Skip to main content

Open Source

OpenCouch

A calm, session-aware mental health support companion that helps you feel heard, make sense of what you're going through, and find one grounded next step — without pretending to be therapy.

Start with QuickstartExplore Safety ModelView Agent Runtime →
Every turnSafety-first runtime pipeline
  1. 01
    User message

    The current turn enters the backend runtime.

  2. 02
    crisis_gaterequired

    Safety check runs before memory retrieval.

  3. 03
    turn_dispatch

    Safe turns route to memory, tools, or support.

  4. 04
    load_memory

    Context loads only after safety clears.

  5. 05
    TherapeuticAgent

    Generates bounded support or guided exercise turns.

  6. 06
    finalize

    Appends response, diagnostics, and transcript state.

OpenCouch loads memory only after the safety gate clears the turn.

Start hereSafety Runtime

Crisis gate, risk levels, routing policy, and safety boundaries.

apps/backend/agent
Core graphAgent Runtime

Graph topology, state reducers, tools, flows, and dispatch.

apps/backend/agent
Context layerMemory System

Semantic, episodic, and procedural memory with retrieval and privacy controls.

apps/backend/agent/memory
RealtimeVoice Runtime

OpenAI Realtime lifecycle, session policy, tools, and persistence.

apps/backend/agent/voice
Product surfaceWeb/API Surface

Backend API contracts and the Next.js web interface.

apps/web · apps/backend
Open source
AGPL-3.0 licensed
Self-hostable
your infrastructure
Crisis-first
safety architecture
Multi-surface
text · voice · messaging

Choose your path

Start with the docs closest to your work

OpenCouch spans safety policy, backend runtime, memory, voice, and web surfaces. Start with the path closest to what you want to inspect or change.

Trust model

Concrete safety and privacy mechanisms

Each principle is backed by a visible implementation artifact instead of a vague promise.

Safety gate before memory

Crisis assessment runs before memory retrieval, so unsafe turns do not load contextual memory first.

crisis_gate → turn_dispatch → load_memoryUser-controlled memory

Memory is documented as inspectable and controllable through explicit user commands and privacy controls.

semantic · episodic · proceduralHonest boundaries

OpenCouch provides support without pretending to be therapy.

onboarding · prompts · crisis responsesSelf-hostable infrastructure

Backend, web, docs, and persistence modes can be run locally or on your own infrastructure.

backend · web · docsAGPL-3.0 licensed

Deployment changes must remain visible, reducing the chance that safety mechanisms are stripped in closed forks.

open source safety preservation

Run locally

Start the stack from your checkout

Use the project virtualenv for backend commands and run each surface from its app directory.

Backend APIcd apps/backend && .venv/bin/python -m uvicorn main:app --reload --host 127.0.0.1 --port 8000
Web appcd apps/web && pnpm dev
Docs sitecd apps/docs && pnpm start
Full quickstart →

Source map

Jump from docs concept to code

The docs map directly to the backend, agent, voice, web, and evaluation areas of the repository.

Backend APIapps/backendAgent runtimeapps/backend/agentVoice runtimeapps/backend/agent/voiceWeb appapps/webDocs siteapps/docsEvaluationeval

Core principles

Built to be trustworthy, not just capable

Most AI products are built for engagement. OpenCouch is built for trust — with clear safety boundaries, user-controlled memory, and an architecture that can run entirely on your own infrastructure.

Private by default

No data sold. No third-party training. Self-hostable so data never leaves your infrastructure if you choose.

01
Safe by design

Crisis detection runs as a hard gate before every response — deterministic patterns plus an LLM classifier, always.

02
Honest boundaries

OpenCouch is not a therapist. It says so clearly, in onboarding, in prompts, and in every crisis response.

03
Where you already are

Web chat and OpenAI Realtime voice are the dogfood surfaces today. Additional messaging channels are on the roadmap.

04
AGPL-3.0 licensed

Not MIT — intentionally. AGPL ensures that anyone who deploys a modified version must publish their changes. Safety features like the crisis gate cannot be silently stripped in a closed fork.

05