Read-only, enforced by SQLite
The database opens with mode=ro and query_only, so even the raw SQL tool can't write. SQLite enforces it, not a check on the query text.
Featured build · Go systems case study
Livepricewatch
A Go CLI that prices an 8,800-card Pokémon collection on a free budget of 1,000 API requests a day, and never guesses which print a card is.
$ pricewatch run --budget 100 --no-wait run 3 (pokewallet) 38 requests, 52 cards ok 50 failed 2 deferred 0 4886 cards not due yet changed since each card's last price BEFORE NOW DELTA % 50.47 60.00 +9.53 +18.9 ex ruby & sapphire 59/109 reverse holo progress saved · next run resumes here
- Role
- Sole developer
- Status
- Live · runs hourly on GitHub Actions
- Outcome
- ~915 requests a day to keep 4,900 prices current
- Stack
- Go, SQLite, GitHub Actions, PokéWallet API, MCP
Why I built it
Learning Go on a problem with real limits
After nine years of C# backend work, I wanted to learn Go on something that would push back. My own collection, 8,800 rows exported from TCG Collector, gave me a real product with a hard constraint. Where two designs worked, I picked the one that exercised more Go.
The problem
8,800 cards, 1,000 requests a day
The free price API allows 100 requests an hour and 1,000 a day. The collection can't be priced in one go, so pricewatch never tries. Each run is a bounded slice of work, and progress lives in SQLite so the next run carries on where the last one stopped.
How a run works
Bounded slices, durable progress
Import resolves the collection once. Every hour after that, a run picks the cards that are due, prices them within the rate limits, saves each observation and republishes the status page.
Step 1
Resolve
Match each CSV row to one card by region, set, number, variant and language.
Step 2
Pick due cards
Never priced first, then most overdue, then most valuable.
Step 3
Pace
Spend against the API's own remaining count, pause at the hour limit.
Step 4
Price
A bounded worker pool. One request prices every variant of a card.
Step 5
Save + report
One observation per card, plus every card that moved.
Step 6
Publish
Rebuild the public status page from derived numbers only.
Hourly loop ↶
GitHub Actions runs at seven past each hour. A missed hour just means the next run has a little more to do.
Value tiers
Spend requests where the money is
A card's value sets how often it's due. Pricier cards move more in dollars, so they're checked more often.
| Card value | Checked every |
|---|---|
| $100+ | 1 day |
| $20 to $100 | 2 days |
| $5 to $20 | 4 days |
| Under $5 | 7 days |
Daily budget
~915 / 1,000
requests a day to keep about 4,900 priceable cards on schedule.
01,000
What it refuses to guess
A wrong match fails silently
The export has no card IDs. When a row can't be matched to exactly one card, pricewatch reports it instead of picking one. On chase cards, the gap between prints dwarfs any price movement.
Tropius 001/084 · Normal
$0.06
Tropius 001/084 · Reverse Holo
$0.20
Ask it anything
Questions in plain English
pricewatch mcp is a local MCP server. Claude starts it, and seven read-only tools answer questions from the same database the hourly job writes.
Which Pokémon is worth the most across every card I own?
pricewatch · queryPikachu, by a long way. 52 cards worth $2,849.93, and one of them, Pikachu with Grey Felt Hat, is over a third of that at $1,015.89.
Pokémon Cards Value Pikachu 52 $2,849.93 Sylveon 21 $2,361.85 Umbreon 8 $1,553.93 Is TCG Collector overvaluing my chase cards?
pricewatch · queryMostly no. Nearly every card worth $100 or more is within a few dollars of market. Four are listed more than $50 high.
Card Listed Market Paradise Resort (Staff) $641.76 $433.51 Umbreon ex $1,410.28 $1,326.66 Pikachu with Grey Felt Hat $1,081.11 $1,015.89 Sylveon ex $533.18 $476.49 data_as_of Oct 5, 2026
Every answer is dated
Results carry data_as_of. If a fresh copy can't be downloaded, the server answers from its cache and says the data is stale.
Prints stay separate
find_cards returns one row per print, and price history takes one exact print. When a name matches several, Claude sees them all and has to pick or ask.
Engineering decisions
Safe to stop, safe to re-run
Two-stage Ctrl-C
The first press stops new cards and lets in-flight ones finish. The second cancels them. Completed work is saved either way, using context.WithoutCancel so the checkpoint outlives the cancel.
Pace against the source's count
The daily count lives in SQLite and is corrected from response headers, so a restart can't overspend. A “too many requests” answer is read, waited out, and the run carries on.
Standard library first
Two third-party packages: pure-Go SQLite and a rate limiter. CSV, HTTP, JSON, logging, embedded templates and fake-clock tests all come from the standard library.
Coming from C#
What surprised me after nine years of .NET
Errors as values
if err != nilthrow / catch
Repetitive, but failure is in the signature and every caller has to acknowledge it. I wouldn't swap back, though I'd take a shorter syntax.
Context vs CancellationToken
context.ContextCancellationToken
Cancellation and deadline travel together, which made the three lifetimes in shutdown easy to keep apart.
Goroutines and select
go + selectasync / Task.WhenAny
No async coloring up the call chain. One select reads as “send the job, or stop.”
Interfaces declared by the consumer
implicit: IStore + DI
Great at this size. For a bigger domain I'd still want C#'s explicit contracts and a container.
Current state
Phases 1 to 3 complete, plus an MCP server
Built
- Collection import with reported, never guessed, matches
- Bounded, resumable pricing runs with a worker pool
- Value tiers and due-card scheduling
- Hourly GitHub Actions runs against a private data repo
- Public status page rebuilt every run
- Offline tests with a fake clock for the worker pool
- Special prints priced from their own products: ball, Energy and Rocket reverse holos, and Prerelease, Staff, Worlds and 30th Anniversary promos
- MCP server so Claude can answer questions about the collection
What’s next
Match the cards it can't price yet
The pipeline is done. The remaining work is matching, so more of the collection gets a price without guessing.
Planned next
- Prints still reported as unsupported: Cosmos, Prize Pack, set stamps and some promos (about 3% of the collection)
- Rows the import reports as unmatched or ambiguous
Explore the work
Built to run unattended
Next in the set
Featured build
Live
Job Hunt Pipeline
A daily AI job search that screens LinkedIn roles against my rules and resume.
- C#
- .NET 10
- Claude Code
Case study 5 of 5. Up next: Job Hunt Pipeline.
Featured build
In Development
PokéJudge AI
An AI rules assistant that asks the right questions before it recommends a cited ruling.
- C#
- .NET
- RAG / Retrieval