GrantKit

The linter and compiler for grant proposals. Grants as files; agents bring the AI.

GrantKit is a stateless, local-first engine. It reads a grant.yaml plus your Markdown responses and it lints them, compiles them into one submission document, and reports a machine-readable status — with no cloud service and no AI calls of its own. Point Claude Code (or any agent) at the files to do the writing; GrantKit keeps them correct.

Why files

Grant portals trap your content in web forms where your tools can't reach it. GrantKit keeps the whole proposal as plain files in a git repo:

• Agents read the full context — every section, limit, and rule is on disk.
• Changes are reviewable diffs, not opaque form edits.
• Git gives you history and rollback.
• You bring the AI you already use — GrantKit itself calls no model.

Think of it as eslint + tsc for a grant: grantkit check is the linter, grantkit build is the compiler, and funder rule packs are the config.

Install

pip install grantkit               # core engine
pip install "grantkit[pdf]"        # + PDF output (WeasyPrint)
pip install "grantkit[docx]"       # + DOCX output (python-docx)
pip install "grantkit[mcp]"        # + MCP server for agents
pip install "grantkit[all]"        # everything

Quickstart

# 1. Scaffold a project from a funder rule pack
grantkit init --funder nuffield-rda

# 2. Write — with Claude Code, Cursor, or your editor
claude "draft responses/b_case_for_importance.md from our repo README"

# 3. Lint against the funder's rules
grantkit check

# 4. Compile the submission document (+ a shareable review page)
grantkit build --format pdf --share

The five verbs

Every verb takes an optional path to the grant directory (default .).

What check catches

• Required sections present and non-empty; word / character / page limits.
• Placeholder text left behind ([TO BE COMPLETED], TODO, lorem ipsum).
• Markdown that a plain-text portal would paste literally.
• Citations ([@key]) that don't resolve against references.bib.
• Budget arithmetic (fringe/indirect) and funder caps; optional BLS salary and GSA per-diem sanity when those API keys are set.
• Funder formatting rules from the rule pack — including the full NSF PAPPG content engine (prohibited URLs/emails, required Intellectual Merit / Broader Impacts, etc.).
• US/UK spelling for the funder's locale.
• Link liveness (--urls, opt-in — the only thing that touches the network).

Funder rule packs

A rule pack is a YAML file under grantkit/data/funders/ describing one funder: its sections and limits, formatting rules (each with a citation), budget caps, portal quirks, spelling locale, and review rubric. Three ship today:

Contributing a pack

1. Copy an existing pack in grantkit/data/funders/ and edit it. The stem of the filename is the pack id (acme-fund.yaml → acme-fund).
2. Only encode values you can source. Leave a limit null if the funder doesn't publish one — never invent a number. Add a provenance: note and comments citing where each value came from.
3. Set locale (en-US/en-GB) and portal.accepts_markdown.
4. Run python -c "from grantkit.packs import load_pack; load_pack('acme-fund')" — it validates against the schema on load — and add a case to tests/test_packs.py.

The full schema is documented in grantkit/packs/schema.py.

CI for grants

Add the composite action to lint every push and publish the review page and status.json as build artifacts:

# .github/workflows/grant.yml
name: Grant
on: [push, pull_request]
jobs:
  grantkit:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: GrantKit/grantkit@v0.2.0
        with:
          path: .          # directory containing grant.yaml
          strict: "false"  # set "true" to fail on warnings too

Errors always fail the job; the assembled.html review page and status.json are uploaded as artifacts on every run.

MCP server for agents

Expose the engine to an agent over the Model Context Protocol:

pip install "grantkit[mcp]"
grantkit-mcp          # stdio transport

Tools: grant_check(path), grant_status(path), grant_build(path, format) — each returning the same JSON structures the CLI emits.

status.json

grantkit build and grantkit status --json always write a status.json describing completion, per-section word counts, and the current check results. It is GrantKit's stable, machine-readable contract for other tools (dashboards, CRMs, agents). The exact shape is documented in docs/artifacts.md.

Development

git clone https://github.com/GrantKit/grantkit.git
cd grantkit
pip install -e ".[dev]"

ruff check . && black --check . && mypy grantkit && pytest

Issues and tasks are tracked in GitHub Issues.

License

MIT — see LICENSE. Created by PolicyEngine.