Skip to content
Switchplinth

Up next

Works today

Section titled “Works today”
  • switchplinth setup sets up a repository or the sample, runs the first scan, and registers the MCP server with Claude Code, Claude Desktop, Cursor, or VS Code.
  • ./scripts/demo.sh runs the food-delivery sample in Docker with one command.
  • Markdown discovery with no template; canonical, portable .decision/ YAML; SQLite and Qdrant indexes that rebuild from those files.
  • Deterministic mock extraction, and model extraction through local OpenAI-compatible servers. A local Qwen3.8 model with reasoning off found all 7 decisions in the sample.
  • Confirm, reject, and correct for claims, aliases, and identity links; the Facts page shows the review log with Undo on each entry.
  • An architecture map with layered layout and fullscreen mode, plus What-if previews that switch decision options in place.
  • An Architecture log with each decision’s rationale, why alternatives were not chosen, and consequences, linked to source passages.
  • Grounded “why” answers with cited passages.
  • Scenario diffs between any two selection maps, with relationship-based downstream impact, over the API and MCP.
  • Input receipts on graph, scenario, search, and diff results, for checking whether the inputs changed.
  • Capability checks: a component can require a capability such as event replay, and What-if, the API, and MCP report when a switch leaves no provider reachable within four relationship hops. Switching event delivery from Kafka to SQS in the sample reports that the Order Service needs replay and SQS lacks it. The check uses extracted requires_capability, provides_capability, and lacks_capability claims; it does not analyze code or runtime behavior.
  • Five read-only MCP tools over streamable HTTP (loopback) and stdio.

Up next

Section titled “Up next”

These are the next steps, in no particular order and without dates.

Next step Where it stands today
Test hosted providers (OpenAI, Anthropic) Adapters exist; no hosted credentials or diagnostics have been exercised.
Measure extraction recall on more repositories Local-model results cover small fixtures and the bundled sample only. Claim-level repeatability is unmeasured.
Extraction beyond the sample The one-command demo uses mock rules tuned to the sample’s text; it shows the decision and what-if model, not extraction quality on other documents.
Authentication for remote API and MCP access Both are unauthenticated and bind to loopback; keep them local.
More source formats Markdown only; no code, PDFs, or other formats.
Conflict detection beyond same-ID writes The compiler reports a conflict only when two selected options write the same entity or relationship ID.
Scenario rename, update, and delete Saved scenarios support create, list, and load.

Other current boundaries are documented where they apply: the embedding fallback in Model providers, and write safety and schema versioning in Portable format.