What's new in peek since launch

Since launch, peek has grown from a read-only session-forensics MCP into a supervised, auditable, portable forensics tool for AI coding agents — still local-first, still “debug what already happened in your real browser.” This is the honest account of what changed, grouped by what it lets your agent do.

peek is pre-1.0 alpha and side-project–maintained (Apache 2.0). Everything below ships today on npm (@peekdev/mcp, @peekdev/cli) and the Chrome Web Store extension. Each section links to a recipe that shows it end to end.

1. Deeper forensic reads

The read path got more useful to an agent reasoning about a failure:

• get_user_action_before_error now returns a pre-assembled causal chain — the user actions, DOM mutations, and network/console errors in the window before an error, merged into one time-ordered timeline with a deterministic narrative, so the agent reasons over “action → DOM change → failed request → error” instead of stitching three queries together.
• query_dom_history gained a selector-free window mode: pass a timestamp (and optional window) to get every DOM change around a moment, no selector needed.
• get_element_detail now returns a curated computed-style bag, the accessible description, and effective aria-hidden/aria-disabled inheritance — enough to reason about why an element looked or behaved the way it did.

See it: See the causal chain that led to an error.

2. Repros that lock the regression

generate_playwright_repro got harder to break and easier to trust. It now emits resilient semantic locators (getByRole / getByTestId / getByPlaceholder / getByText, each uniqueness-checked against the snapshot, with a CSS-selector fallback) instead of brittle deep selectors, and it can seed a console-error-absence assertion — so the generated spec doesn’t just replay the steps, it fails if the original error comes back. The agent can re-run it to verify its own fix actually worked.

See it: Generate a Playwright repro from a real browser session.

3. Visible, revocable supervision

When you let the agent act on the page (the opt-in write path), the oversight is now first-class and on-screen:

• A Level-4 control shield — an overlay that makes it obvious peek is driving, with a one-click Stop and an Esc kill-switch that hard-drops the origin back to read-only.
• A first-run trust experience: the permission levels are named and explained, and a persistent indicator shows when “auto-approve” is active so it can never be on without your knowing.
• An apply-and-re-verify convention: after a step, the agent re-reads the page to confirm the change took (and that no error appeared) before advancing — with a terminal success/failure banner and a non-blocking heads-up when you’re about to click something destructive during a hand-back.

See it: Assisted apply, with supervision.

4. Tamper-evident, and portable

Sessions and the actions taken on them became governable and shareable — without a cloud:

• The action audit log (~/.peek/audit.log) is now a hash chain: each entry is linked to the one before it, so peek audit verify can detect any edit, reorder, or truncation after the fact.
• A *.peekbundle export/import packages a session (with an integrity manifest) so a teammate can replay your repro on their machine — no environment spin-up, no account.
• An explicit retention policy bounds the local store — it never deletes silently; you set the rules.

See it: Reproduce a bug from a teammate’s peek session.

5. Find sessions and verify integrity from the CLI

Two CLI additions make the local store easier to triage and audit:

• peek sessions search lets your agent (or you) filter sessions by text (--q), origin, time window (--since / --until), status, and error type (--errors console|network|any) — narrowing thousands of recorded sessions down to the one that matters, without opening a viewer. The search_sessions MCP tool exposes the same facets to agents directly.
• peek audit verify recomputes the hash chain of the action audit log and exits non-zero on any gap, reorder, edit, or truncation. The verify_audit_log MCP tool exposes the same check so an agent can assert chain integrity as part of a workflow.
• peek audit bundle / peek audit verify (the bundle variant) packages the audit log alongside a session export so a recipient can verify the chain independently.

6. Delegated connectors — supervised background daemons

peek connect is a supervised connector daemon: register a surface (currently Slack), run it once interactively to capture tokens, then peek connect start spawns a detached supervisor that keeps each connector alive and auto-restarts it on exit. Per-connector logs stream to ~/.peek/connect/logs/; the supervisor PID and each connector’s state are visible via peek connect status. Connectors extend peek’s reach beyond the browser — a Slack connector turns channel messages into queryable local sessions the same way the extension turns browser activity into sessions.

7. Claude Code plugin

peek is now a one-line Claude Code plugin install:

/plugin install peekdev@peek

This drops the peek MCP server and the Claude Code skill into place in a single command — no peek init required for MCP configuration. The plugin is hosted in the public repo and installable via the community catalog (/plugin marketplace add Cubenest/rrweb-stack).

The honest part

peek is local-first: peek uploads nothing — what your MCP client does with the data is up to you. That last clause matters: anything your agent reads from a session becomes context for whatever model your MCP client is using. And while peek masks passwords, PII-class values, and fields you mark private at capture time, the underlying rrweb full-page snapshots can still contain non-password free-text values that were on screen — so treat a recording, and a *.peekbundle you share, as you would the page itself. The read tools are non-mutating; the write tools require explicit per-origin consent.

Something out of date or inaccurate? Open a PR on this page — the goal is accuracy, not advocacy.