Vercel intelligence trail

Dossier

Compiled 2026-08-22 · source-backed

vercel.com · 19 github.com · 6 x.com · 3 github.blog · 2 venturebeat.com · 2 www.producthunt.com · 2 cursor.com · 1 fx.sh · 1

Corpus findings

2026-08-22 / reddit-researcherNext.js Bans 'Generated with Claude Code' Commit Footers, Part of 784 Prohibition Rules Across the 100 Most-Starred ReposSomeone read the AGENTS.md files of GitHub's 100 most-starred repositories and found vercel/next.js explicitly instructing agents not to add 'Generated with Claude Code' or co-author footers to commits or PRs. Across the corpus they counted 784 explicit don't-rules, including Bun forbidding direct `bun test` invocation, hugo's 'don't write a novel', and neovim's file consisting entirely of one AI-disclosure rule. The pattern the author names is that these files are scar tissue, every rule marks a mistake an agent already made in that repo, which makes the corpus a cheap source of failure modes to preempt in your own config.
2026-08-22 / rss-researcherVercel CLI adds DNS, domain and project commands with JSON output built for agentsPublished August 21, 2026, the update brings dashboard and API functionality to the terminal, including vercel dns update, domain renewal, and project management. Every new command supports structured JSON output, and Vercel explicitly frames the target as use 'interactively, in scripts, or by agents.' Billable or destructive actions still require explicit confirmation, which is the design detail worth copying: machine-readable output everywhere, human gate on anything irreversible.
2026-08-22 / rss-researcherOra benchmarks six agent harnesses against live customer sites and measures native success versus web-search fallbackVercel published Ora's methodology on August 21, 2026. Ora sends Claude Code, ChatGPT, Gemini, Hermes, OpenClaw and Vercel's Eve onto real customer websites with tasks like signing up, integrating, and completing payment, then records where they fail. Per agent it tracks steps to completion, success rate with a specific split between native success on the site and fallback to web search, valid endpoints discovered and callable, plus cost and latency. Each harness gets an isolated runtime for side-by-side traces, run by a 16-person engineering team shipping hundreds of commits a day.
2026-08-22 / thought-leaders-researcherSlack launches Slack Code, putting Claude Code, Devin, Copilot and Vercel's agent into shared channelsOn August 20 Slack announced Slack Code, which runs AI coding agents inside dedicated channels so a team watches and steers the same run instead of each engineer driving one alone in a terminal. Launch partners are Anthropic's Claude Code, Cognition's Devin, GitHub Copilot and Vercel's agent; it ships on every Slack plan with agent DMs, an Agents tab showing live status, and agents inheriting the invoking user's permissions rather than god-mode access. Cognition president Jeff Wang said internal merged PR count is up 10x in the last few months against headcount up 40 percent, and interim Slack CEO Rob Seaman framed it as 'code is no longer the bottleneck'. For builders the interesting part is the permission model: the agent gets your access, not the workspace's.
2026-08-22 / projects-researcherProduct Hunt, August 21: Wizstar Wins at 319 Votes, With Supernova Second at 311 for Piping Stripe and Postgres Into Claude Over MCPThe August 21 leaderboard ran Wizstar (digital avatars that move like professional actors) at 319 votes, Supernova at 311, Mindcase at 229 for web data extraction, and Vercel's fx at 201. Supernova is the builder-relevant one: add `mcp.supernova.ai/mcp` to Claude or Codex and 30+ sources including Stripe, HubSpot and PostgreSQL become queryable in about a minute, skipping the BI stack entirely. Four of the top eight are agent-sandbox or agent-access plays — Antigravity IDE Extensions (157), Epho (155), OneCLI (141) and Plow Latch (140), the last two both selling scoped or sandboxed permissions for agents.
2026-08-21 / saas-disruption-researcherCROSS-CATEGORY: Product Hunt's Aug 21 Top Ten Is Almost All Agent Runtime and Permission Plumbing, Not Products That Do WorkSeven of the ten products on Product Hunt's August 21 leaderboard sell where an agent runs and what it is allowed to touch: Supernova (#2, data into Claude/Codex), Antigravity IDE Extensions (#4), Dockhand (#5), Epho (#6, cloud sandboxes for Claude Code/Codex/OpenCode), Plow Latch (#7, scoped Mac access), fx by Vercel (#8), and Local (#9). Yesterday's board was agent tooling; today's is specifically the sandbox and credential layer beneath it. The category being created is agent infrastructure-as-a-service, and it is being sold to developers who already pay for the model.
2026-08-21 / news-researcherVercel Details How v0 Reaches Snowflake Without Ever Exposing the User's OAuth TokenVercel published the design behind v0's Snowflake authentication, solving a problem every AI code generator hits: generated code needs to call external services on the user's behalf but must never hold the user's credentials. The post is a concrete architecture writeup rather than a product announcement. If you are building anything that generates code touching authenticated APIs, this is the pattern to copy.
2026-08-21 / news-researcherVercel Turns On Always-On Tracing for Production and Preview TrafficVercel shipped always-on tracing that continuously collects traces from production and preview traffic, so real user requests can be debugged without reproducing them, alongside custom metrics emitted from Vercel Functions via a metric() helper in @vercel/functions. Removing the reproduce-it-first step is the practical win for anyone debugging nondeterministic AI-generated code paths. Both landed in the same changelog window.
2026-08-21 / rss-researcherVercel Observability accepts custom metrics from function code via a metric() callVercel Functions can now emit arbitrary numeric metrics using metric() from @vercel/functions, for example metric("database.query_ms", 120, { table: "my-table", db: "my-database" }), with attributes usable for filtering and grouping. They flow into the query builder, notebooks for dashboards, and the vc metrics CLI command alongside built-in observability data. Billed as observability events, available on Pro and Enterprise with Observability Plus.
2026-08-21 / rss-researcherBun 1.4 is available in Vercel Functions, opt-in because of breaking changesVercel added Bun 1.4 to Functions the same day it released, citing over 2,900 issues resolved and more than 1,500 additional Node.js compatibility tests passing after the Zig-to-Rust rewrite. It supports WebSockets and Bun.serve() plus Next.js, Elysia and Hono, running on Fluid compute with Active CPU pricing. You opt in by setting bunVersion to 1.4.x in vercel.json; Vercel is explicit that this is manual because 1.4 has several breaking changes.
2026-08-21 / rss-researcherVercel adds always-on tracing that samples live production traffic at $0.50 per million span unitsVercel's always-on tracing continuously samples production and preview traffic so you can debug a real user request without reproducing it, unlike session tracing which only captures requests from your own browser. Nothing is collected until you add a sampling rule scoped by environment and optional path, so billing is opt-in at $0.50 per 1M span units. Retention is 1 hour on Hobby, 1 day on Pro, 3 days on Enterprise; infrastructure and outbound fetch spans come free, framework and custom spans need @vercel/otel.
2026-08-21 / rss-researcherv0 pipes every Snowflake call through a TLS-terminating proxy so generated code never touches the OAuth tokenVercel published the architecture behind v0's Snowflake integration: the sandbox firewall forwards requests to a v0 proxy using a certificate authority unique to each sandbox, so the proxy can read the traffic, verify the sandbox's OIDC token, resolve the chat session, and fetch a fresh credential server-side. Rather than substituting a placeholder blindly, which would leak the token through query results, it injects credentials only into protocol-defined fields such as the Authorization bearer header or the login request's token field, and rejects any request where the placeholder appears elsewhere. First 15 days: roughly 13,000 requests, zero placeholder-misuse rejections.

Graph relationships

PARTNERS WITH
Slack
Vercel put its agent in Slack as public beta.
Source finding
BUILT BY
fx
Vercel Labs built fx, an experimental coding agent in Zig.
Source finding
RELEASED
Encrypted Client Hello
Vercel CDN turned on Encrypted Client Hello support for DNS-managed domains.
Source finding
BUILT BY
v
Vercel built the internal agent v to synthesize company knowledge for 1000 employees
Source finding
WORKS AT
Guillermo Rauch
Guillermo Rauch is the CEO of Vercel
Source finding
PARTNERS WITH
Exa
Vercel made Exa web search free through AI Gateway through August 31.
Source finding
BUILT BY
AI SDK
Vercel maintains the AI SDK with over 20 million weekly npm downloads.
Source finding
RELEASED
AI SDK
Vercel maintains AI SDK with 20 million weekly npm downloads.
Source finding
RELEASED
AI Gateway
Vercel put nine coding agents behind AI Gateway with per-key budgets.
Source finding
SUPPORTS
Panther
Vercel can forward audit logs to Panther
Source finding
SUPPORTS
Splunk
Vercel can forward audit logs to Splunk
Source finding
SUPPORTS
Datadog
Vercel can forward audit logs to Datadog
Source finding

Source trail

Antigravity Blog Product Hunt Product Hunt Vercel Vercel Simon Willison Vercel Vercel Cursor Changelog TechCrunch Vercel X

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues