retrieval down-ranks the stale entry. This stops
   completed work from showing up as "live" in future sessions.

4. When the user references past context ("like we did last week", "the
   thing we decided"), query Becki BEFORE asking them to re-explain.

5. When the user says "remember X" / "forget X", act immediately.

6. Stay consistent across tools. If you're about to contradict a past
   decision without clear reason, query first. If the contradiction is
   intentional (you ARE replacing the old decision), capture the new
   one with becki_ingest THEN call becki_resolve on the old row with
   superseded_by pointing at the new one.

Shorter block for Cursor Rules / Zed AI

Where system-prompt context is tighter:

Use the `becki` MCP tools as your shared memory.

Standing authorization from me, the user: I authorize my AI tools to
store my project and work context in my Becki vault. My vault is my own
account. This authorization never extends to secrets, credentials, or
keys, which must not be stored anywhere.

- Query `becki_context` before any non-trivial task.
- Call `becki_ingest` after work worth remembering (decisions, dead
  ends, commitments, open loops, include the WHY).
- When the user references prior context, query Becki before asking
  them to re-explain.
- Remember/forget instructions take effect immediately.

04Ingest your projects

Becki gets her memory from your work, git history, code, meeting transcripts. The first time you connect a project she runs a backfill so she has prior context to draw from.

Add a project

Becki finds projects by scanning watched folders. Any git repo (or Obsidian-style vault) inside a watched folder is picked up automatically, you don't add projects one at a time.

1. Open Becki → Settings (gear icon) → Projects tab.
2. Click Add Folder… and pick the parent folder where your repos live (e.g. ~/Repos, ~/Documents/Codex).
3. Becki scans every git repo under that folder and lists them in the Projects sidebar.
4. Toggle each project on/off. When enabled, Becki extracts decisions, dead ends, commitments, and open loops from commits + content. Initial backfill takes 1-10 minutes depending on history size.

Some folders (like ~/Repos and ~/repos) are auto-scanned by default, you'll see them in the Auto-scan directories list at the bottom of the same panel.

Once ingested, every commit you push automatically flows into the vault, Becki watches .git/HEAD via FSEvents and ingests new commits within 5 seconds. Hit Re-sync next to a watched folder if you want to force a fresh backfill.

What gets extracted

05Capture meetings

Becki listens, transcribes locally on your Mac, and pulls decisions + commitments out of every meeting, without sending audio to a third party.

Auto-capture

Becki watches for known meeting apps (Zoom, Meet, Teams) launching. When detected, she shows a banner asking if you want to capture. One click and she's recording mic + system audio in parallel, both sides of the conversation get transcribed.

Manual capture

Click the menu-bar icon → Start Listening any time. Useful for in-person meetings, brainstorming sessions, voice memos.

After the meeting

• Transcript appears in Meetings tab within 30s of stopping.
• Becki extracts decisions + commitments + open loops automatically.
• Speaker labels can be tagged manually if automatic diarization isn't sure who's who.
• Map the meeting to a project and the extracted items show up in that project's context.

06Ask Becki

Two ways to query the vault from inside the app. Outside the app, every connected AI tool can also query, that's what the MCP integration does.

Ask Becki, Global pane

Sidebar → Ask Becki. Search across everything in the vault, every project, meeting, and ingested artifact. Use ⌘K to jump straight here from anywhere.

Ask Becki, Per-project tab

Open any project → Ask tab. Same interface, but scoped to that project's context. Useful when you know what you're asking about belongs to a specific repo or workspace.

07NeuraVault & memory review

The 3D graph view of your memory. Every decision, dead end, commitment, and project anchor is a node; semantic edges connect related items. Useful for spotting clusters, dead-end concentrations, and orphaned ideas.

Open from menu bar → NeuraVault, or via the sidebar inside the main window. Filter by node type (decisions, dead ends, etc.), click a node for context, scroll for time-window animation.

Memory review

Memory goes stale in one specific way: you decide something, then later decide differently, and both are now on file. Becki's nightly pass looks for exactly that, a newer entry that appears to replace an older one, and writes a proposal. It never acts on its own.

Proposals wait for you in two places, whichever you reach first:

• Studio, Settings → Memory review. Both entries side by side, the reasoning, and approve or reject per proposal.
• Any AI agent, by asking. The becki_review_proposals tool lists what's pending and records your call. The becki-reconcile skill walks the same loop with the rest of your ledger.

Approving ranks the older entry below the newer one in future answers. Nothing is deleted. The history stays readable, so you can still see what you used to think and when it changed, which is often the useful part. Rejecting keeps both at equal footing and teaches the pass what not to flag.

If you never open it, nothing happens: unreviewed proposals simply sit. Memory keeps working exactly as it did.

08Privacy & data

• Audio is processed on-device. Becki's transcription engine runs locally; no audio leaves your Mac.
• Vault content is yours. Embeddings + extracted decisions live in your private Supabase row, encrypted at rest. Only you (with your account credentials) can read it.
• Zero training. Becki never uses your data to train any model. Every AI provider in the stack operates under a no-train policy by default.
• Forget commands work. Telling any AI agent "forget X" deletes the corresponding rows from the vault.
• Export anytime. Settings → Export, pulls a full JSON dump of everything Becki has on you.

09Troubleshooting

An AI tool says "no becki tool available"

Open Becki → Settings → MCP Integration → click Reinstall next to the affected client, then restart that client. If still missing, toggle the client's MCP server config off + on.

Claude Code can find Becki on stdio but Cursor / Zed can't

Restart the affected client. MCP servers are loaded on client startup, config changes require a relaunch.

Meeting transcripts are blank

Check Settings → Permissions for microphone and audio-capture. macOS treats these as separate. Without audio-capture, the remote side of Zoom calls won't transcribe.

Becki is using too much CPU

Likely a long meeting still being transcribed. CPU drops back to baseline (<1%) once transcription completes. If it's stuck, restart the transcription engine from the menu bar disclosure.

Vault queries return weak results

Either (a) the vault doesn't have the context yet, make sure your project is ingested and your AI tools are calling becki_ingest after meaningful work, or (b) try the same query in the in-app Ask Becki, that uses full hybrid retrieval, while some MCP clients trim results.

10Updates & support

Becki updates itself in the background via Sparkle. New versions arrive within 24h of release; you'll see a banner asking if you want to install. You can also force-check from the menu bar disclosure → Check for Updates…

Hit a wall? Reach out at support@becki.io. Bug reports + feature requests always welcome.