Website · Docs · Try it in 60 seconds
PAL
A personal agent on your own Mac, in the chats you already use. It knows your people, your calendar and your mail, and it never says yes on your behalf.
- Brief. Type
? and get what needs you across every thread, ranked with a reason. React to act: 👍 do it, 👎 skip, ⏰ snooze.
- Holds. Money, commitments and claims it can't verify are held at the send boundary, with an always-on backstop that works even when the model doesn't.
- Sounds like you. Every draft is scored against messages you really sent; a draft that drifts from your voice is rewritten before it goes out.
- Remembers your people. One memory per person across iMessage and WhatsApp.
- Knows what's true now. A dated picture of your life from mail, calendar and your own notes, with a source on every line, and nudges before you'd remember.
PAL Runtime
The layer PAL runs on, open source, for your own agents. Durable runs, microVM sandboxes, hard budgets and signed receipts, with the data plane in your cloud.
- Agents that finish. Every step is journaled; after a crash the run resumes where it stopped and finished steps are never re-run. External side effects carry an idempotency key.
- Contained. Untrusted code runs in a Firecracker or Kata microVM that you can pause, resume with the same id, or fork into many that share memory.
- Nothing leaks. In broker mode the VM has no network device: a host broker enforces the allowlist, injects credentials the guest never sees, and scans for card numbers and keys.
- Proof. Each run can carry an Ed25519 receipt over the hash of its journal. Edit one row and it fails to verify.
- Spend you set. Budgets refuse a run with HTTP 402 before the model is called. Ask for a capability, not a vendor. A regressed eval returns 422 and blocks the deploy.
Try it in 60 seconds
Docker only, no API key. make demo starts the control plane, the console and a recorded-answer model server (it also needs openssl). Each visitor gets a private workspace, deleted after 30 minutes; nothing calls a model provider.
git clone https://github.com/dshakes/pal.git && cd pal
make demo # prints "demo: http://localhost:3001/kiosk"
make demo-down # stop and delete all demo data
Run your own agent
Needs Docker, Go 1.26+ (CI and the control plane pin go1.26.10; the CLI module needs at least 1.25), and an OpenAI or Anthropic key.
make dev # Postgres, Redis, MinIO, control plane, console
( cd packages/cli && go install ./cmd/pal ) # the CLI (or: brew install dshakes/tap/pal, see below)
pal doctor # health, auth, model provider, a live run
pal onboard # describe an agent in one line, get a running agent
Just the CLI, without a checkout: brew install dshakes/tap/pal (builds pal from the latest tagged release with Go; the release workflow publishes Formula/pal.rb to the dshakes/homebrew-tap tap). If the tap has no pal formula yet, use the go install line above.
Console: http://localhost:3001, sign in as [email protected] / pal (local dev only). Sandboxes, the egress broker and real Firecracker on a Mac: Runtime quickstart.
Use it from code
The SDKs are not on a registry yet; install from a checkout (npm install /path/to/pal/packages/sdk-ts, or pip install /path/to/pal/packages/sdk-python).
import { PalClient } from "@pal/sdk";
const pal = new PalClient({ baseUrl: "http://localhost:8080", apiKey: process.env.PAL_API_KEY });
await pal.agents.create({ name: "triage", description: "Classifies support emails" });
await pal.budgets.upsert("triage", { maxCostUsdPerDay: 25, hardFail: true });
const run = await pal.runs.create({ agent: "triage", input: { email: "..." } }); // 402 if over the daily cap
Pass baseUrl (or set PAL_API_URL): the client's default is not your self-hosted stack. Python and Go SDKs cover the same management surface.
What's in the repo
| |
|---|
| Control plane, workflow engine, schedulers | Go |
| Model router, runtime manager, in-VM harness, gateways | Rust |
| Console, website, TypeScript SDK | TypeScript |
| SDKs | TypeScript, Python, Go |
| PAL's bridges (iMessage, WhatsApp) | TypeScript, on your own Mac |
Learn more
We started as Lantern: a flame you could carry because glass kept it contained. As the product became a companion, we kept the light and dropped the lamp. The flame became a dot, a single moment of attention. पल means "a moment." A PAL is someone who shows up in it.
