telepace

Voice-native, Agent-first user research infrastructure.

Your Claude / Cursor / Codex can now interview 100 users while you sleep, and wake up to structured insights it can act on.

What this is

telepace is a user research platform designed for AI agents as the first-class user. Three surfaces:

1. MCP Server — Claude / Cursor / Codex call create_campaign, get_insights, etc.
2. Web App — traditional research team dashboard with chat-first study composer
3. REST API + Webhook — embed in your own agent or product

Positioning

• vs Outset / Listen Labs: Agent-first, not dashboard-first
• vs Perspective / Voicepanel: Voice-native (real-time voice interviews, not text-only)
• vs Tally / Typeform MCP: Research closed-loop (moderate + analyze + push back to agent), not form CRUD

Architecture

Ingress (Marketing / Web App / Respondent UI / MCP)
    ↓
Contract Layer (Pydantic — single source of truth)
    ↓
Harness (Orchestrator + Router + Memory + Policies + Observability)
    ↓
Agents (Designer / Interviewer / Analyst / Coordinator)
    ↓
Domain Services (voiceflow / analysis lib / channel workers)
    ↓
Event Store (Postgres append-only) + Projections + pgvector

See docs/architecture.md for the full design.

Repo layout

core/              domain models, events, protocols (contract layer)
harness/           orchestrator, router, memory, policies
agents/            designer, interviewer, analyst, coordinator
interfaces/        mcp_server, rest_api, realtime
voice/             wrapper around telepace/voiceflow (Go)
analysis/          sentiment, clustering, persona synthesis
storage/           event store + projections + vector
frontend/          single Next.js app (marketing + app + auth + respondent), shared UI
eval/              prompt evals + CI
deploy/            docker-compose, fly.io, github actions
docs/              architecture, agents, protocols, roadmap

Quickstart

One command (starts docker services → backend :8010 → frontend :3300):

scripts/up.sh          # bring the whole dev stack up
scripts/doctor.sh      # check the environment first (toolchain, ports, services)

Or step by step:

# Backend
uv sync
docker compose -f deploy/docker-compose.dev.yml up -d postgres redis
uvicorn interfaces.rest_api.main:app --reload --host 127.0.0.1 --port 8010

# MCP server
python -m interfaces.mcp_server.server

# Frontend
cd frontend && pnpm install && pnpm dev   # → http://localhost:3300

No migration step: the database schema is created automatically on backend startup (CREATE TABLE IF NOT EXISTS), so there is no alembic to run.

Behind a SOCKS proxy? If your shell exports ALL_PROXY/http_proxy, the LLM SDK crashes on startup with a missing-socksio error and local curl hangs. Strip the proxy for the backend process (or install httpx[socks]). Full walkthrough + an end-to-end register → create → respondent self-test: docs/local-fullstack-selftest.md.

License

MIT (planned).