PandaClip

One local-first MCP server + a desktop activity lens. No daemons, no cloud, SQLite everywhere.

@vektorgeist/pandaclip is a single server with four tool families:

All data lives under ~/.panda/<area>/ (override with PANDA_HOME). One registration, one process, 40 tools.

How it works

PandaClip is a single stdio MCP server (TypeScript). Your agent's MCP client spawns it, it exposes the 40 tools, and every tool is a small, deterministic operation on a plain SQLite database — one store per family under ~/.panda/<area>/, WAL mode, no background daemons, no network, no cloud. Nothing happens unless a tool is called:

• clipboard — clip_push appends to history with TTL classes, tag/contains filters, and secret screening (obvious keys and tokens are refused before they are ever stored). Snippets are permanent named clips; channels are named label lanes inside the same local store, so an agent can be pointed at exactly the clip you mean.
• cache — a namespaced TTL cache with canonical hash keys, for expensive lookups an agent shouldn't repeat.
• bamboo — tags, notes, and metadata overlaid on files you already have, in place; nothing is moved or copied.
• garden — a typed knowledge graph: plant/grow/prune nodes, typed edges, BFS traversal, per-node history.

The desktop lens reads those same stores from the other side. The Electron app spawns a read-only watcher child on your system Node (>= 22.5 — it needs node:sqlite's WAL-aware reads), which polls the stores plus any sources you configure, diffs against the last state, and streams change events as NDJSON up to the live feed. The server stays the sole writer; the lens can only look.

The desktop lens

apps/bamboo-clipboard-ui/ is an installable Electron app: a live, read-only feed of what your agents are doing. Out of the box it watches PandaClip's own stores — the clipboard (clips, snippets, channels), cache writes, and garden plantings. Three more sources are blank slots you can point at your own stores in ~/.panda/config.json:

{
  "lens": {
    "notes": { "dir": "~/my-notes", "skipDirs": ["drafts"] },
    "git":   { "dir": "~/code" },
    "kg":    { "db": "~/.local/share/kg/facts.sqlite3",
               "vectors": "~/.local/share/kg/vectors.sqlite3",
               "groupKeys": ["project"],
               "typeLabels": { "journal_entry": "journal written" } }
  }
}

notes feeds markdown edits, git feeds new commits from repos under the dir, and kg understands a knowledge-graph on-disk layout (a triples DB plus a vector store). The optional kg keys shape how vector-store records appear in the feed: groupKeys/tagKeys pick metadata fields for the title and tags, and typeLabels/idPrefixLabels map a record's type field or id prefix to a friendly label — unset, records show generically. Unconfigured slots are simply off. The watcher is strictly read-only. Requires a system Node.js >= 22.5. Run with npm start --workspace pandaclip.

Inspired by jacobian-lens

The lens borrows its shape from Anthropic's jacobian-lens — companion code to the paper "Verbalizable Representations Form a Global Workspace in LMs" — which reads a transformer's internal activations to surface what the model is disposed to say before it says it. PandaClip applies the same move one level up the stack: instead of weights and gradients, it gives you a live, read-only window onto what your agents are doing in the moment — every clip, cache write, planted fact, note edit, and commit surfaces in the feed as it happens. A model lens reads the model's working state; PandaClip reads your agents'.

Better together with magpie-search

PandaClip deliberately ships no search engine. Pair it with magpie-search (PyPI, npx -y magpie-search-mcp) and the two cover each other: magpie finds things (transcripts, history, deep lookup), PandaClip holds things (clipboard, cache, file overlay, knowledge graph).

In practice the pairing looks like this: magpie-search finds — it fans out over transcripts, local files, and the web and brings back the answer. PandaClip keeps — cache_put the expensive lookup so it isn't repeated, clip_push the excerpt that mattered, garden_plant the durable fact with a typed edge to what it relates to. Next session that fact is one garden_search away instead of a fresh search, and the lens shows the whole loop happening live.

Run them side by side as separate MCP servers — there is no code coupling and neither requires the other. Without magpie, everything still works; your agent just does its lookups manually.

Develop

npm install
npm run build
npm test

Use (Claude Code / Desktop)

See examples/mcp-config.json. Point the entry at servers/pandaclip/dist/index.js (or npx @vektorgeist/pandaclip once published).

Design rules

• One server; the four families share only @vektorgeist/panda-core (SQLite helpers) and the ~/.panda/ home.
• Tool names are literal (cache_get, not feed_panda); panda theming stays in docs.
• Cache is an optimization, never a source of truth.

About

PandaClip is built by VektorGeist LLC.

We build local-first tools for people who run their own AI. PandaClip is the working-state toolbox; our agent platform is at vektorgeist.com.

• Website: vektorgeist.com
• Contact: floukie@vektorgeist.com
• Issues & contributions: open an issue or PR on this repository.

License

Licensed under the Apache License 2.0 — see LICENSE. Copyright © 2026 VektorGeist LLC.

"PandaClip" and the panda mark are trademarks of VektorGeist LLC. The code is open under Apache-2.0; the brand and name are reserved.