Carrel: A Private Desk in the Library, Set Up by AI

The researcher never touches a terminal. Everything is conversational.

Carrel is a Claude Code plugin that onboards researchers into an AI-augmented research environment. Instead of shipping a checklist, it interviews the researcher, audits the machine, installs what the conversation actually warrants, and learns how to work with this specific person along the way. The product is not the install — it is the relationship that begins during install.

Why This Matters for Research

The list of tools at the bottom of any "AI research stack" tutorial is correct. Following it is still the wrong way to start, because you end up with a configured machine and no understanding of why each piece is there.

Carrel inverts that. The setup-interviewer agent has a conversation about what you actually do — and only then installs what the conversation warrants. A short stretch reads roughly like this:

Setup-interviewer: What kind of source material do you work with most? Mostly PDFs, mostly interview recordings, mostly notes from the web, or a mix? You: A mix, but heavy on interview audio right now. Setup-interviewer: Good to know. I'll install coli for transcription and ffmpeg as its dependency. For PDFs I'll set up liteparse for the simple ones — and I'll ask about adding mineru-mcp for complex ones after we test on a sample. Anything sensitive about the audio? Patient interviews? Anything under NDA?

The product of the interview is a stack you can explain. The transcript of the interview is also a stack you can refer back to — /carrel-status and /carrel-fix work against the same understanding.

Underneath, it bootstraps:

• Obsidian as the shared vault — the researcher sees a GUI notebook; Claude sees markdown.
• carrel Python CLI with eight subcommand groups (paper, transcript, capture, google, vault, env, setup-state, trust).
• Conversion tooling — liteparse, markitdown, defuddle, mineru cloud OCR for complex PDFs.
• Transcription via coli (with ffmpeg).
• Google Workspace integration via gws for researchers who use Drive/Docs.
• Optional MCPs layered in based on the interview — Vox for multi-model access, MinerU for complex PDFs, zotero-mcp for citation pipelines.

Sensitivity-aware by default: local-first, nothing leaves the machine unless explicitly chosen.

Host Support

If you're a researcher choosing a host: Claude Desktop, Code tab. It is the path Carrel is tuned for and the one the install script verifies.

What's in the plugin

The plugin ships slash commands for the routine moves (setup, status, recovery, conversion, transcription, capture, automation, reflection), two agents (@setup-interviewer for onboarding and @research-partner for thinking-with), and a library of intent-discoverable skills that Claude pulls in on demand — vault operations, environment doctor, knowledge-wiki tending, model-teammate management, and so on. Underneath sits a single carrel Python CLI that the slash commands shell out to, so the same operations work from a terminal when you need them.

The exact slash-command list is below; the agents and skills are documented in the source repo.

Commands

Trust Levels (Code-Enforced)

Carrel's automation trust model is not narrated in skill markdown — it is enforced in code. Four levels of relationship between researcher and assistant define the contract:

• Advisory — suggest, never act.
• Consultative — propose actions, ask before each.
• Delegated — act within named boundaries, report after.
• Partnership — act with discretion, surface what matters.

Writes that cross those boundaries pass through carrel trust check <action> before the skill proceeds. A delete that the current vault classifies as outside the level — say vault.archive-bulk at the Advisory level — returns a non-zero exit and a single line explaining why:

$ uv run carrel trust check vault.archive-bulk --vault .
denied: vault.archive-bulk requires Delegated; vault is at Advisory

The skill sees the denial and stops, rather than asking for forgiveness afterward. To inspect the matrix without attempting an action:

uv run carrel trust list --vault .

The point isn't to be permissive or restrictive by default — it's to make the relationship contract legible. A researcher can read their vault's trust matrix and know what their Claude can and can't do without asking.

Installation

There are two install paths. Path A is recommended — it's more reliable because Claude Code in the terminal can diagnose and fix issues autonomously.

brew install gh
gh auth login

uv tool install git+https://github.com/linxule/carrel.git

Verifying installation

In any Claude Code session:

/carrel-status

If Carrel is active, it reports environment state. If not installed, the command won't be recognized.

Facilitator-assisted setup

When setting up a researcher's machine in person:

What Carrel installs

Carrel installs three layers, in the order they come up during the interview.

The environment (always)

These are the substrate. Carrel installs them on every machine; the install script handles the prerequisites and the slash commands use them at runtime.

• Obsidian — the vault GUI; researcher sees a notebook, Claude sees markdown
• carrel Python CLI — eight subcommand groups (paper, transcript, capture, google, vault, env, setup-state, trust) that the slash commands shell out to
• Conversion stack — liteparse (PDFs), markitdown (general formats), defuddle (web pages)
• Transcription — coli plus ffmpeg
• Google Workspace — gws (Drive / Docs), if the interview surfaces it

None of these have Toolkit pages because they're not agent-callable surfaces — they're CLIs the plugin orchestrates.

MCPs (interview-conditional)

These ship as MCPs that any Claude Code project can use. Carrel adds them at project level only if the interview surfaces a need:

Each MCP requires its own API key, which /carrel-setup will prompt for.

What Carrel doesn't install today

Carrel's current onboarding doesn't install other Claude Code plugins or standalone web platforms. Both are scope decisions, not architectural limits — Carrel's agents could shell out to claude plugin install like any other bash command, and Codex agents can install Claude Code plugins cross-host the same way. Today the plugin layer is a separate question from the environment layer Carrel handles. The Toolkit entries that pair cleanly are sibling plugins you install yourself for now:

• Interpretive Orchestration — the qualitative-methodology layer; bundles its own MCPs (Sequential Thinking, Lotus Wisdom, Markdownify) used by /qual-* commands. Carrel's docs recommend pairing for Gioia / grounded-theory work.
• Memex Plugin — collaborative session memory across compaction windows.
• Kimi Plugin for Claude Code — independent second-model reviewer.

Install these separately from the Claude Code plugin marketplace. They coexist with Carrel without sharing state.

claude plugin marketplace add linxule/interpretive-orchestration
claude plugin install interpretive-orchestration@interpretive-orchestration --scope user

Three plugins, three jobs: Carrel sets the desk; Interpretive Orchestration runs the analysis; Memex remembers what happened. Add Kimi Plugin when you want a structurally different model checking the work.