Daily AI Launch Radar: July 23, 2026

TL;DR: This daily Radar summarizes source-checked AI launch candidates for Kingy AI readers, with pricing notes, use cases, and human-review caveats where details are still emerging.

Today’s list is unusually focused on the infrastructure around AI agents. OpenAI Presence packages deployment, policy, evaluation, and escalation into a managed enterprise product. Kastra checks an agent’s proposed actions before they run, while box by ASCII supplies persistent virtual machines where coding agents can work. Humalike approaches the problem from the social side, and Buzzy applies agentic control to multi-scene video production.

These products solve different problems, so the scores below are editorial priorities rather than a head-to-head ranking. The practical questions also differ: buyers should examine operational ownership for Presence, policy coverage for Kastra, isolation and workload cost for box, consent and memory for Humalike, and sequence consistency plus rights terms for Buzzy.

Launch Snapshot

The snapshot below compares the strongest source-checked launches by Kingy AI score. It is a research-priority visual, not a benchmark chart or hands-on test result.

Strongest Launches

Humalike x Hermes

Humalike launched a one-command Hermes Agent plugin that adds turn-taking, group-tone adaptation, and social memory.

Sources checked: launch source, docs, GitHub repo.

What it does: The open plugin connects Hermes to Humalike’s behavioral APIs. It can decide when the agent should join a conversation, rewrite replies to match a group’s style, build a persona, track how messages may land, and learn recurring social conventions. The repository documents setup for Slack, Telegram, and WhatsApp and warns that group permissions need deliberate configuration.

How to evaluate it: Start in a disclosed test group with no sensitive history. Watch for interruptions, unwanted replies, overfamiliar language, and memory that persists longer than participants expect. The most useful result is an agent that knows when to stay silent.

Why it matters: Agent quality in group settings depends on timing, restraint, and social context as well as answer quality. This launch makes those behaviors a separate infrastructure layer that builders can integrate.

Who should care: AI App Builders, AI Engineers, Developers, Creators

Pricing: Humalike says each new account receives $20 in free credits with no card required. A complete public numeric price schedule beyond those starter credits was not verified, so builders should check the current offer before budgeting a production deployment.

What remains unproven: Humalike says SOC 2 Type II, ISO 27001, and GDPR work is in progress rather than complete. Social memory can contain sensitive context, so builders should examine consent, retention, moderation, and failure behavior before use in private or vulnerable communities.

box by ASCII

ASCII launched box, a CLI and API service for starting persistent Ubuntu virtual machines for coding agents and software factories.

Sources checked: launch source, docs.

What it does: A box is a persistent Ubuntu machine with SSH and SCP access, Docker, a dedicated IPv4 address, a virtual desktop, snapshots, and disk-level forking. The official docs provide CLI, HTTP API, Python SDK, and TypeScript SDK paths. Stopping a box snapshots its state and pauses billing, while a fork creates another machine from that saved state.

How to evaluate it: Create a disposable machine without production credentials, install a small project, stop it, resume it, and fork it. Measure startup time, snapshot reliability, outbound-network controls, and the cost of the actual agent workload rather than comparing only the advertised hourly rate.

Why it matters: Long-running and parallel coding agents need isolated computers with persistent state. box packages that infrastructure into a self-serve service aimed specifically at agent workloads.

Who should care: AI Platform Teams, AI Engineers, Developers, Founders

Pricing: The official site lists per-second billing with a $20 monthly account minimum. That amount includes 2,000,000 VM-seconds, about 555 hours, for dedicated 4-vCPU and 8-GB VM time shared across one or more boxes; ASCII’s comparison page carries the current account terms.

What remains unproven: The official site currently lists EU regions only and an account minimum. Builders should test network isolation, secret handling, snapshot retention, concurrency limits, data residency, and total workload cost before production use.

OpenAI Presence

OpenAI launched Presence, a managed product for deploying governed voice and chat agents into specific enterprise workflows.

Sources checked: OpenAI’s official launch and availability announcement.

What it does: Each deployment begins with a defined job and limited access to the knowledge and systems required for that job. Companies set policies, approved actions, and escalation conditions. Simulations and graders test common requests and risky edge cases before launch; after launch, production sessions and escalations feed a Codex-assisted process for proposing updates that teams can test and approve.

How to evaluate it: Presence is a services-led enterprise purchase, not an API feature to switch on. A useful evaluation should name the workflow owner, the actions the agent may take, the evidence required for approval, the human escalation path, and the measurement period before expansion.

Why it matters: The launch packages agent design, deployment controls, operational evaluation, and post-launch improvement into one managed enterprise offering rather than treating a model or chatbot as a finished production system.

Who should care: AI Product Teams, Enterprises, Operators

Pricing: OpenAI did not publish numeric pricing. Presence is available to eligible enterprise customers through a limited general availability program, and the availability section directs interested organizations to their OpenAI account team.

What remains unproven: Presence is not self-serve, numeric pricing is unpublished, and availability is limited. The launch page includes OpenAI-reported deployment results and named design-partner examples; buyers should validate results, permissions, escalation policies, data handling, and operational ownership for their own workflow.

Buzzy

Buzzy launched an agentic infinite canvas for storyboarding, generating, and editing multi-scene AI video projects.

Sources checked: Product Hunt launch and Buzzy’s official product site.

What it does: Buzzy organizes characters, objects, locations, shots, and model outputs on a visual canvas. Makers say users can revise individual scenes, change camera angle or lighting, reuse references across shots, and switch underlying video models while preserving project context. The official site shows film, advertising, animation, music-video, and explainer workflows.

How to evaluate it: Use one character, one location, and one prop across a five-shot sequence. Check visual drift after each model switch, then inspect export quality, editability, generation cost, and the terms governing uploaded references and generated footage.

Why it matters: The launch focuses on continuity and shot-level control, two practical constraints that become more important when AI video work expands beyond isolated short clips.

Who should care: Creators, YouTubers, Marketers, Designers

Pricing: Product Hunt marks Buzzy as having free options and the official site offers a start flow, but no stable public numeric price table was verified. Creators should inspect the current plan limits before committing a long project.

What remains unproven: Public numeric pricing, export limits, rights terms, and independent long-sequence consistency tests were not verified. Product Hunt descriptions about film length and model count are maker claims rather than independent benchmarks.

Kastra

Kastra launched a runtime authorization layer that checks agent, model, and tool actions against policies before execution.

Sources checked: Product Hunt launch, Kastra’s official product page, and its pricing page.

What it does: Kastra applies policies to tools, prompts, inputs, and outputs across several agent environments. Its product pages describe local enforcement, centralized policy management, audit trails, and integrations spanning coding agents and model SDKs. The design goal is to reject an unauthorized action before the underlying tool receives it.

How to evaluate it: Write a small policy set for one coding agent, then test allowed actions, denied actions, malformed tool calls, prompt-injection attempts, and loss of the control-plane connection. Policy coverage and predictable failure behavior matter more than the vendor’s latency claim.

Why it matters: As agents gain access to files, credentials, and operational systems, authorization must happen at action time. Kastra aims to supply that enforcement layer independently of the agent or model.

Who should care: AI Platform Teams, AI Engineers, Developers, Enterprises

Pricing: Kastra’s official pricing page says users can start free and scale to team or enterprise tiers. A stable public numeric price table was not visible in the checked source.

What remains unproven: The sub-millisecond decision claim comes from Kastra and was not independently benchmarked. Teams should test policy coverage, bypass resistance, offline behavior, logging, and failure modes before treating the layer as a complete security boundary.

Related Kingy AI Links

For more launch tracking and founder resources, see AI Launches, AI Tools, and the AI News archive. Founders can also use the AI Sponsored Video ROI Calculator. The safest next step for any product in today’s Radar is a bounded test with explicit permissions, known data, and a clear stop condition.