05 — ProviderCatalogCli (bonus)

A non-interactive SQLite CRUD CLI for an "LLM provider catalog" — the table an agent reads to know which model endpoints it can talk to.

What it shows#

A provider catalog is just a SQLite table plus a thin CLI. Every other piece of the agent (resolver, dispatcher, health probe) just queries this table, so "agent configuration" really does collapse to CRUD over a handful of rows. This demo is the skeleton of OpenClawNet's ModelProviderDefinitionStore with the production complexity stripped away — same shape, same primary-key-on-name, same single-default invariant, none of the orchestration. Contrast with demos 02 and 03: those were REPLs because the lesson was interaction state; this one is one-shot subcommands because the lesson is configuration storage.

Prerequisites#

.NET 10 SDK
(No Ollama, no network — pure local SQLite.)

Run#

dotnet run -- seed
dotnet run -- list
dotnet run -- show local-llama
dotnet run -- add my-claude --kind anthropic --endpoint https://api.anthropic.com --model claude-3-5-sonnet --key-env ANTHROPIC_API_KEY
dotnet run -- update local-llama --model llama3.2
dotnet run -- set-default openai-gpt4
dotnet run -- remove my-claude

The DB lives next to the built executable as providers.db. Override the location with the PROVIDERS_DB environment variable:

$env:PROVIDERS_DB = "C:\temp\my-providers.db"
dotnet run -- list

Sample output (example)#

After seed, then list:

NAME         KIND          MODEL         ENDPOINT                                DEFAULT
-----------  ------------  ------------  --------------------------------------  -------
azure-prod   azure-openai  gpt-4o        https://my-resource.openai.azure.com
local-llama  ollama        llama3.2   http://localhost:11434                  ✓
openai-gpt4  openai        gpt-4o-mini   https://api.openai.com/v1

show local-llama:

name:        local-llama
kind:        ollama
endpoint:    http://localhost:11434
model:       llama3.2
api_key_env: (none)
is_default:  yes
created_at:  2026-04-27T14:02:11.4421137Z
updated_at:  2026-04-27T14:02:11.4421137Z

Asking for something that isn't there:

> dotnet run -- show nope
❌ no such provider: nope
(exit code 1)

Commands#

Command	Description
`list`	Print all providers as a table, sorted by name.
`show <name>`	Print every column of one provider, one per line.
`add <name> --kind … --endpoint … --model … [--key-env …]`	Insert a new provider; errors if name exists.
`update <name> [--kind …] [--endpoint …] [--model …] [--key-env …]`	Partial UPDATE; only supplied fields change.
`remove <name>`	DELETE a provider; warns if it was the default.
`set-default <name>`	Clear all defaults, set this one (in a transaction).
`seed`	Insert 3 sample providers if the table is empty.
`help` / `-h` / `--help`	Print usage with one example per command.

Exit codes: 0 success, 1 user error (missing flag, name not found, duplicate), 2 unknown subcommand.

How it works#

One table, providers, primary-keyed on name — so duplicates can't silently appear, and lookups are trivial.
Single-default invariant enforced in code: set-default runs UPDATE … SET is_default = 0 then UPDATE … SET is_default = 1 for the chosen row, both inside a SQLite transaction so the table is never in a two-defaults state.
Partial updates: update builds the SET clause dynamically from the flags that were actually supplied, so --model llama3.2 won't clobber endpoint or kind.
Manual flag parser (ParseFlags) — ~10 lines, no CLI library. Long-form flags only (--kind, --endpoint, --model, --key-env).
No orchestration logic anywhere — just CRUD. A real agent reads the same table, picks the row where is_default = 1 (or the one named in its config), and dials the endpoint. That's it.
Mirrors production: this is the shape OpenClawNet's ModelProviderDefinitionStore reduces to once you remove validation, multi-tenancy, and the resolver layer. Definitions table + a thin manager CLI, nothing fancier.

Try this#

Add a priority INTEGER column and let the agent fall back to the next provider if the default is unreachable.
Add a last_used_at TEXT column, stamped every time the agent picks a provider — gives you free usage telemetry.
Add a sibling provider_health table (one row per name + a status/checked_at pair) and write a check <name> subcommand that pings the endpoint and upserts the result. This is exactly how the real codebase grows: schema first, command second.