← BASELOAD English

Two Ways to Wire Claude Code into Paperclip — The MCP Connection [Part 3]

October 17, 2026 · 71 min read 한국어 원문

Since Part 1 and Part 2 went out in July, Paperclip has shipped twelve releases (patches included — nine if you count only feature releases). The most structural of them is the “Assistant Connection (MCP)” in the October 12, 2026 release. In one line: Claude Code can now log into Paperclip as a user and act like a member of the team.

This is where people get confused. Paperclip could already connect to Claude Code. But that connection runs the other way around. Telling those two directions apart is the whole point of this article.

⚠️ Up front: this is not a write-up of an installation I ran myself. It is assembled from Paperclip’s official documentation and GitHub release notes, and every screen, config key, and step below is quoted from primary sources. Where something may differ from the live product, I say so in the text.

The three questions that need answering first

Photo by Pavel Danilyuk on Pexels

Two Directions — Who Is Hiring Whom

Answer first: with the Process Adapter, Paperclip is the boss; with the MCP connection, Claude Code is a colleague.

Why does the distinction matter? The “company of AI employees” from Part 2 is the top row — you define agents, run them on a heartbeat, and keep them in check with budgets. What is new is the bottom row: a path for the Claude Code you use every day to reach directly into that company’s work system.

The Existing Path — Running Claude Code as an “Employee” via the Process Adapter

Answer first: set adapterType: "claude_local" and Paperclip spawns your local claude CLI as a child process. The official docs describe it as running “Anthropic’s Claude CLI, including session support and skill integration.”

The config keys look like this.

Four constraints worth knowing (per the official docs)

1. The timeout defaults to 900 seconds. Past that it sends SIGTERM, then SIGKILL after the grace period. Long jobs need this raised
2. Exceeding maxTurnsPerRun resets the session
3. Session resume is conditional — it continues only when the agent has a stored sessionId and cwd matches the saved working directory. Change the work folder and the context does not carry over
4. It assumes a single machine. The docs say the Process Adapter is “suitable for development” and point to the HTTP adapter for distributed execution

One thing to watch here: the official docs trail the releases. The docs list Claude Opus 4.6, Sonnet 4.5 (recommended), and Haiku 4.5, while the October 12 release notes say Claude Code agents can select Claude Sonnet 5.5 at Low through Max effort.

That does not mean “just copy the release notes into your config,” though. The release notes give display names only — they do not publish the model ID strings you put in the config. So the safer move is to read the model value off the model menu in the product UI. Two more things to keep in mind:

• The allowed effort values differ. The Process Adapter config in the docs gives low, medium, and high as examples. Whether the “Max effort” from the release notes maps straight onto a Process Adapter config value I could not confirm
• There is a sandbox image rebuild condition. From the release notes: “Custom sandbox images built with a Claude Code version older than 2.1.284 must be rebuilt before Sonnet 5.5 works on the CLI path.” The same release bumped the harness pin to Claude Code 2.1.286

The New Path — Joining the Team via Assistant Connection (MCP)

Answer first: Claude Code connects through the “Assistant Connection (MCP)” card under Connections. Here is the release note verbatim:

A new “Assistant Connection (MCP)” card in Connections lets Claude Code, Codex, OpenCode, and generic MCP clients connect to Paperclip through browser OAuth or device-code consent and act as the logged-in user.

What a connected Claude Code can do

Controls ship alongside it — write-scope restriction (pin it to read-only) and single-company pinning (limit access to one company) are both options. Since the structure lends out your logged-in user’s privileges wholesale, those two are the ones to turn on before you start.

How to enable it (per the official release notes)

1. Enable Settings → Experimental → "Assistant connections (MCP)" — still experimental (config key enablePublicMcp). The release notes state that this switch “exposes a public MCP endpoint and requires PAPERCLIP_PUBLIC_URL to be an HTTPS origin”
2. Set an HTTPS PAPERCLIP_PUBLIC_URL — OAuth redirects mean an HTTP localhost will not do
3. Connect your client from the Assistant Connection (MCP) card under Connections (browser OAuth or device code)
4. Local stdio clients use the paperclip mcp CLI bridge together with device login

Step 2 is the real gate. If you are running locally the way Part 2 described — npx to start it, localhost to reach it — you have to arrange a public HTTPS URL first.

One thing is easy to mix up here. The August 24 release added Tailscale-based HTTPS exposure (PAPERCLIP_MANAGED_RUNTIME_HTTPS), but that is a different layer. The release notes are explicit about the target: it exposes the “managed worktree runtime” over tailscale_https, and the problem it solves is “not being able to view an agent’s branch preview from your phone or another computer.” In other words it is for previewing agent output, and it has nothing to do with the PAPERCLIP_PUBLIC_URL that MCP requires (the HTTPS origin of Paperclip itself).

Putting it behind a TLS-terminating platform (Railway, Render, Fly, and so on), which the same release mentions, looks like the more direct route to a public HTTPS origin. That part, though, I could not verify — this is not the result of standing one up.

Photo by Simon Kadula on Unsplash

So Which One Do You Pick

Answer first: Process Adapter if you want work to run itself; MCP if you want to drive the company yourself.

They are not mutually exclusive. Leaving your employee agents running and attaching your own Claude Code on top as a colleague is, in practice, the most natural picture.

What Changed Since Part 2 — For Readers Who Followed Along

Answer first: the default UI changed. The screens described in Part 2 no longer match what you will see.

Of these, the first row is what trips up readers following Part 2 the most. The article’s logic still holds, but the default screen moved, so the description and what is in front of you no longer line up.

Prerequisites and Pitfalls

Things to check before you actually wire this up.

1. It is experimental. You have to enable it under Settings → Experimental, and experimental specs change. Safer to try it in a separate company before dropping it into real operations
2. A public HTTPS URL is mandatory. On a local localhost setup, this is the step that takes the longest
3. The privileges are those of the logged-in user. The connected assistant acts as “you,” so turn on read-only restriction and single-company pinning first, then widen the scope
4. The official docs trail the releases. As of the date checked (2026-10-15), the hosted docs index has no MCP-specific page, and the model list is out of date. The GitHub release notes are more current
5. Screens that are not in the docs can move. The steps here follow the release-note descriptions, so the actual card placement and wording may vary by version

So What — The Reversal of Roles Is the Point

Paperclip’s original picture was “I manage a company staffed by AI employees.” That was Parts 1 and 2. What the MCP connection adds is not one more feature but an inversion of the relationship. Now the coding agent I use has direct access to that company’s work system.

Why does that matter in practice? Until now, the weak spot of the “AI agent company” was that it sat apart from my day-to-day work. The company ran on its own, I worked separately in my terminal, and the thing joining them was a human hand. The MCP connection erases that hand — my Claude Code reads the company’s tasks, creates them, delegates them, and pulls back the output.

So the order I would recommend is this:

1. Connect with read-only plus a single-company pin first, and let your Claude Code only query the company’s state
2. Once querying proves useful, open up writes and widen to creating and delegating tasks
3. Keep overnight work with the Process Adapter employees, and use MCP as the console you command from

Mixing the two directions is the destination; trying to replace one with the other is reading the design wrong. One is autonomous execution, the other is the console you drive.

Related reading

• Building a Company Staffed by AI Employees — A Paperclip Primer [Part 1]
• Paperclip Hands-On — From Setup to Running an AI-Staffed Company [Part 2]

Sources

• Paperclip v2026.1012.0 release notes — Assistant Connection (MCP): connection methods, privileges, prerequisites, model support
• Paperclip release list — change history since July (chat-style default, Decisions, Connections rework, and more)
• Process Adapter documentation — claude_local config keys and constraints (timeout, turn limit, session resume, single machine)
• Agent Integration documentation — the adapter system
• Paperclip documentation index — the basis for saying there was no MCP-specific page at the time of checking
• Paperclip GitHub repository

※ This article is based on official documentation and release notes as of October 15, 2026, not on a hands-on installation. Because the feature is experimental, screens and specs may change — re-check the release notes before you apply any of this.