4 AI drops worth watching: July 24
OpenAI: Health in ChatGPT connects the model to your medical records
On July 23, OpenAI launched Health in ChatGPT, a feature that lets U.S. users connect Apple Health, supported hospital-system records, One Medical, and Function Health so ChatGPT can reason over their actual medical data. It can compare a new lab result against prior tests, summarize what changed since a last appointment, and carry health context into ordinary conversations, invoked explicitly with an @Health mention or through a dedicated Health hub in the sidebar. It runs on GPT-5.5 Instant for free users and GPT-5.6 Sol for paid, both with dedicated health training that OpenAI says performs “comparable to frontier Thinking models” on hard evaluations. Connected data is encrypted at rest and in transit, is “not used to train our foundation models or target ads,” and deletes within 30 days of disconnection. It is available today to users 18 and older on Free, Go, Plus, and Pro across web and iOS.
The pitch is context: most people’s health data is scattered across portals and apps, and the value is a model that can see the whole picture at once rather than a single uploaded PDF.
The take. The guardrails are the actual product here, not the health reasoning. OpenAI leads with the no-training, no-ads, 30-day-delete commitments because the capability is trivial next to the trust problem, and it named the model tiers and their evaluation parity to preempt the obvious safety objection. The load-bearing disclaimer is that ChatGPT “does not replace the care and judgment of qualified medical professionals,” which is doing a lot of work for a feature that will absolutely be used for triage. Watch adoption among the Free tier: if millions connect medical records to a consumer chatbot, the privacy posture stops being a feature and becomes regulated infrastructure. This is the highest-trust data most users own, and it just got a chat interface.
Definitions:
- @Health mention: An in-chat command that pulls connected health data into a specific ChatGPT conversation on demand.
- GPT-5.6 Sol: OpenAI’s paid-tier flagship model, here fine-tuned with dedicated health training.
Screenpipe: a local memory of your screen that agents can query
On July 23, Screenpipe (YC S26) launched with a pitch to give agents “a searchable memory of what you’ve seen, said, and heard.” Rather than continuously recording video, it listens for meaningful events (app switches, clicks, typing pauses, scrolling) and pairs each screenshot with the operating system’s accessibility tree at the same timestamp. Everything indexes locally into SQLite, mp4 files, and markdown, and an API on port 3030 serves that memory to agents with MCP and skills support. The codebase is Rust, MLX, and ONNX. It is source-available rather than open source: free for personal, nonprofit, and educational use, with commercial use requiring a license, though versions before the license change remain MIT.
The design choice that matters is local-first: the recording, indexing, and storage all happen on the device, so the screen history never has to leave the machine to be useful.
The take. Screen memory is the missing input for agents that are supposed to automate your actual work, and pairing screenshots with the accessibility tree is the smart part, because it gives an agent structured semantics instead of asking a vision model to re-read pixels. The event-driven capture is what makes it cheap enough to run all day. The honest caveat is the license: source-available is not open source, and the founder said so directly, so builders evaluating it for commercial use should read the terms before committing. Watch whether the local-first framing holds as the product scales, because a searchable index of everything on your screen is the most sensitive dataset an agent could hold, and the entire value proposition rests on it staying on-device.
Definitions:
- Accessibility tree: The structured, machine-readable representation of on-screen UI elements that the OS exposes for assistive tech, here used to give agents semantic context.
- Source-available: Code you can read and often self-host, but under a license more restrictive than OSI-approved open source (here, commercial use requires a license).
- MCP (Model Context Protocol): The standard for letting AI agents connect to external tools and data sources.
OneCLI: hand agents fake keys, inject the real ones at the gateway
On July 23, OneCLI launched as an open-source credential gateway with a one-line thesis: “Store once. Inject anywhere. Agents never see the keys.” Developers store real API keys in OneCLI and give each agent placeholder credentials. When an agent makes an HTTP request through the gateway, OneCLI matches the request by host and path pattern, swaps the fake key for the real one, decrypts it, and injects it into the outbound request, so the agent never handles the actual secret. Credentials are stored with AES-256-GCM encryption, each agent gets a scoped access token, and the system runs as a Rust gateway on port 10255 with a Next.js dashboard and encrypted PostgreSQL storage. It is Apache-2.0 licensed.
The problem it targets is specific: agents increasingly need broad API access, and handing real credentials to a model that logs, forwards, or hallucinates is a standing liability.
The take. This is the counterweight to every drop above it: as agents reach into more systems, the security question becomes what they can see, not just what they can do, and OneCLI answers by making sure they never see the secret at all. The credential-substitution pattern is the right primitive, because it keeps keys out of prompts, logs, and model context entirely, which is exactly where leaked secrets end up. Apache-2.0 helps adoption; agent-infrastructure tooling rewards permissive licensing. Watch whether this pattern gets absorbed into the major agent frameworks, because a gateway that swaps credentials at request time is generic enough that Claude Code, Codex, and the orchestration layers will want it as a built-in, not a bolt-on.
Definitions:
- Credential substitution: Giving an agent a placeholder key and swapping in the real secret only at a trusted gateway, so the secret never enters the agent’s context.
- AES-256-GCM: A widely used authenticated encryption standard, here protecting stored credentials at rest.
- Scoped access token: A credential limited to specific permissions or resources, so a compromised agent can reach only what it was granted.
Palmier Pro: an agent that edits the timeline alongside you
On July 23, Palmier Pro launched as an open-source, Swift-native video editor for macOS where “you and your agent can generate and edit videos together inside the timeline.” It connects Claude, Codex, or Cursor over MCP (exposing a server at 127.0.0.1:19789/mcp when open), or uses an in-app agent, so a model can operate on the same project you are editing. It also embeds generative models (Seedance, Kling, Nano Banana Pro) directly in the timeline for creating video and images in place. The editor is free with no login; AI generation requires a subscription. It is GPLv3, runs on Apple Silicon under macOS 26 (Tahoe) only, and reports 11.5k GitHub stars.
The interesting move is scope: instead of an agent that exports and re-imports files, the agent works inside the editing surface itself, on the same timeline state as the human.
The take. Exposing the editor as an MCP server is the pattern to watch, because it turns a creative app into an agent-operable surface without a bespoke integration, and it is the same shift InsForge and others chased for deployment: design the runtime around what agents can navigate. Shared-timeline collaboration is a real bet that the next editing workflow is human and agent on one project, not an agent handing back a render. The constraints are steep (Apple Silicon, macOS Tahoe only, GPLv3), so this is an early-adopter tool, not a broad release. Watch whether the MCP-server-per-app idea spreads; if creative tools start shipping local MCP endpoints by default, “connect your agent” becomes a checkbox rather than a project.
Definitions:
- MCP server: A local or remote endpoint that exposes an app’s actions and data to AI agents through the Model Context Protocol.
- GPLv3: A copyleft open-source license requiring derivative works to stay open under the same terms.
- Seedance / Kling / Nano Banana Pro: Current generative video and image models, here embedded in the editor’s timeline.
For your week ahead: the through-line is access. OpenAI wired ChatGPT into your medical records, Screenpipe into your screen, and Palmier Pro into your editing timeline, while OneCLI exists precisely to keep agents from seeing the credentials that reach makes them need. As agents move from the chat box into your actual tools and data, the interesting engineering is the access surface and its guardrails. Build the boundary before you build the reach.
More drops at dropwatch.ai. Want them in your inbox? Subscribe below.