What PAL is

PAL is two products on one platform. PAL is a personal agent that answers your iMessage and WhatsApp threads as you, from your Mac, and holds anything that commits you until you say yes. PAL Runtime is the platform PAL runs on: a runtime for AI agents that makes runs durable, budgeted and sandboxed. PAL talks to the same sessions API you use for your own agents.

	PAL	PAL Runtime
What it is	A personal agent for your own iMessage and WhatsApp	A platform for running agents, yours or PAL’s
Who it is for	You, on your Mac	Developers and platform teams
What it does	Drafts, checks and sends replies in your voice; holds the risky ones	Runs agents durably, within budget, with untrusted code in a microVM
Where it runs	Two channel bridges on your Mac, next to a control plane	Your laptop, your cluster, or a managed control plane with the data plane in your VPC
Start with	How PAL works	What PAL Runtime is

What PAL does

Replies in your voice, holds what commits you, briefs you on what needs you, and keeps one memory of your people across iMessage and WhatsApp. Each capability, with its own diagram: How PAL works.

What PAL Runtime does

Build

• PAL RuntimeModel routingAsk for a capability, not a model; failover and idempotency keys built in.
• PAL RuntimeTool calls in a VMA workflow tool step executes inside the microVM via the runtime manager.
• PAL RuntimeRun lifecycleFrom POST /v1/runs to a streamed result, including crash re-drive.
• PAL RuntimeInvocation modesThe ways a run starts and where its code executes.
• PAL RuntimeConnectors18 real connector APIs with credentials encrypted at rest.
• PAL RuntimeWebhook triggersA signed public URL per agent; duplicate deliveries are dropped.
• PAL RuntimeCron schedulesTimezone-aware schedules that fire runs.
• PAL RuntimeSessions and memoryMulti-turn sessions with events over SSE and a shared memory API.
• PAL RuntimeToken streamingTokens flow from provider to browser with no buffering in between.
• PAL RuntimeAgents talking to agentsAttach MCP tools, discover agents over A2A, invoke marketplace agents.
• PAL RuntimeAgentic onboardingDescribe a goal; get a reviewed spec, a tested key and a first streamed run.

Make it reliable

• PAL RuntimeDurable executionA crash mid-run resumes from the journal; finished steps are not re-run.
• PAL RuntimeBudgets and eval gatesOver a hard cap is HTTP 402 before any provider call; a regression is 422.

Make it safe

• PAL RuntimemicroVM isolationUntrusted code runs in its own microVM; untrusted and hostile classes are never downgraded to runc.
• PAL RuntimeEgress brokerNetwork calls pass an allowlist; credentials are added outside the VM.
• PAL RuntimeSandbox lifecyclePause, resume, fork and idle-suspend a running microVM.
• PAL RuntimeTenant isolationRow-level security on every tenant table and a token-gated gRPC boundary.
• PAL RuntimeIdentity and secretsSecrets are references until the harness vends them to an authenticated peer.
• PAL RuntimeSigned receiptsAn Ed25519 signature over the run’s journal hash; tampering breaks it.

Ship and operate

• PAL RuntimeTracingEvery span carries tenant, run and step ids, into the VM and back.
• PAL RuntimeDeploy anywhereControl plane managed or self-hosted; the data plane in your VPC.

What PAL is not

• Not a chat client. PAL acts through the iMessage and WhatsApp accounts on your Mac. It does not replace either app.
• Not fully unattended. PAL sends ordinary replies on its own. It holds anything that commits you, and it does not verify a held claim; it holds it.
• Not a hosted service you sign up for. A fully managed PAL cloud is the intended offering, not a shipped one. Today you run the control plane yourself (Deploy to your cloud).
• Not a model. You bring provider keys. Agents ask for a capability, and PAL maps it to a model from your providers.

Next: for your own messages, How PAL works. To build and run agents, What PAL Runtime is. To look first with no keys, Try it: the demo.