k12-lesson-toolkit-for-claude-code

Standards-grounded K-12 lesson tooling for Claude Code: an MCP standards server built from openly-licensed public data, two reference wikis, and a renderer that ships a Teacher Edition and a Student Edition a teacher can actually hand out.

"The skills are written to make effective use of the Learning Commons Knowledge Graph Claude connector that is included for all teachers using Claude for Teachers. This connector provides Claude with access to academic standards across all 50 states and progressions beneath them; adapt the skills if your environment does not have access to these tools."

— anthropics/k12-teacher-skills

This is that adaptation.

Not affiliated with, endorsed by, or supported by Anthropic. Built on their Apache-2.0 licensed k12-teacher-skills, which remains theirs. Anthropic's copyright and NOTICE are preserved wherever their code is vendored here.

---

The problem

Anthropic's two K-12 teaching skills ground their output in a Learning Commons connector that ships only inside Claude for Teachers. Outside that account the connector is absent.

The skills handle that honestly, and it is worth being precise about this rather than overselling the gap. They check whether the connector's tools are available, they state that they are fully functional without it, they forbid inventing citations, and when the connector is missing they stamp the lesson plan with a visible footer:

"Generated without the Learning Commons Knowledge Graph. Standards and misconceptions reflect general best practice."

That is good engineering. It is also a real cost. A teacher outside Claude for Teachers gets a lesson plan explicitly marked as not grounded in the standard it names — the graceful degradation, every time, forever. The disclaimer is honest precisely because the grounding is genuinely absent.

This project supplies that grounding from data anyone can ship, so the skills take their connected path instead of their fallback path. Same skills, same contract, no gated account.

What this adds, and what each addition is for

A standards server the skills already know how to use. It registers the same seven tool names the skills look for:

find_standard_statement                     find_curriculum_lessons
find_standards_progression_from_standard    find_materials_for_lesson
find_misconceptions_for_standard            list_standards_for_mathematical_practice
find_learning_components_from_standard

The skills branch on whether those names are present, so nothing here forks, patches or shims them. Use it to: install Anthropic's k12-teacher-skills exactly as published, register this server, and get lesson plans grounded in the actual standard text instead of the not-grounded footer.

The standards themselves, in the repository. 2,303 California mathematics standards as a filtered derivative of the Learning Commons public export, CC BY 4.0, each record keeping its own license and attributionStatement. No account, no API key, no gated connector. Use it to: clone and have grounding work offline, immediately.

Published curriculum aligned to those standards. 3,301 Illustrative Mathematics lessons and 12,599 activities and assessments, CC BY 4.0. Use it to: ask "what has someone already written for this standard?" and get real lesson names back, then pull the activities inside one.

Teacher Edition and Student Edition documents. The renderer emits both as editable .docx, answer keys in one and not the other, with figures carrying screen-reader alt text — upstream's block vocabulary has no image type at all, which is workable for prose subjects and useless for geometry. Use it to: hand a student the Student Edition without hand-editing anything out.

Two reference wikis, 208 pages. docs/sourcing/ answers may I use this source? per host, with the verbatim licence sentence, URL, HTTP status and date behind every verdict — because licences expire, and two this field relies on were withdrawn inside six months during 2026 while OER indexes went on publishing the old answer. It also keeps citing, quoting and adapting apart, which is where most open-education licence trouble starts: ShareAlike and NoDerivatives do not touch citation, and quoting does not trigger ShareAlike. docs/udl/ answers does this design remove a barrier, or does it change what I am measuring? Use them to: settle a sourcing question in a minute, and as standalone reading — they are the part most likely to be useful even if you never run the server.

A slice builder for other states and subjects. The export covers 52 jurisdictions and 4 subjects. Use it to: build Texas Science or New York ELA instead of California mathematics.

Getting it running

You do not need to run any of this yourself. Clone the repository, open it in Claude Code (or any coding agent), and say:

Read SETUP.md and set this up, then verify it and tell me what works.

SETUP.md is written for the agent, not for you. It carries the install steps, the expected numbers at every stage, how to register the MCP server, and — the part that matters — the three ways this system returns a confident empty answer that looks like a real one. It ends by telling the agent exactly what to report back.

If you would rather drive it by hand, SETUP.md is perfectly readable; it is just longer than you need.

What is tested, and what is not

	Status
The seven MCP tools against the shipped data	Tested. 115 automated tests, run on every change.
The shipped data reproducing from its own script	Tested. Re-extracting California mathematics reproduces data/ca-math/ record for record.
Teacher and Student .docx output	Tested. The worked example was built end to end and both editions are in examples/.
Figures on a non-macOS machine	Partly. The example was rebuilt with the macOS rasterizer disabled and matched exactly — but on macOS. Never run on Linux or Windows.
Anthropic's skills taking their connected path	Not tested end to end. The tool names and response shapes are pinned by tests/test_contract.py against the contract extracted from the skills. Nobody has watched the real skills run against this server and confirmed the footer disappears.
Any state or subject other than California mathematics	Not tested as content. The extractor is tested and builds other slices; no unit has been written from one.
Misconceptions	Empty by measurement. The upstream export contains none, anywhere. That tool returns [] and the skills fall back to their own knowledge.

Layout

Path	What
src/k12_toolkit/mcp/	the 7-tool MCP standards server
src/k12_toolkit/ingest/	builds the SQLite store from the shipped export
src/k12_toolkit/docgen/	Teacher/Student edition renderer, vendored from upstream plus a figure block
data/ca-math/	the standards export, CC BY 4.0 — see LICENSE-DATA
docs/sourcing/	reference wiki: what you may cite, quote or adapt, per host, with dates
docs/udl/	reference wiki: Universal Design for Learning as a design discipline
docs/reference/	the engine map, the export schema, the sourcing verdict, the document standard
docs/design/	why the MCP server has the shape it has
scripts/	extract_standards.py builds a slice for any state and subject; extract_curriculum.py adds the curriculum layer
tests/	the suite
SETUP.md	install and verification, written for an agent

Both wikis start at their own README.md and read as standalone references.

The worked example

The two finished documents are in examples/hs-geometry-similarity-trig/ — open those first; they are what a teacher actually receives.

hs-geometry-similarity-trig is the full source: a complete two-week grade 9-10 geometry unit built with this toolkit, CC BY 4.0, with ten lesson packages, three quizzes, a final exam, a parallel-form practice exam, 73 accessibility-validated figures, and a construct-register row for every assessable item written before the item existed.

What it cost to make

Measured from commit timestamps, not recalled:

Phase	Wall clock
Standards research, licence sweep, reference-wiki build, unit design	6h 22m
Unit build — 10 lessons, 5 instruments, 73 figures	2h 59m
Document rendering — Teacher and Student editions	1h 05m
First unit, starting from nothing	10h 26m

Output: 143 files, 285,583 words of markdown, 73 validated SVGs, two rendered editions.

A second unit is a projection, not a measurement. Only one has been built, so treat the number accordingly. It would reuse the wikis, the figure validator, the renderer, the document standard and the construct-register schema, and reuse none of the topic analysis, unit design or content. On that basis, roughly 5 hours. If that turns out wrong, the honest number will replace it here.

Building a different state or subject

California mathematics is shipped because a clone has to stay small, not because it is the only slice that works — the whole export would be roughly 280 MB. It covers 52 jurisdictions and 4 subjects, and scripts/extract_standards.py builds any pair.

It needs the Learning Commons knowledge-graph export (nodes.jsonl + relationships.jsonl, ~812 MB), which is public and not redistributed here. Ask your agent for it — "build me the Texas Science slice" — and point it at the head of scripts/extract_standards.py, which documents the one non-obvious rule: every slice also carries the Multi-State standards for its subject. That is not optional. A state's own standards hold no progression edges and reach them only through the crosswalk, so a slice built without Multi-State loads perfectly and answers every progression query with an empty list.

Re-running the script for California mathematics reproduces the shipped data/ca-math/ exactly, and --verify data/ca-math checks that — so the script is not asking to be trusted.

Honest limits, in detail

The table above is the summary. These are the five that will bite you specifically, and the first four share a shape: they return an empty list rather than an error. An empty result from this system is a boundary, not an answer about the standard you asked for.

• A standards code must be in the exact form the store holds. HSG-SRT.C.6 returns rows. G-SRT.6 and HSG-SRT.6 return zero rows and no error — the failure most likely to read as "this standard has no content" when it means "you typed it differently".
• Zero misconception records, and this one is not fixable here. Measured rather than assumed: across all 247,786 nodes of the export, no property key and no relationship label matches misconception, error or mistake. misconceptions.jsonl is 0 bytes for every slice, not just this one. Every misconception in the worked example was authored and cited by hand.
• No progressions outside CCSS mathematics. All 757 buildsTowards edges in the export run between Multi-State mathematics standards. A Science or ELA slice loads and answers every other tool, and returns nothing on progressions; the extractor says so at build time rather than letting the tool look empty later.
• Curriculum reaches a state standard by inference, not by statement. All 561 standards the curriculum aligns to are Multi-State CCSS nodes; zero are California's. A California lookup therefore travels the export's own evidence-based crosswalk to its CCSS twin, and every such result carries alignedVia naming the standard it came through. 525 of California's 1,467 standards reach curriculum this way; the other 942 reach none — most have no crosswalk edge at all. Treat an alignedVia result as a strong suggestion rather than as the publisher's own alignment.
• Licence verdicts are dated 2026-08-07 and 2026-08-08. See the note at the foot of LICENSE-DATA about why that matters more than it looks.

Three limits listed here previously have been closed and are recorded rather than erased, because in each case the reason for the limit was wrong in an instructive way: find_curriculum_lessons and find_materials_for_lesson were stubs "because the data is not in the public export" — it was, and the join failed only on an identifier-space mismatch that returns empty without raising (docs/reference/sourcing-verdict.md). list_standards_for_mathematical_practice was a stub justified by its consumer rather than by its data; MP1 to MP8 were in the shipped export all along. And the .docx figure path was macOS-only; it now runs a backend chain and was verified end to end with QuickLook disabled.

Licence

Code is Apache-2.0 (LICENSE). Data under data/ca-math/ is CC BY 4.0 and carries a separate set of obligations, including a CCSS rider that is not Creative Commons at all (LICENSE-DATA). Third-party attribution is in NOTICE.