← All changelog entries
August 15, 2026 · Content

Agent-native browsers get a recipe, seven tool entries, and the first scripts that skip Astro DB entirely

A new orientation on the browser-shaped body for agents — the one that can pull four years of portco correspondence out of Gmail as filed PDFs — plus the seven browsers it surveys, added to the registry. Underneath: the stack rows behind it were written to Turso directly, with no `astro db execute` anywhere, which is the first piece of getting off a layer Astro is deprecating.

Authors
Michael Staton
Augmented with
Claude Code on Opus 5
Tags
#Recipes#Tool-Registry#Agentic-Browsers#Turso#Astro-DB#Technical-Debt

Agent-native browsers get a recipe

Why Care?

The Dojo’s orientation content has argued that you’re never comparing intelligences, only embodiments — Claude Desktop gets a chat window, Claude Code gets your file system. A third body has been arriving all year and nobody wrote it down: the agent that wears your logged-in browser.

That shape matters more in venture than almost anywhere else, because so much of the job lives behind a login with no API you can get a key for. Affinity, PitchBook, Carta, DocSend, fund-admin portals, LP portals, a dozen data rooms with bespoke viewers. “Just use the API” quietly doesn’t apply to most of the surface area. An agent driving the browser you already authenticated into routes around all of it.

The flagship recipe is the one worth stealing: every email to or from a portfolio company, exported as correctly-named PDFs, then handed to Claude Code to build a timeline of what they said they’d deliver. The correspondence history with a portco is a real asset that usually exists in exactly one partner’s inbox and dies when they leave.

What’s New?

  • New Recipe — Orientation: Agent-Native Browsers — the third-body framing, why the shape fits venture, a seven-browser field table, the portco-email-to-PDF recipe with its two-pass inventory-before-export technique, four more venture-shaped tasks, and a security section that doesn’t flinch.
  • Seven tool-registry additions under a new Agentic Browsers category — Comet, Aside, Dia, Opera Neon, Fellou, Sigma, and Tabbit.
  • Claude Desktop vs Claude Code gained three sections: running many sessions at once (one folder each), an honest account of where Desktop does and keeps its work, and a “For the Geekmaxxers” section on agent-aware multiplexers (herdr, muxel, rmux, Claude Squad).
  • Two scripts that talk to Turso directlypnpm add:stack:turso and pnpm sync:stacks:turso, using @libsql/client with no @astrojs/db in the path.
  • A context-v issue doc planning the full retirement of the Astro DB layer.

The security section is the point

It would have been easy to ship a cheerful recipe about letting an agent into your Gmail. Prompt injection against these browsers is not a solved problem and may not become one — Brave demonstrated a working attack against Comet, OpenAI has written that it’s “unlikely to ever be fully solved,” and Gartner has told enterprises to block the category pending upgrades. The reason is structural: the agent’s usefulness is its access.

So the recipe carries five working rules — a dedicated browser profile, never putting banking or LP or cap-table sessions in it, treating read-and-export as a different risk class from send-and-post, watching the first run of any new task, and knowing whether your browser processes page content on-device before pointing it at portco material. The compressed version: you’re the account holder, move data outward into files you own, keep irreversible actions off the table.

The Astro DB thread

Adding Aside and Comet to a stack meant writing Stack rows, and Stack rows live in Turso. Rather than reach for astro db execute — which Astro 7 already broke once, and which now runs through a hand-written shim into @astrojs/db’s internals — the two rows went in over @libsql/client directly.

That turned into the beachhead for a larger cleanup. 41 files still import astro:db: 20 scripts, 16 runtime files, the schema, and the seed. Turso was always the actual database; Astro DB was only ever the access layer, and it’s being deprecated. The new issue doc inventories all of it and proposes a four-phase retirement ordered by risk — scripts first, src/lib/ next, API routes and SSR pages last, teardown after.

The new sync script’s YAML helpers are ported verbatim from the old one so materialized participant markdown stays byte-identical. A port that “cleans up” the formatting would produce a meaningless diff across every participant file.

What’s Next?

  • Fix six wikilinks in tools/mcpmarket.md that resolve to /people/<handle> instead of /tools/<handle>.
  • Decide the n8n stack row: markdown had it carrying current-shaped fields in the aspirational bucket; Turso’s version won the materialization and an added date was dropped.
  • Banners and favicons for the seven new tool entries, per the August 2 self-hosting convention.
  • Phase 1 of the Astro DB retirement — port the 20 scripts, starting with the ones that run on a cadence.