Plug the open knowledge-coding toolchain into your chat LLM. Exposes KCF semantic modeling and code generation tools over the Model Context Protocol so an LLM can build a validated, coverage-assessed, source-traceable semantic model of a domain and generate code from it.
KCF MCP server demonstrates solid definition quality with 14 well-named, verb-driven tools (example_model, authoring_reference, compile, assess, coverage, etc.). All tools have descriptions (avg ~150 chars, within 10-1024 baseline). Input schemas are present and typed for all tools. However, output schemas are not explicitly documented in the visible code, responses are described narratively ('returns {ok, source}') rather than as formal JSON Schema. Tool descriptions are good but could be more prescriptive about when to call each tool vs alternatives. Parameter descriptions are present but sometimes terse (e.g., 'profile' in scaffold lacks enum values). Error handling guidance is minimal, tools return {ok, ...} but recovery paths are not documented.
The readiness verdict: `valid`, `ready`, coverage gap counts. Your primary "is the model good enough?" check.
Return a compact reference for writing `.kcf` model text — the model skeleton, the concept / relationship / action vocabulary, and what `valid` vs `ready` means. READ THIS FIRST, before drafting, so the model you write compiles. No inputs. Returns `{ok, reference}` (Markdown).
A self-describing manifest of everything this server does: the full pipeline, tools, coverage model, and provenance vocabulary in one call. Start here.
Assembles the system + user prompt that generates the app. Run the returned prompts yourself.
`.kcf` -> semantic IR + diagnostics + a `valid` flag.
Apply the decisions; returns the governed IR. Confirmed fills become fact; rejected are removed.
Output schemas not formally documented. Tools describe return values narratively ('{ok, source}') rather than as JSON Schema. LLMs cannot reliably parse or chain results without explicit output type definitions.
Parameter 'profile' in scaffold tool lacks enum constraint. Description says 'One of the six domain presets' but does not list valid values. LLMs will hallucinate invalid profile names.
Error handling guidance missing. Tools return {ok: true/false} but do not explain what to do on failure, whether errors are retryable, or how to recover. Agents cannot self-correct.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
The itemized gap to-do list (missing obligations per construct; add `by_concept=true` for a per-concept view).
HOW gaps are derived from constructs (the reference), and the provenance vocabulary for tagging synthetic fills.
The process for interviewing the user and drafting a model dimension by dimension (plus any house conventions).
Return a small, ready-to-run sample `.kcf` model. Call it to learn the syntax by example and to copy a working skeleton before drafting the user's domain. No inputs. Returns `{ok, source}` where `source` is the `.kcf` text.
The backend/frontend stacks you can generate for.
THE agent driver: the single best next step + whether the model is ready to generate. Call it after every edit to loop autonomously.
Tier the synthetic (LLM-proposed, tagged) fills into a `bulk` chunk (offer for one-click MASS approval - fast) and a `review` chunk (decide INDIVIDUALLY - rigorous).
A starting brief for one of six domain presets. Takes a profile name and returns a starter template.
Tool composition clarity: 'next_action' is described as 'THE agent driver' but its relationship to compile/assess/coverage is not explicit. LLMs may not understand when to call next_action vs assess directly.
Parameter descriptions for complex objects (model_ir in review_queue, confirm_synthetic) lack structure hints. LLMs do not know what fields model_ir contains or how to construct confirm/reject arrays.