A small lab in the open

Open-source AI with depth.

Local-first, security-first, modular AI tooling — composable parts for people who want to look beneath the surface.

~/recall

# Local memory for Claude Code, Codex, Cursor.
# Three lines, no API key, no cloud.
$ git clone git@github.com:fantazmai/recall.git
$ cd recall
$ make setup

  model      all-MiniLM-L6-v2 (local ONNX, 57 MiB)
  wired      claude-code + codex + cursor
  backfill   ingested 1,247 prior sessions

# In your next Claude Code session — auto-injected:
› remember we're using pnpm not yarn
(decision recorded; will surface in future sessions)

What we've shipped

Open source. Self-hosted. No accounts required. Each piece stands alone — wire them together as you like.

Pre-launch

Fantazm Recall

Memory for agents.

Auto-ingests sessions from Claude Code, Codex, Cursor into one local SQLite store. Recall via MCP + hooks. 98.9% R@5 on LongMemEval, no API key.

fantazm.ai/recall

Coming soon

Fantazm

The agent runtime.

One Go binary, multi-LLM. Encrypted local vault, sandboxed skills, hash-chained audit log, scheduled tasks, native Mac menu-bar app.

shipping after Recall

How we build

A short list of things we won't compromise on.

Local-first by default

Your data, your machine. Nothing leaves the device unless you explicitly opt in. Even our cloud features (when we add them) will be optional, not the default.

Security as scaffolding

Hash-chained audit log, encrypted credential vault, egress proxy, sandboxed skills. Not features — load-bearing structure. If it's not auditable, we don't ship it.

One binary, no Kubernetes

Most agent platforms ship a control plane. We ship a single static binary. If you can run `curl | sh`, you can run our stack.

Multi-LLM, no lock-in

Claude, OpenAI, Ollama — your choice per session. The agent loop is provider-agnostic. Switch in one line. Bring your own keys.

Composable, not monolithic

Each tool we ship is a standalone binary with a wire-protocol API. Use one. Use all. Swap any of them for someone else's. Standard formats (OTLP, OAuth, OIDC) so nothing's a dead end.

Built in the open

Permissive licenses, every commit public, no closed-source core. We'd rather ship a smaller thing that you can fork than a bigger thing you can't see. Issues and PRs are how we ship the roadmap.