VektorGeist Documentation
What each VektorGeist product is, how to get started, and where to find the source. Everything here is local-first — it runs on your machine.
Aviary — Starter $19.99 · Pro $39.99 · Complete $59.99 (one-time)
Local-first security for the AI agents you run
You run AI agents — Claude Code, Cursor, an MCP server, your own Python. They read your files, send mail, run commands, and reach the network on their own. Aviary is the security layer that watches them: agent governance for the individual, the kind of supervision enterprises spend six figures on, running entirely on your own machine. It's a set of small engines ("birds"), each guarding one thing, all offline — no server, no account, nothing leaves your box.
Highlights
- Canary — catches prompt-injection by measuring behavioral deviation, not keyword matching, so it flags novel attacks
- Stork — canonicalizes what your agent sends and strips PII, secrets, hostnames, and IPs before anything leaves the machine
- Crow — real-time alerts the moment something looks wrong, pushed straight to you
- Runs entirely offline — no server, no account, no telemetry
- One-time purchase; higher tiers (Pro, Complete) unlock more birds — see the marketplace for each tier
Get started
Pick a tier on the VektorGeist marketplace, download the package, and install it locally. It runs offline from the first launch.
magpie-search — Free · open source (Apache-2.0)
Local-first federated search your AI agent can call
The search engine an AI agent or LLM reaches for when it needs to find something true to reason over. magpie-search takes one question and fans it across five sources at once — your AI's entire conversation history, the files on the machine, a structured knowledge graph, a vector store, and the live web — then fuses everything into a single ranked answer where every result carries a trust tier (fact > reference > lead > stale). It runs entirely on your machine: no server, no account, no telemetry, and your transcripts never leave. It plugs into whatever AI you run over MCP.
Highlights
- Five sources, one call — transcripts, files, knowledge graph, vector store, live web
- Five search modes — grep (exact/regex), lexical (BM25), semantic (embeddings), hybrid (RRF), and cross-encoder rerank
- Trust-weighted ranking with cross-source dedup and corroboration counts
- Local SQLite index (FTS5 + sqlite-vec, 384-dim on-device embeddings); ~30 classes of secrets scrubbed at ingest
- Listed in the official MCP Registry; MCP tools: search, recent, session, list_sessions, stats, reindex
Get started
Add it to your agent over MCP with `npx -y magpie-search-mcp` (or `pip install magpie-search`), then point it at your sources. Full setup is in the GitHub README.
PandaClip — Free · open source (Apache-2.0)
A local-first MCP toolbox your agents share — plus a live lens on what they're doing
Your agents produce working state constantly — the excerpt that mattered, the expensive lookup, the fact that took a whole session to learn. PandaClip keeps all of it on your machine: a clipboard with history, TTL classes and named snippets; a namespaced cache so lookups aren't repeated; a file-organizer overlay that tags files without moving them; and a knowledge graph where facts grow, connect, and get pruned. One server, one registration, 40 tools, SQLite everywhere — no daemons, no cloud, no network. The desktop lens is the part you watch: an installable app that streams your agents' activity live, read-only by construction.
Highlights
- Clipboard with history, TTL classes, and named snippets
- Namespaced cache so expensive lookups aren't repeated
- File-organizer overlay — tags and annotates files without moving them
- Knowledge graph — facts grow, connect, and get pruned
- 40 tools in one MCP server, SQLite everywhere, fully offline
- Desktop lens (Windows, Linux, macOS) streams agent activity live — read-only
Get started
Run the MCP server with `npx -y @vektorgeist/pandaclip`, then grab the desktop lens installer for your OS from GitHub Releases. Pairs with magpie-search — magpie finds, PandaClip keeps (neither requires the other).
Fledge — Free · open source (MIT)
Raise your own AI agent on Claude Code
Fledge installs a clean, proven agent architecture — ICM file-memory plus operating discipline — into your home folder, with a visual walkthrough that teaches how it actually works instead of hiding it behind a "magic brain." It's the same operator-and-agent workflow VektorGeist is built around, set up for you in one command.
Highlights
- Clean, proven agent architecture — ICM file-memory + operating discipline
- A visual walkthrough that teaches how it works, nothing hidden
- One command on macOS, Linux, or Windows
- Free and fully open source (MIT)
Get started
Grab it on GitHub and run the one-command installer for your OS. Everything is open source, so you can read exactly what it sets up.
PR Triage — Free (3/day) · Pro $19/mo · Team $49/mo
AI pull-request triage for maintainers
PR Triage automatically scores pull requests by complexity, risk, and urgency so maintainers focus review time where it matters most. Connect GitHub and it evaluates incoming PRs for you.
Highlights
- Automatic scoring by complexity, risk, and urgency
- GitHub integration
- Free: 3 evaluations/day with basic risk scoring
- Pro ($19/mo): unlimited evaluations, priority analysis, custom scoring rules, Slack notifications
- Team ($49/mo): org-wide analytics, workload balancing, priority support
Get started
Open the web app, connect your GitHub, and start on the free tier (3 evaluations/day). Upgrade in-app when you need more.