The DNA Cloud starter catalog & BYO-DNA

DNA Cloud ships two things a customer consumes over MCP:

1. a starter catalog — a few generic, ready-to-use agents in the shared _lib scope that every tenant inherits (the on-ramp), and
2. BYO-DNA — each customer authors their own agents as a tenant overlay on top of that base (the product, and the moat).

Both are the same mechanism you already know: _lib inheritance for the catalog, tenant overlays for the customer's DNA, and the PricingPlan caps for plan gating. There is zero new architecture here — see adr-dna-cloud-content.

The three starter agents

They live in examples/dna-cloud/.dna/_lib/agents/ — each a real, composable agent (a Soul + baseline Guardrails + a focused instruction), inheritable by every tenant:

Agent	What it is	Composes
assistant	A clean, general-purpose helpful assistant — the solid default. The "hello world" of the catalog.	soul helpful-assistant + guardrail baseline-safety
code-reviewer	Reviews code and diffs for correctness, security, and clarity — structured findings, no rubber-stamping.	soul senior-engineer + guardrails baseline-safety, review-integrity + skill structured-code-review
dna-copilot	Coaches the user to author their own DNA — the on-ramp from using the catalog to building it.	soul dna-mentor + guardrail baseline-safety

None of them pins a model: the model falls through to the consuming scope's Genome.default_llm, so a tenant runs the catalog on their own model without editing the agent.

Preview any of them — the persona and guardrails are composed in by build_prompt, not pasted:

from dna.kernel import Kernel
mi = Kernel.quick("_lib", base_dir="examples/dna-cloud/.dna")
print(mi.build_prompt("code-reviewer"))

Multiple personas in one scope

The shared _lib catalog packs three different Souls into one scope. The built-in layout: persona-first reads a single top-level {{{soul_content}}} that flattens last-wins across every Soul in scope, so with more than one Soul the personas would collide. The catalog agents therefore pin their Soul with a small section-based promptTemplate ({{#soulspec-soul}}…), which resolves the agent's declared soul. When you author in your own scope with a single persona, just use layout: persona-first — it is the cleaner form and composes identically. (Tightening the flatten to honour the declared soul is tracked as a follow-up.)

BYO — author your own agent and push it live

Your DNA is your tenant overlay: you author an agent and push it to the hosted source, and the MCP server serves your version — author once, any runtime, no deploy. An overlay of a catalog agent wins for your tenant while every other tenant still sees the base.

# 1. Scaffold your own agent in your scope (dna-copilot will walk you through this)
$ dna new agent assistant --scope my-scope --soul my-brand-voice --layout persona-first

# 2. Author it — fill the AGENT.md instruction + wire your Soul / Guardrails / Skills.

# 3. Preview the composed prompt before you ship
$ dna doc show Agent assistant --scope my-scope

# 4. Push it to the hosted DNA source as your tenant overlay
$ DNA_TENANT=my-tenant dna doc apply agents/assistant/AGENT.md \
      --scope my-scope --source "$DNA_SOURCE_URL"

Now the MCP server resolves assistant for my-tenant to your overlay; another tenant with no overlay keeps reading the base catalog agent. That isolation is proved by packages/sdk-py/tests/test_dna_cloud_catalog_overlay.py.

Gated by your plan

Authoring + emit is a Pro capability; Free reads the base catalog. The caps are declared on the PricingPlan Kind (examples/dna-cloud/.dna/_lib/tiers/):

Plan	feature_families	Catalog	BYO overlay
Free	definitions, sdlc	read the base	—
Pro	+ memory, emit	read the base	author own + emit

If you hit a cap, upgrade rather than working around it — dna-copilot will point you there.

Seeding the catalog into the hosted source

These _lib agents are exactly what the hosted deployment seeds into its DNA source (the same "Phase 2" seed step as the demo scope in Hosting the MCP server on Azure):

# push the whole _lib catalog to the hosted Postgres source
$ dna doc apply examples/dna-cloud/.dna/_lib --source "$DNA_SOURCE_URL"
# …or, on the ACA deployment, the mounted-source helper:
$ ./scripts/push-scope.sh examples/dna-cloud/.dna

After the seed, every tenant inherits the three agents immediately, and Pro tenants can start overlaying their own.