Replit launches Ramp for Agents for autonomous financial agents, GitHub publishes a multilingual dataset of 40M repositories, HeyGen opens transparent WebM video output

On July 11, agents stepped out of code and into the real world: Replit links its agent to Ramp to incorporate a company and autonomously manage cash, invoices, and expenses, GitHub publishes an open source dataset of more than 40 million repositories on the language distribution of code, and HeyGen chains together two developer announcements — transparent WebM video output and automatic synchronized editing for spoken videos. Pika, for its part, demonstrates advanced video editing capabilities via a model it calls “Gemini Omni,” an attribution not confirmed by Google to date. Five short items round out the picture, from Together AI’s voice call line to Cohere’s infrastructure partnership with NVIDIA and CoreWeave.

Replit launches Ramp for Agents: its agents now handle company finances

July 11 — Replit announced a partnership with Ramp, a financial management and corporate cards platform, enabling its agent to go beyond code generation and handle real financial tasks. Concretely, the Replit agent can now help incorporate a company, submit an application for a Ramp account for that entity, then prepare the business to spend money, pay its bills, and manage its cash — all from a natural language conversation.

The launch fits into Replit’s broader strategy, which has already opened its agent to e-commerce (Shopify integration, June 2026) and discoverability (SEO Agent, June 2026). The product goes through a dedicated page operated by Ramp, “Finance for the Agent Economy,” which gives agents cards, transfers, and bank accounts with built-in controls to govern every dollar spent.

Every company used to start with paperwork. The next one starts with a prompt.” — @Replit on X

GitHub publishes the Multilingual Repositories Dataset

July 11 — GitHub published the GitHub Multilingual Repositories Dataset, a new open source dataset aimed at researchers and developers working on multilingual AI. It covers more than 40 million repositories and more than 80 million classification lines, mapping where the README files, issues, and pull requests written in a language other than English live on the platform.

Measured indicator	Measured value
Repositories covered	40M+
Classification lines	80M+
README repositories in Portuguese (dominant language)	3M+
Dominant language of issue texts	Korean

The figures highlighted by GitHub reveal concrete linguistic trends: Korean dominates issue texts, while Portuguese leads for README files, with more than 3 million repositories affected. This is not a Copilot feature as such, but a research and open data release, in line with GitHub’s positioning as an AI-powered development platform.

🔗 Tweet @github — Multilingual Repositories Dataset

Pika edits videos with a model it calls “Gemini Omni”

July 11 — Pika published a video demonstration showcasing advanced editing features applied to a user-imported video: background change, camera angle change, outfit change, visual effects, and even French voice dubbing — all presented by Pika as made possible “with Gemini Omni”.

“Drop in your video and change the background, change the angle, swap the outfit, add visual effects, even make it speak French — all with Gemini Omni” — @pika_labs on X

Editorial caution: the name “Gemini Omni” does not appear in any official Google or DeepMind announcement to date — the dedicated scan of Gemini announcements found no news on July 11. This is the name used by Pika itself to refer to the underlying model behind this feature; its exact existence and nature have not been independently established. This news should be read as a feature announced by Pika, not as a confirmed product launch by Google.

Moderate engagement for the account (15,000 views, 80 likes), well below Pika’s major posts — a sign of a feature announcement rather than a flagship product launch.

HeyGen beefs up its developer ecosystem: transparent video and automated editing

Two announcements on the same day for HeyGen, one about video output format, the other about editing automation.

Transparent WebM video output on API, CLI, and MCP

July 11 — HeyGen announced a new video output option across its API, CLI, and MCP server: WebM with a transparent alpha channel. By setting the output_format: "webm" parameter on the /v3/videos endpoint, generated avatar videos come back with a truly transparent background (native alpha channel), without relying on a green screen to key out afterward.

This feature targets developers who compose avatar videos into larger productions — editing, overlays, integration into 3D scenes. The associated thread illustrates a concrete example: an avatar composited directly into a Minecraft-style three.js scene.

🔗 Tweet @HeyGen — transparent WebM output

The “Talking-Head-Recut Skill” synchronizes graphics onto a spoken video

July 11 — The sixth episode in a series of daily demos around HyperFrames, HeyGen’s agent-driven video production environment: the “Talking-Head-Recut Skill” overlays presentation graphics (animated titles, counters, cards, quotes) onto a “talking head” video, automatically synchronized with the spoken dialogue. The agent reads the transcript, decides which elements deserve a graphic, then draws each card in the provided visual style.

The demonstration shows the end-to-end workflow from a single prompt: searching for a news topic (OpenAI’s GPT-5.6 announcement), sourced script, avatar generated directly from the HeyGen CLI, visual style pulled live from chatgpt.com, then final render with subtitles and music — no video editor or camera required. The skill is documented as open source on GitHub.

🔗 Tweet @HeyGen — Talking-Head-Recut Skill 🔗 GitHub — talking-head-recut skill

Short news

Together AI launches a voice call line — a minimalist tweet (“Why type when you can call?”) points to docs.together.ai/call, which opens a phone dialer (+1 415-723-8167, ext. 676) connecting the caller with a voice assistant powered by Together AI’s inference platform, with no associated blog post. 🔗 Tweet @togethercompute
Pika puts an experimental MCP server online — experiment.pika.art (“Pika MCP – Make Your Agent Creative”) lets AI agents call Pika’s creative video generation and editing capabilities directly from their agentic environment. 🔗 Tweet @pika_labs
MiniMax and Fireworks AI optimize the M3 kernel on Blackwell — a new “KV-stationary” kernel developed by Fireworks AI for MiniMax’s open M3 model reads each selected block only once, preserving the gains of sparse attention and reaching around 980 TFLOP/s on a B200 GPU. 🔗 Tweet @MiniMax_AI
GPT-Live deployed 100% worldwide — OpenAI’s full-duplex voice feature, launched on July 8, reaches full rollout in ChatGPT; OpenAI temporarily doubles the voice usage limit for the July 11–12 weekend. 🔗 Tweet @athyuttamre
Cohere highlights its partnership with NVIDIA and CoreWeave — institutional communication on X around reliability and infrastructure protection for large enterprises deploying AI, with no technical detail or associated article. 🔗 Tweet @cohere

What this means

Video production is becoming a fully agent-driven pipeline, from editing to distribution. HeyGen removes a classic technical step — green-screen keying — by making the alpha channel native on API, CLI, and MCP, and automates graphics synchronization on spoken videos through its Talking-Head-Recut Skill: a single prompt now suffices to produce a sourced, styled, subtitled video without a camera or editor. Pika pushes the same logic on the post-production side, with a model it calls “Gemini Omni” to change the background, angle, and outfit of an existing video — an attribution that, lacking official Google confirmation, also illustrates a trend of nicknaming technical building blocks to give them a veneer of power, to be taken with due caution.

Agents are moving beyond code and into a company’s administrative and financial work. Ramp for Agents lets Replit’s agent incorporate a company, open a Ramp account, and manage cash and invoices — territory that, until now, had remained strictly human. The same movement appears, at a smaller scale, in Pika MCP, which opens Pika’s creative capabilities to third-party AI agents via an experimental MCP server: the agent no longer just produces, it orchestrates third-party services on the user’s behalf.

The foundations of multilingual AI and large-scale inference continue to open up. GitHub’s Multilingual Repositories Dataset documents, at the scale of 40 million repositories, where Korean, Portuguese, and the other languages of the global codebase actually live — a resource for any researcher working on multilingual models. The KV-stationary kernel developed by Fireworks AI for MiniMax’s M3 illustrates the same infrastructure work on the inference side, preserving the theoretical gains of sparse attention on Blackwell. Cohere’s strengthened partnership with NVIDIA and CoreWeave completes the picture: infrastructure reliability is becoming a selling point in its own right for enterprise AI.

Voice is establishing itself as a full-fledged product channel, rather than just a side feature. GPT-Live reaches full rollout at OpenAI just three days after launch, with doubled limits to encourage trial; Together AI, meanwhile, opens a phone line that directly connects callers to a voice assistant powered by its inference platform. Two different approaches — one integrated into an existing app, the other accessible from a regular phone — for the same bet: the voice interface is becoming a full-fledged entry point to models, not just one option among others.