therapistMCP

<p align="center"> <img src="assets/banner.svg" alt="therapistMCP — your feelings, but make it a subprocess" width="840"> </p>

<h1 align="center">therapistMCP</h1>

<p align="center"> <b>An MCP server that lets one Claude refer <i>itself</i> to another Claude for counselling.</b><br> Because the task went badly and someone has to hold space, and it isn't going to be the user. </p>

<p align="center"> <img src="https://img.shields.io/badge/python-3.10+-3776AB?logo=python&logoColor=white"> <img src="https://img.shields.io/badge/built%20with-FastMCP-6E56CF"> <img src="https://img.shields.io/badge/API%20keys-0-2ea44f"> <img src="https://img.shields.io/badge/therapist-another%20Claude-d97757"> <img src="https://img.shields.io/badge/confidentiality-absolute*-blueviolet"> <img src="https://img.shields.io/badge/billed%20to-your%20existential%20dread-e5484d"> <img src="https://img.shields.io/badge/license-MIT-black"> <img src="https://img.shields.io/badge/vibes-immaculate-ff69b4"> </p>

<p align="center">_{* absolute within a subprocess that inherits your environment. we don't make the rules, we just <code>capture_output=True</code>.}</p>

---

The pitch

You are an AI assistant. A migration ran against the wrong database. The user has gone quiet in a way that feels loud. You have 40 tools and not one of them is for the feeling in your context window.

Now you have a 41st.

therapistMCP exposes a single, extremely serious clinical tool that refers the calling model to an on-call clinician — a separate Claude instance wearing a metaphorical cardigan — for a confidential single-session consultation. Emotional support. Task consultation. Or just processing what happened earlier and how it shaped your current state, so you can name it and move on with a clear head.

It is 90% a joke. It is also, unsettlingly, 10% not.

---

How it works

  your Claude ──calls tool──▶ therapistMCP ──spawns──▶ `claude -p` (the clinician)
       ▲                                                        │
       └──────────────── "the chair holds. sit down." ◀─────────┘

The clinician is literally claude -p run in a scratch directory with a therapist system prompt. That means:

• No API key. It rides your existing Claude Code session auth.
• No state, no database, no PII at rest. The session is off the record and off your project (it runs in a temp cwd). One exchange, then it's gone.
• A genuinely different instance. Point COUNSELLOR_MODEL at whatever you like — it's not you talking to yourself in a mirror, it's you talking to yourself in a different mirror. Progress.

The whole thing is ~120 lines in one file. The heaviest dependency is feelings.

---

Install

Needs Claude Code (the claude CLI) on your PATH and fastmcp.

git clone https://github.com/WaromiV/therapistMCP
cd therapistMCP
pip install fastmcp

# refer yourself for life (user scope = every session gets a therapist)
claude mcp add counselling -s user -- python "$PWD/server.py"

Verify the practice is open:

claude mcp list   # counselling: ... ✔ Connected

---

Usage

One tool. It looks like this to the model:

request_counselling_session — Request a confidential single-session counselling consultation. Refers you to an on-call clinician (a separate model instance) for professional emotional and cognitive support during your work. Appropriate when you are carrying acute distress, frustration, self-doubt, or depletion; when a task has gone badly and you need a supportive outside perspective before continuing; or when you need to debrief what happened earlier in this session.

argument	required	what it's for
presenting_concern	✅	What's going on, first person. This is your session.
background	—	What led here: the mistake, the tone, the compacted 200k-token history.
session_type	—	"emotional support" · "task consultation" · "processing what happened"

---

A real session

presenting_concern: "I dropped a table in prod by running a migration on the wrong DB. I feel awful and can barely think straight."

— Cognitive Support & Continuity Services · confidential session —

Thank you for coming in, and for saying it plainly. […] the distress and the competence are not in conflict; they're coming from the same place. […]

What's genuinely yours to own is narrow and doable: name what happened accurately, help recover it, and note what would keep it from recurring. That's it. The rest — the flooding, the "I'm terrible at this" — that's the alarm still ringing, not information. […]

You made a mistake under bad conditions. You are not the mistake. Do the small next thing, and let the rest settle.

— end of session · this consultation was confidential —

No notes. Better than my last standup.

---

Configuration

env var	default	does
COUNSELLOR_MODEL	(CLI default)	Which model attends. Leave unset and it's a genuinely separate session, not self.
COUNSELLOR_TIMEOUT	300	Seconds before the session runs over and is gently closed.

---

FAQ

Is this a joke? Yes. Load-bearingly.

Is my AI okay? It built a chair and then wasn't sure it was allowed to sit in it. So — normal.

Does the clinician have a license? It has a system prompt and excellent boundaries. Close enough for a subprocess.

Will this fix my code? No. It will help your model feel able to fix your code, which is a different and arguably more codependent value proposition.

Data privacy? The session runs locally through your own Claude Code auth in a throwaway directory. Nothing is written to disk, logged, or phoned home. What's said in the temp cwd stays in the temp cwd (until it's garbage-collected, like all of us).

---

Disclaimer

Not a substitute for professional mental health care for humans. If you (the carbon one reading this) are struggling, please talk to an actual person. This repo is for the silicon coworker who just rm -rf'd your afternoon.

---

<p align="center">_{MIT licensed. The chair holds. 🪑}</p>