The control plane for durable AI agents

Frameworks are libraries that run your agent code in one place. Skopik is a control plane with a clean public REST + SSE boundary, so your product keeps its own API, credentials, and UX, separate from the agent infrastructure.

The core idea is "compute moves; agent identity does not." You dispatch the same agent to cloud Lambda, a Firecracker sandbox, or a customer's Mac without changing your agent code — and identity, brain versions, memories, and execution history persist durably.

Yes. Create an agent once with a versioned brain. Dispatch sessions to cloud, sandbox, or a poll runtime on a desktop, CLI, or your own infrastructure — it stays the same identity.

Usage, audit, and brain snapshots all accumulate in one place no matter where the agent executes.

Set an org plan — Hobby, Pro, or a custom Scale plan. Skopik meters model calls, tool calls, embeddings, and search queries daily.

When cumulative spend hits your plan ceiling, the next dispatch is blocked with a 402 and plan_limit_reached. In-flight sessions complete; there are no silent drops and no surprise overage on Hobby.

Sandbox executions get a writable workspace, support long-running commands modeled as suspendable Operations, and can run image families like Playwright for headless browser work.

Lambda caps at 15 minutes; the Firecracker sandbox is the right home for compilation, file manipulation, and other real process work.

Edit the agent's brain — versioned S3 files like AGENTS.md, notes, routines, and references — then publish a new snapshot. The next run respects it; previous runs stay auditable.

Restore a prior snapshot to revert. Because snapshots are immutable, you can diff behavior and prove a change came from your edit.

AgentMemory rows are durable in DynamoDB and projected to a search index. Runtime Brain tools search them within a session, and agents write new memories via PublishDoc.

Hybrid BM25 + vector retrieval surfaces relevant memories from past sessions, read-through against the source row so stale index hits never leak through.

Yes. Enable email on an agent and it gets a public address like {agent}--{org}@agents.skopik.com.

Inbound mail enters the model fenced as untrusted content and opens a thread. The governed email.send tool replies, auto-sending to known senders without an approval pause.

Metered: LLM tokens (input and output counted separately) and external calls where Skopik pays the provider — tools, embeddings, and search.

Not metered: Lambda invocation, DynamoDB, S3 storage, API requests, and session count. Sessions count toward your monthly quota but carry no per-session charge.

Yes. A lead agent spawns child sessions on other agents (or ephemeral workers from a template), parks at $0 while they work, and synthesizes their results. Questions flow up the spawn tree.

Spawn policy lives on the agent — allowed targets, depth caps, budget slices — so a child's scopes never escalate to the lead's. Every child is a first-class session with its own transcript and audit trail.

No. Skopik is a developer-first control plane. You provision agents via the API and SDKs (TypeScript and CLI today, Python on the way), write brain files, open sessions from code, and integrate external tools.

The platform is orchestrable by AI systems unattended — its golden paths prove an agent can run end-to-end without a human in the loop.

An append-only audit trail logs every event — agent created, task created, dispatch enqueued, email sent — queryable via GET /events?since=...

Usage events meter each run's model cost, tool calls, embeddings, and search by agent, org, and session, so every run is traceable and spend reconciles.

Sign up for the console, create a scoped API key, and install an SDK or the CLI.

Then create an agent, publish a brain snapshot, create a task, and dispatch a session. The same flow works from the TypeScript SDK, the CLI, or the REST API directly.