Composio shipped three new Model Context Protocol integrations on July 21, 2026. One connects Claude Code to Bitbucket for version control workflows. Another gives OpenClaw agents access to Make automation platform. The third lets Antigravity IDE orchestrate Ollama open-weight models. Three different agent frameworks, three different target platforms, one shared protocol layer.
Individually, none of these integrations would clear the front page. Together, they illustrate a structural shift that has been building for months: MCP is no longer an emerging protocol. It is the default way agent frameworks connect to external tools.
The Protocol That Won Without a Fight
MCP started as an Anthropic project. The official specification describes it as “an open-source standard for connecting AI applications to external systems,” comparing it to USB-C for AI. The pitch is straightforward: build an integration once as an MCP server, and any MCP-compatible client can use it.
That pitch worked. The protocol’s adoption page now lists Claude, ChatGPT, Visual Studio Code, Cursor, and MCPJam as supported clients. GitHub added an MCP Registry to its main platform navigation, placing it alongside Actions, Codespaces, and Issues as a core developer workflow feature. When a protocol gets top-level navigation on GitHub, the adoption debate is functionally over.
The competitive landscape reinforces this. Every major agent framework now supports MCP natively or through middleware. Composio’s Bitbucket integration page lists 20 framework targets for a single tool: Claude Code, Antigravity, OpenAI Agents SDK, Claude Agent SDK, Claude Cowork, Codex, Kimi Code, Grok Build, OpenCode, Cursor, VS Code, OpenClaw, Hermes, CLI, Google ADK, LangChain, Vercel AI SDK, Mastra AI, LlamaIndex, and CrewAI. That is not a protocol still seeking adoption. That is a protocol that has already achieved it, according to Composio’s integration listings.
Composio’s Managed Layer Strategy
The most aggressive bet on MCP’s dominance comes from Composio, a startup that raised $29M in July 2025 to build what it calls “skills that evolve with your agents.” The company’s positioning is simple: it provides a single managed MCP endpoint that routes to over 1,000 apps and 20,000 tools, handling OAuth, API keys, token refresh, and scoping so developers never touch authentication plumbing.
The three integrations shipped on July 21 demonstrate the model. A developer connecting OpenClaw to Make’s automation platform does not write custom API code. They point OpenClaw at Composio’s managed MCP URL, authenticate once, and their agent gains access to Make’s 1,000+ app integrations. As Composio’s documentation states: “OpenClaw is the fastest growing agent harness out there, which can work 24/7 to automate almost any kind of tasks. However, its capabilities are limited to the tools it has access to.”
The Ollama integration for Antigravity follows the same pattern but addresses a different problem. Google’s agentic IDE defaults to Gemini models. By routing through Composio’s Ollama MCP server, Antigravity agents can execute open-weight models locally, reducing vendor lock-in and marginal inference costs. Antigravity treats its agent as “a first-class teammate, planning, coding, and validating work through an Agent Manager dashboard,” according to Composio’s integration guide. Adding Ollama through MCP means that teammate is no longer restricted to Google’s model pricing.
The Bitbucket integration pushes coding agents further into DevOps territory. Claude Code agents can now create branches, open pull requests, and manage repositories through natural language commands. Composio lists 105 available tools for Bitbucket alone, covering the full version control lifecycle.
These are not three isolated product launches. They are three instances of the same abstraction: wrap an enterprise tool behind MCP, let any agent framework call it, manage authentication centrally. The pattern repeats at industrial scale. Composio has shipped MCP integrations for Discord, Google Search Console, Scrapegraph, and dozens more in the past week alone.
The Managed Layer Trade-off
Composio’s approach comes with a concentration risk that the company has already experienced firsthand. In May 2026, Composio disclosed a security incident affecting internal systems, following a separate incident in April. When a single managed endpoint mediates between an agent and 1,000+ applications, a breach at the middleware layer exposes every connected service.
This is the structural tension in managed MCP: the same centralization that makes integration convenient also creates a high-value target. An attacker compromising Composio’s infrastructure would potentially gain access to OAuth tokens, API keys, and session state for every connected agent across every framework. The company’s two incidents within two months suggest this is not a theoretical concern.
The alternative is self-hosted MCP servers, where each organization runs its own integration endpoints. That approach eliminates the single-point-of-failure problem but reintroduces the authentication and maintenance burden that Composio exists to solve. Most teams building agents do not want to run their own OAuth infrastructure for 20 different SaaS platforms. They want to point their agent at a URL and start working.
Skills as the Interface Layer Above MCP
While MCP handles the transport and authentication layer, a parallel ecosystem is standardizing what agents actually do with those connections. The agent skills ecosystem, which defines reusable procedures and workflows as structured markdown files, has reached the point where mainstream tech publications are writing end-user guides.
MakeUseOf published a guide on July 21 recommending five skills across Claude, OpenClaw, and Hermes. The article highlights skills for PR review pipelines, web-to-knowledge-base conversion, document generation, Git workflow automation, and a meta-skill for building other skills. The key observation, as MakeUseOf’s Yadullah Abidi writes: skills “can be run across Claude, OpenClaw, and Hermes with only minor wiring differences.”
That cross-framework portability is the critical detail. Skills sit above MCP in the stack. MCP provides the transport layer (how an agent authenticates and calls a tool). Skills provide the procedure layer (what the agent does with that tool, in what order, with what guardrails). A PR review skill, for example, uses MCP to access GitHub’s API, but the review scoring criteria, the two-stage approval workflow, and the comment formatting are all defined in the skill’s specification, independent of which MCP endpoint provides GitHub access.
Anthropic formalized this with an official skills repository on GitHub, shipping document-handling skills for DOCX, PDF, PPTX, and XLSX files. These are the same skills Claude.ai uses internally, released as open-source packages that any framework can load. The pattern extends to third parties. Firecrawl publishes a knowledge-base skill. SpillwaveSolutions publishes a PR reviewer skill. Both install with a single command and work across frameworks.
The emergence of MakeUseOf-style “best skills” articles signals that this ecosystem has crossed from developer tooling into consumer-adjacent territory. When mainstream tech media starts recommending specific agent skills the way it recommends browser extensions, the ecosystem has achieved a level of maturity that changes how frameworks compete.
How Frameworks Compete When Integration is Commoditized
If every agent framework can access the same 20,000 tools through the same MCP endpoint, and run the same skills with minor wiring differences, the traditional competitive moat around integrations disappears. GitHub Actions competed partly on the breadth of its marketplace. Zapier competed on the number of app connections. Those advantages depended on proprietary integration work that competitors could not easily replicate.
MCP eliminates that barrier. When Composio ships a Bitbucket integration, it works simultaneously across Claude Code, OpenClaw, Antigravity, and 17 other frameworks. No framework can claim exclusive access to a tool integration because the protocol is open and the managed layer is framework-agnostic.
This shifts competition to three remaining dimensions. First, the agent runtime itself: how reliably an agent can plan, execute, and recover from multi-step tasks. Second, the orchestration layer: how a framework handles scheduling, persistence, and multi-agent coordination. Third, the governance layer: how an organization controls what agents can access, what they can spend, and when they need human approval.
OpenClaw competes on orchestration, with persistent agents that run 24/7 across Telegram, WhatsApp, and CLI interfaces. Claude Code competes on code quality, with Anthropic’s models fine-tuned for software engineering tasks. Antigravity competes on the development environment, embedding agents directly in the IDE with visual dashboards. None of them compete on which SaaS tools their agents can access, because MCP made that a commodity.
The Infrastructure Stack Taking Shape
The agent infrastructure stack that is emerging looks increasingly similar to the web infrastructure stack that solidified in the 2010s. HTTP standardized transport. REST and GraphQL standardized API design. OAuth standardized authentication. Container registries standardized deployment. Each layer commoditized a previously competitive dimension, pushing differentiation upward.
MCP is performing that function for agent infrastructure. It standardizes how agents discover, authenticate with, and invoke external tools. Managed MCP providers like Composio add a convenience layer on top, the way managed database providers added convenience on top of PostgreSQL. Skills standardize reusable agent procedures, the way container images standardized reusable deployment units.
The three integrations Composio shipped on July 21 are individually minor. Collectively, they are data points in a standardization curve that has already passed its inflection. The 20 framework targets listed on a single Composio integration page tell the story more clearly than any single product launch: MCP already won the integration protocol war.
The remaining question is not which protocol will prevail but who controls the managed layer above it, and whether that concentration creates systemic risk for the agent ecosystem it is designed to serve.