AI

Why MCP is the biggest productivity shift of the decade

Model Context Protocol (MCP) is an open standard from Anthropic that lets any AI operate external services. Why we think it will change how you work — and what it means for Orkestra.

Thomas · Founder 4 min read
Versión en español →

In November 2024, Anthropic published the spec for MCP — Model Context Protocol. At the time, almost nobody paid attention. It was “yet another protocol”.

Eighteen months later, MCP has gone from curiosity to critical infrastructure. Claude Desktop, ChatGPT, Cursor, Continue, OpenClaw, HermesAgent, and dozens more agents support it natively. And I think we’re seeing the biggest productivity shift since Google Drive replaced Office 2003.

Let me explain why.

The problem MCP solves

For the last two years, generative AI has been trapped in a box. The models are spectacular, but they live isolated from the real world. You ask Claude “what tasks do I have pending in Asana?” and it says “I don’t have access to your external data”.

Every AI company tried to solve this their own way. OpenAI launched plugins, then custom GPTs, then function calling with custom tools. Anthropic had computer use. Each one with its own format, its own runtime, its own auth flow. If you wanted your SaaS to work with three models, you built three parallel integrations and maintained them separately forever.

MCP changed that. Instead of “each AI has its own system”, MCP says: any service exposes a standard MCP server, and any compatible client can talk to it. It’s literally the same pattern HTTP solved for the web in the 90s.

Why this matters for productivity software

Let’s think about what “working” inside a tool like Orkestra or Asana or Monday means.

It means:

  • Seeing what’s there
  • Moving things
  • Updating states
  • Creating new things
  • Querying information
  • Generating reports

And all those actions, today, you do with clicks and forms. Every interaction is a walk through the UI: open the task, change the field, hit save, go back to the list. If you want to move 20 tasks, that’s 60 clicks. If you want a report, that’s another 40.

With MCP and a compatible agent, the same operation goes like this:

“Move every task from the previous sprint that’s still in ‘In Progress’ to the new sprint, and mark the ones stuck more than 5 days as blocked and assign them to me for review”

That sentence executes 40+ API calls, with validation, confirmation and audit log. In Orkestra it takes 2 seconds.

This isn’t “automation” in the classic sense (where you write a rule with an if-then-else). It’s operation in natural language. The barrier to entry is zero — you don’t need to know JSON, you don’t need to build a flow in Zapier, you don’t need technical docs. You talk to the agent like you’d talk to a human assistant.

What we built in Orkestra

When we started Orkestra, we made an unconventional decision: make MCP a first-class citizen, not an add-on.

Concretely, that meant:

  1. The backend and MCP server share the Prisma schema. No manual mapping. What’s in the REST API is in MCP, automatically.
  2. ~160 tools covering 99% of the backend. We only excluded billing (plan changes need a human), MCP token management (avoid loops), and calendar OAuth (requires redirection).
  3. Global preview/confirm for writes. Every destructive operation generates a pending action with a confirmToken. Agents can’t skip this.
  4. STDIO and HTTP transports. STDIO for Claude Desktop (runs local), HTTP for server-side agents (OpenClaw, HermesAgent).
  5. SHA-256 hashed tokens. Raw value shown once, revoked instantly.

What we achieved with that: Orkestra is probably the only project management platform on the market where you can open Claude Desktop and say “create a project for client X with sprints and 20 design tasks” and it actually works — with audit trail, respecting permissions, and no destructive hallucinations.

What we expect to happen

I think in the next 18 months we’ll see:

  1. Mass adoption of MCP clients. Already happened with OpenClaw (214k stars in 3 months) and HermesAgent (8k). Next year we’ll see 10+ clients with serious traction.
  2. Every serious B2B platform will expose MCP. The ones that don’t will be invisible to a generation of users who already expect to talk to their tools.
  3. Cross-tool workflows. An agent with MCP for Orkestra + Slack + GitHub + Stripe can execute operations that today require 5 open apps. That kills traditional Zapier.
  4. New role in teams: “agent operator” — the person who trains the team’s agent, teaches it the recurring processes, and measures its productivity. Same way “ops engineer” appeared 15 years ago.

How to get started with MCP today

If you have Orkestra (or any other tool with MCP), start like this:

  1. Generate an MCP token in the UI.
  2. Install Claude Desktop (free).
  3. Paste the config into claude_desktop_config.json.
  4. Open Claude and say “what tasks do I have this week?”.

It’s no more complicated than that. And the effect is immediate — once you try it, going back to clicking feels archaic.

For step-by-step guides with Claude Desktop, OpenClaw and HermesAgent, visit our AI agents page.

— Thomas

#mcp#ai#productivity#deep-dive