SShortSingh.
Back to feed

BigCommerce's AI store server gives agents 16 tools with no per-action decision scores

0
·6 views

BigCommerce has opened its Store Management MCP server in early access to all stores, letting AI agents such as Claude, ChatGPT, Cursor and Grok work with catalogs, pricing, inventory, orders, promotions and profiles. Its documentation lists 16 tools, of which 9 are mutating writes, and a refund quote tool that moves no money. Only two writes are staged; create_refund, price list replacement and absolute inventory sets are not, and the sole named guard is a per-user area scope. Per-action decision scores are named zero times in that documentation. A separate test by ScriptMasterLabs, run with an uncalibrated local heuristic harness, auto-acted on catalog searches while holding every refund shape at 0.50 advisory, and the firm itself notes the harness's limits.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Dropping the Enter Key: A Developer's Lesson from 58 Players of a 60-Second Math Game

A developer built a daily 60-second mental math game in 2022 that drew little interest and was shelved for four years. Revisiting it, they concluded that friction, not the math itself, was the main reason players disengaged. They rebuilt the game as a single HTML file with no login and a 60-second clock. Every answer is now a single digit that submits on tap, removing the Enter key so players can keep a rhythm. Fast tapping exposed UI delays, which led to a "ghost cell" feedback layer that shows correctness without pausing input.

0
ProgrammingDEV Community ·

Canada-wide dataset maps 4,030 licensed cannabis stores from provincial registries

No single national list of licensed cannabis retailers exists in Canada, since retail is regulated provincially and territorially. A team compiled thirteen separate provincial and territorial registries into one open dataset of aggregate store counts. As of November 20, 2026, it recorded 4,030 licensed, open stores across 1,027 cities, with Ontario accounting for 1,867 of them. The compilers note that registries differ in how they define a store, with Quebec counting only government-run outlets while some other provinces mix government and private retailers, so comparisons are imperfect. Most cities have very few stores, with the median city having just one.

0
ProgrammingDEV Community ·

Webhook Signature Failures After Deploy Traced to Raw Body Handling and Admission Controls

A marketplace webhook signature check can start failing after a deploy, leaving platform events unable to reach the backend. The article attributes this in Express to JSON middleware consuming the raw body before verification, since re-serializing parsed data alters the signed bytes. It recommends capturing the original bytes before parsing, checking the registered signing secret for rotation issues, and rejecting failed verifications with a non-retryable status. The author frames this as a proposed incident drill rather than a documented outage, emphasizing a trade-off between spending limits and refused legitimate traffic. The piece also suggests tracking acceptance and processing as separate objectives, tying verification errors to registration IDs, and avoiding urgent pages from a single invalid request.

0
ProgrammingDEV Community ·

karasu weekly: ADR search experiment, self-modeled architecture, and new entity-view layout

The seventh weekly update from karasu, a text modeling language for system architecture inspired by the C4 model, reports 61 merged PRs and seven ADR decisions. A 150-run experiment compared three setups and found that a skill for finding ADRs did not improve correctness, which was already near ceiling, though it cut cost by about 20%. The same single-sentence CLAUDE.md instruction achieved most of the savings at about 14%, so the skill was moved to a separate tool, adr-tools. The team also reverse-engineered the karasu repository itself into a 4,453-line model, showing the language can describe non-web systems, and added storage IDs that let use cases reference specific resources. Finally, the entity view now groups related out-of-domain entities into per-domain frames, with layout rules tested on three large real-world models.