DISPATCH
Wire
MCP in 2026: the protocol grew up The Dispatch Weekly: MCP momentum, lean infra, and the agent shakeout Shipping production MCP servers without the footguns MCP vs raw tool-calling: when the protocol earns its keep The MCP servers worth your stdio Context engineering ate prompt engineering The PaaS-on-a-VPS renaissance Git-push deploys for monorepos that don't fight you MCP in 2026: the protocol grew up The Dispatch Weekly: MCP momentum, lean infra, and the agent shakeout Shipping production MCP servers without the footguns MCP vs raw tool-calling: when the protocol earns its keep The MCP servers worth your stdio Context engineering ate prompt engineering The PaaS-on-a-VPS renaissance Git-push deploys for monorepos that don't fight you
Weekly Digest

The Dispatch Weekly: MCP momentum, lean infra, and the agent shakeout

MCP hardens into real infrastructure, AI coding moves onto the meter, and venture capital quietly abandons the wrapper layer.

By Tishan David 4 min read

The week in review

Welcome to the week to 26 June 2026 (AEST). The interesting action sat less in flashy model launches and more in plumbing: how MCP servers scale, who pays for agent tokens, and which agent startups are quietly running out of runway. Three threads ran through almost everything I read this week, so that’s how I’ve framed it below — MCP hardening into real infrastructure, the cost of AI tooling moving onto the meter, and capital fleeing thin wrappers for orchestration and interop.

MCP’s next spec bets on statelessness

The release candidate for the next Model Context Protocol spec — locked on 21 May, finalising 28 July — drops the initialize handshake and the Mcp-Session-Id header entirely, so any request can hit any server instance (the breakdown on mcp.directory is worth reading carefully). Protocol version and capabilities now ride in _meta on every request, with new Mcp-Method/Mcp-Name headers for routing and ttlMs/cacheScope for HTTP-style caching. In practice a remote MCP server can sit behind a plain round-robin load balancer instead of sticky sessions and deep packet inspection — a real drop in operational tax. My take: this is the change that makes MCP boring in the best way, the kind of decision you only earn the right to make once a protocol has enough production deployments to feel the pain.

Salesforce makes MCP a default, not a demo

Salesforce’s Summer ‘26 release began rolling out from 15 June with MCP wired straight through the stack: hosted MCP servers now generally available, Tableau MCP letting agents query the analytics engine directly, and Agentforce reaching 30-plus partner connectors over MCP and Agent-to-Agent messaging (Salesforce Ben has the rundown). When a vendor this size treats the protocol as table stakes for its agent platform rather than a side experiment, the “will MCP stick?” debate is over. The sharper question now is governance: who audits which tools an enterprise agent is actually allowed to call, and who’s on the hook when one of those tools writes to a production record.

Copilot puts AI coding on the meter

As of 1 June, every GitHub Copilot plan moved to usage-based billing. Premium request units are gone, replaced by GitHub AI Credits — 1 credit = US$0.01 — metered on input, output and cached tokens at API rates (GitHub’s own announcement). Code completions and Next Edit Suggestions stay unlimited, but everything agentic now draws down a monthly allotment: 1,500 credits on Pro, 7,000 on Pro+, 20,000 on the new Max tier. This is the end of the all-you-can-eat era for agentic coding, and teams that never modelled their token spend are about to learn what their workflows truly cost — a gap I keep running into on real builds, which I’ve written up in a couple of recent client case studies.

The agent shakeout starts at the wrappers

Funding coverage this month told a consistent story: capital is concentrating in orchestration engines, enterprise security layers, and interop frameworks, while thin application wrappers struggle to raise bridge rounds — with a meaningful share of early-stage agent startups projected to exhaust reserves by year-end as token costs and slow enterprise sales cycles bite. Meanwhile TechCrunch’s running tally of 2026 layoffs where employers cited AI kept climbing through its 22 June update. My read: the wrapper-over-someone-else’s-API business was always going to compress once the base platforms shipped their own agents — see Copilot and Agentforce above. Defensibility now lives in distribution, proprietary data, or orchestration that’s genuinely hard to copy. Everything in between is acquisition bait.

Governance catches up to the protocol

Late last month the NSA published a 17-page cybersecurity information sheet on MCP security design, flagging uncontrolled automated actions, unscreened inputs, and susceptibility to overload, and recommending least-privilege access plus local-first data handling (the NSA release is public). It landed alongside broader Five Eyes guidance on careful adoption of agentic AI, co-signed by Australia’s ASD. None of the individual recommendations are exotic, which is precisely the point: the agencies are codifying what disciplined teams already do, and giving auditors a document to point at. For anyone shipping MCP servers into a regulated environment, this is now the text procurement will quote back to you.

What I’m testing next

I’m porting a small internal MCP server to the stateless RC pattern to see how well the round-robin promise holds under real load — and whether dropping session state genuinely simplifies the gateway or just relocates the complexity into _meta and request handles. In parallel I’m rebuilding our Copilot cost model against the new credit rates before the September promotional allowances expire, because the default assumption that “agentic mode is free” is about to get expensive. If you’re in Sydney and want to compare notes on either, I’ll be at a few of the meetups on this month’s events roundup.

← Back to the feed