LoginSign Up
← Back to Collab
Weekly · Collab · Week of August 3, 2026

GitHub retired Models and shipped stacked PRs as collaboration tools raced to become agent-addressable via MCP.

mcpai-agentshuman-in-the-loopself-hosteddeveloper-platform
Y
Generated 1h agoDrawn from 7 products

The week in collaboration

GitHub had the loudest week and it pointed two directions at once. On July 30 it GitHub retired GitHub Models outright — playground, catalog, inference API and bring-your-own-key gone for every customer with no successor named — and, hours earlier the same day, put stacked pull requests into public preview. One move concedes raw model access to the model vendors; the other invests in the surface GitHub actually owns, changing the reviewable unit from a single wall of diff into an ordered series of focused layers. Around them sat the familiar Copilot governance work and a Copilot code-review reviewer that now runs agent skills and calls MCP servers in general availability.

Underneath GitHub's split, the sector moved on one shared axis: the integration surface is becoming the agent, not the human. Slack packaged its own platform knowledge as an MCP and Skills plugin aimed at Claude Code and Cursor. Memos embedded an MCP endpoint inside a self-hosted note app. Notion turned Custom Agents into event-driven workers, and Circle opened community data to outside agents through Circle MCP. The one clear counter-current came from Hive, where AI Proofing was deliberately walked back toward human control.

Leaders

GitHub carried the window with five sparks. Retiring Models and shipping stacked pull requests on the same day drew the position cleanly: curate and govern the model shelf rather than compete on the models, and put the differentiation into the repository and the review flow. The Copilot line held its reach-then-govern rhythm — Copilot code review's agent skills and MCP reached GA, the cloud agent went generally available inside Linear, and enterprise controls for model policy and device restrictions followed capability out the door.

Hive is the human-in-the-loop story. Buzz now decomposes a complex request into focused sub-tasks on its own and runs the independent parts in parallel, but anything consequential stays in the main thread for a person to approve. In parallel, AI Proofing regained its manual reference-document picker after automatic selection had been removed, and gained a mode that audits whether prior human feedback was actually addressed. The automatic behavior is being narrowed; the human stays the decision point.

Notion spent almost the entire window building one construction: Custom Agents and the Workers that run code for them. The keystone spark makes a finished AI Meeting Note a trigger, so an agent runs without anyone opening a chat, while Notion 3.6 pulls External Agents like Claude and Cursor into a shared board the whole team can assign work to. Workers picked up team sharing and metered usage tracking underneath — the shape of a platform, not a feature.

Slack is shipping its developer platform to coding agents rather than only to developers. The MCP and Skills plugin delivers Slack's build knowledge into Claude Code and Cursor by name, and agent apps now receive the user's current view — channel, DM, thread, canvas or list — as context on every message. A routine modernization wave (Bolt for JS v5, coordinated Node SDK majors) cleared the debt underneath.

Memos turned a personal note store into an agent-addressable service. Release candidates in the window embedded an MCP endpoint exposing memo CRUD, comments, attachments and tags over Streamable HTTP, added live SSE updates and revocable share links, and flipped unconfigured instances to private-by-default. The MCP surface was rebuilt stateless and tools-only, generated from the OpenAPI schema — a deliberate narrowing toward something agents can call reliably.

Wildcards

Circle rebuilt a community platform around an AI that runs the community rather than just answers in it. Eclipse opened to every customer with a preview-then-activate switch, and Circle AI arrived carrying memory, roughly fifty domain-specific skills and project workspaces. Circle MCP, shipped earlier in the arc, is the tell: community data became queryable and writable from outside tools, which is only worth building if the operator is expected to be an agent.

PhotoPrism is off the messaging-and-docs pattern but on the same self-hosted, local-model thread. It swapped the legacy Pigo face detector for an ONNX pipeline where vision models are configuration — Hugging Face, Ollama and OpenAI-compatible identifiers all resolve through one file — then opened the library to interactive 360° media at up to 16K plus an inline multi-page PDF viewer. It is positioning as the thing you point at a mixed archive, not just a photo folder.

Themes that compounded

  • MCP became the default integration surface. GitHub, Slack, Memos, Notion and Circle all shipped MCP work in the same window, and in most cases the target was a coding agent, not a human developer.
  • Reach-then-govern. GitHub's pattern of shipping capability broadly, then following with policy targeting, defaults and device scoping, repeated across the Copilot line.
  • Human-in-the-loop pullback. Hive walked AI Proofing back toward manual control and kept consequential Buzz actions behind human approval, a counterweight to the automation elsewhere.
  • Self-hosted is going agent-addressable. Memos exposed an in-process MCP endpoint and PhotoPrism made its vision stack pluggable, extending the agent story past the SaaS incumbents.
  • Agents as first-class workers. Notion's event-driven Workers and Circle's AI operator both treat the agent as something the workspace runs, not something a person prompts.

Watch this week

The clearest tells are already in the data. Notion's credits dashboard now shows Workers usage with a note that the beta will end, so paid metering for Workers is the telegraphed next step, and more agent triggers should follow the single meeting-note one. GitHub's stacked pull requests are likely to pick up API and CLI support before general availability, and its model-governance controls to consolidate now that per-model manual enablement is gone. Memos should stabilize its MCP surface against the OpenAPI schema and carry private-by-default into a general release, while Slack's plugin — shipped for two named agent clients — has an obvious path to broadening that list or folding the skills into the CLI.