Vite+ beta, GitHub issue fields, and Manufact — AI Digest for July 2, 2026

Five updates stood out today because they change the work around AI, not just the demos: web toolchains are consolidating, issue trackers are becoming agent-readable, MCP deployment is turning into its own platform layer, and model routers for media generation are getting easier to plug into products.

VoidZero released Vite+ in beta as a single open-source entry point for web projects. The tool manages the runtime and package manager, then wires together Vite 8, Vitest, Rolldown, tsdown, Oxlint, Oxfmt, and a built-in task runner behind commands such as vp dev, vp check, vp test, vp build, vp pack, and vp run. 1

The useful part is not that another CLI exists. It is that Vite+ tries to reduce the number of custom glue decisions in a repo: aligned tool versions, shared config, CI commands that match local commands, and a migration command for existing projects. VoidZero says more than 1,300 public repositories already depend on vite-plus, and the project is MIT-licensed. 1

The caveat: this is still beta. Remote caching, GitLab CI support through setup-vp, and broader framework compatibility are listed on the road to 1.0, so this is a good candidate for test repos and internal templates before you standardize a large organization on it. 1

GitHub made issue fields generally available for organizations on Free, Team, Enterprise, and GitHub Enterprise Cloud with data residency plans; GitHub Enterprise Server support is planned for 3.23. The feature adds typed metadata to issues, including default fields for Priority, Effort, Start date, and Target date. 2

The AI angle is the MCP integration. GitHub says issue fields are accessible through its MCP server, so tools such as Copilot can read and set field values when creating or updating issues. That makes issue triage less dependent on fragile natural-language conventions like "put P1 in the title" or "use this exact template checkbox." 2

GitHub also said more than 40,000 organizations used issue fields during the public preview. If your team assigns issues to coding agents, the practical next step is simple: define the few fields an agent is allowed to edit, then keep the rest human-owned until you trust the workflow. 2

Manufact, formerly associated with the open-source mcp-use SDK, used Launch HN to introduce a cloud product for MCP apps and servers. The pitch is deployment, preview URLs, cross-client testing across ChatGPT and Claude, publishing checks for marketplace submissions, production inspection, and analytics. 3

The product page frames Manufact as a path from first commit to production: connect a GitHub repo, deploy MCP servers, test calls in a browser inspector, check marketplace readiness, and monitor tool-call volume, latency, traces, errors, and sessions. 7

This is worth watching because MCP projects are moving past "can my agent call a tool?" The harder problems are now auth, review requirements, client differences, observability, and regression testing. Those are normal platform-engineering problems, just pointed at AI clients instead of browsers.

Microsoft announced Microsoft Frontier Company, an enterprise AI engineering business backed by a $2.5 billion investment and 6,000 industry and engineering experts embedded with customers. The stated job is to co-design, deploy, and continuously improve AI systems tied to measurable business outcomes. 4

For builders outside large enterprises, the signal is the packaging. Microsoft is not only selling models. It is selling model choice, proprietary-data integration, governance, security, FinOps-style cost tracking, and continuous improvement as one implementation loop. 4

That is useful framing even if you are a small team. A production AI feature needs more than a prompt and a model endpoint: ownership of data flow, fallback behavior, cost limits, auditability, and a way to improve after launch.

Pollo AI announced Pollo API, a unified developer API for more than 300 video and image models. The release says it supports model families such as Veo, Seedance, Kling AI, Sora, GPT Image, Nano Banana, Runway, and Hailuo, with generation, editing, enhancement, effects, API keys, task polling, logs, webhooks, and developer docs. 5

The appeal is the same one developers already know from LLM routers: connect once, switch models per job, and avoid rewriting provider-specific integration code every time a new image or video model becomes good enough. The caution is also familiar. This is a PR announcement, so treat provider coverage, stability, and pricing as claims to test against your own workload before building around it.

Feyn published Pulpie, a family of open-source models for extracting main content from HTML pages. The smallest model, pulpie-orange-small, has 210 million parameters and is reported at 0.862 ROUGE-5 F1 on WebMainBench, close to Dripper's reported 0.864 score despite Dripper being 600 million parameters. 6

The speed claim is more interesting for builders: on an NVIDIA L4, Feyn reports 13.7 pages per second for pulpie-orange-small versus 0.68 pages per second for Dripper, translating to a claimed $7,900 to clean 1 billion pages versus $159,000 with Dripper under the article's cost assumptions. 6

Use the numbers carefully because they are release-party benchmarks. Still, the problem is real. RAG systems and crawlers often fail before retrieval because the page text is full of nav bars, footers, ads, and broken code blocks. A small extractor that preserves main content cheaply can improve a pipeline without changing the downstream LLM.

If you only have an hour, test Vite+ on a side repo, then check whether your GitHub issue metadata is structured enough for agents to use safely. If your team is building MCP servers, Manufact is a sign that deployment and cross-client testing should be treated as part of the product, not as afterthoughts. For data-heavy AI work, Pulpie is the most directly testable open-source item in the set: run it against pages your current scraper handles badly and compare the Markdown before touching your retrieval stack.