Example Projects¶
Five complete, runnable projects live in the public
tuvl-io/examples repository. Each was
built from a project-specification.md in its directory, passes
tuvl validate with zero warnings, and runs with tuvl dev — clone one, point
it at your Postgres, and make it yours.
| Project | Difficulty | What it demonstrates |
|---|---|---|
invoice-extraction-api |
Easy | Raw invoice text → validated Postgres records: one Agent step with typed JSON output, a deterministic totals check, enum currencies, a secure tax id |
knowledge-base-qa |
Easy–Medium | RAG on the built-in rails — DataIngest/DataSearch over pgvector, cited answers, zero custom Python, a @tuvl/client demo script |
content-moderation-pipeline |
Medium | LLM classification, deterministic match: region routing, and group-gated human review — the submitter cannot approve their own content |
mcp-research-agent |
Medium–Complex | An AutonomousAgent driving an MCP fetch tool to a cited research brief, bounded by a token budget, with live loop progress over the SDK |
kyc-onboarding |
Complex | A supervised autonomous investigator (fail-closed LLM judge), policy RAG, compliance-gated approval, PII masking, versioned risk schemas |
Together the five exercise every step kind and subsystem: all nine step kinds,
RAG, spec.supervisor, HumanInTheLoop with auth.required_group, MCP over
stdio, model versioning, secure: true masking, IAM scopes and groups, and the
tuvl test framework with LLM-judge evaluations.
Running one¶
uv tool install "tuvl[standard]>=2026.2.6.1"
git clone https://github.com/tuvl-io/examples.git
cd examples/<project-name>
cp .env.example .env # fill in DATABASE_URL, GEMINI_API_KEY, …
tuvl validate
tuvl dev # → http://localhost:8000 (+ /insight)
Per-project infrastructure needs (database, pgvector, API keys, MCP tooling)
are catalogued in the repo's
REQUIREMENTS.md.
Contributions follow
AGENTS.md —
each project ships a specification a human or coding agent can implement
unaided.