PUNKthe adaptive runtime

//DOCS What's New

User-facing record of what shipped in Punk, newest first.

What's New

The running record of what shipped in Punk, in plain terms. Newest first. For the engineering-level history, see the repository changelog.

June 2026

Accounts, teams, and billing

  • Organizations and teams. Every account belongs to one or more organizations.
  • Invite teammates by email, switch between organizations from the sidebar, and manage members and roles per organization. Owners can remove members and delete an organization outright, and a removed member loses access immediately.

  • Sign in, your way. Email-and-password accounts for the dashboard sit alongside
  • API keys for your agents, SDK, and tools. Self-serve signup with email verification can be switched on when you want open access; invitations always work.

  • Bring your own provider keys. Add your own OpenAI or Anthropic key per
  • organization through the encrypted vault, instead of relying on a platform key.

  • Plans and usage. A usage dashboard shows month-to-date runs against your plan,
  • spend, savings, and a 30-day trend. Free, Pro, and Enterprise tiers carry sensible limits, enforced fairly even under heavy concurrent load.

Workflows, chat, and agents

  • Chat that gets cheaper. Every reply in the chat view shows the route it took
  • and what it cost; repeat questions are served from cache at no cost. Turn a good conversation into a reusable agent in one click.

  • Scheduled agents. Single-task runners with a prompt, a model, and an optional
  • schedule, built on the same engine that powers workflows.

  • Visual workflow builder. Compose multi-step workflows on a canvas: model steps,
  • tool calls, web fetches and actions, branches, and a new map step that runs a body over every item in a list. Notify Slack or a webhook on completion. Every model step inside a workflow is a real governed, cached run, so workflows get cheaper the more they run.

The runtime, in your editor

  • Punk Runtime Engine for Claude. Punk runs as an MCP server inside Claude Code
  • and the Claude app: run workflows, fetch the semantic web, check savings and approvals, and explain any run, all as native tools. An optional hook records each prompt as an observed run, so your normal work gains Punk's observation and audit without changing how you work.

Trust, safety, and reliability

  • Provider failover. If your primary model provider has an outage, requests
  • transparently fail over to a backup, and the route explanation says so.

  • Tripwires and honeytokens. Plant decoy credentials and detect when an agent
  • reads or leaks them, with an alert or a hard block.

  • Live data-loss protection. Optionally redact sensitive values out of streamed
  • responses as they are sent, not just in stored traces.

  • Memory quarantine. Stop low-trust context (like untrusted web content) from
  • silently triggering high-impact actions such as payments, until a human approves.

  • Approvals everywhere. A policy that requires sign-off opens an approval you can
  • act on from the dashboard or the macOS menu bar, with time-limited exceptions.

Optimization that compounds

  • A glass-box learning view. See exactly why a pattern did or did not become a
  • reusable artifact, with the compiler's own reasoning, replay and shadow results, and a confidence trajectory.

  • Tenant-local preference. As your own results accumulate, Punk tunes how readily
  • it promotes and routes work for you, always within the same safety gates.

  • Opt-in shared learning. Choose to benefit from anonymized, shape-only signal
  • across organizations. It only ever shares hashed pattern shapes and success rates, never your prompts, outputs, or identity, and it never replaces your own proof.

Foundations

  • Persistent, scalable storage. Runs on Postgres in production and a single file
  • locally, with the same behavior either way.

  • Semantic web for agents. Pages are compiled into a compact Semantic Object
  • Model, far smaller than raw HTML, with meaningful change detection. Agents can read and act on the web through it.

  • Marketing site and docs. A public story at the root domain, the dashboard at
  • app, and this documentation on both.


Punk is in active development. This page covers user-facing milestones; smaller fixes and internal changes land continuously.