Bilig WorkPaper

bilig

Run workbook-shaped business rules inside Node.

Bilig gives services, queue workers, tests, MCP servers, and tool integrations a typed WorkPaper object: write inputs, recalculate formulas, read outputs, persist JSON, restore, and verify. It fits pricing models, quote approval, payout checks, import validation, forecasts, and formula-backed workflow steps.

Run the no-project service check from any Node project:

npm exec --yes --package @bilig/workpaper@latest -- bilig-evaluate --door workpaper-service --json

Expected WorkPaper service result:

{
  "schemaVersion": "bilig-evaluator.v1",
  "door": "workpaper-service",
  "verified": true,
  "evidence": {
    "editedCell": "Inputs!B2",
    "dependentCell": "Summary!B2",
    "before": 24000,
    "after": 38400,
    "afterRestore": 38400,
    "persistedDocumentBytes": 999
  }
}

For TypeScript services that should own the workbook model:

npm create @bilig/workpaper@latest pricing-workpaper
cd pricing-workpaper
npm install
npm run smoke

For MCP clients or other tool integrations, run the same proof loop through the MCP evaluator before adding host-specific config:

npm exec --yes --package @bilig/workpaper@latest -- bilig-evaluate --door agent-mcp --json

Evaluator examples live in examples/bilig-evaluator-proof. Integration matrix docs and host-specific config files are available when a team needs them, but the public proof starts with WorkPaper service readback. File compatibility diagnostics are a separate path for import/export boundaries, not the default runtime story.

Project site: https://proompteng.github.io/bilig/

Start Here

Pick the path that matches the job:

You have...	Start with	You should see
A Node service, route, queue, test, or tool needs workbook logic	Node service WorkPaper evaluator	input edit, recalculated output, serialized JSON, restore check, and verified: true.
An MCP client or tool integration needs workbook commands	MCP tool evaluator	tool discovery, cell edit, formula readback, export, restart check, and verified: true.
You want a starter project with the runtime installed	90-second Node quickstart	a local package smoke test that edits one input, recalculates, saves JSON, and restores the WorkPaper.
An imported file is the integration boundary	Workbook Compatibility Report	unsupported functions, external links, macros, pivots, volatile formulas, and import/export risks.

If you are not sure which one fits, start with the thing that owns state. Use WorkPaper when your service or tool should own the workbook model. Use file diagnostics only when import/export compatibility is the actual contract.

Good fits: pricing, quote approval, payout checks, import validation, forecasts, CI fixtures, formula-backed workflow steps, and tool integrations that need exact cell addresses plus readback. Bad fits: manual spreadsheet editing, Office macros, desktop Excel automation, or one-off arithmetic where a workbook would be ceremony.

If You Only Try One Thing

Run the WorkPaper service proof at the top of this README first. It is the shortest proof that Bilig gives backend code a workbook object it can change, recalculate, read back, save, and restore without driving Excel, LibreOffice, Google Sheets, or a browser grid.

If an MCP client or tool integration owns the workflow, run the MCP door:

npm exec --yes --package @bilig/workpaper@latest -- bilig-evaluate --door agent-mcp --json

Trust boundaries:

• Runs locally in Node or in your GitHub Actions runner; no hosted workbook upload is required.
• Does not claim Excel parity. Start with where Bilig is not Excel-compatible yet before using it for irreversible workflows.
• File import/export diagnostics are available when an imported file is the contract, but they are a separate path from service-owned WorkPaper state.

Which Path Should I Install?

Problem you have right now	Install or use	First proof
Formula workbook state belongs inside a Node service, route, queue, test, or tool	npm install @bilig/workpaper	Node service WorkPaper evaluator
An MCP client or tool integration needs workbook tools with computed readback	npm install @bilig/workpaper	MCP tool evaluator
Import/export compatibility is the integration boundary	compatibility diagnostics	Workbook Compatibility Report

Advanced adapters are still available when the boundary is already specific: SheetJS, ExcelJS, external workbooks, MCP/tool integrations, @bilig/workbook when a runtime needs transport-neutral plan data and command receipts, and runtime provenance.

MCP And Tool Integrations

Use the WorkPaper host handoff when a tool host needs workbook reads, writes, recalculation, JSON export, and restore proof. The first check is always the no-key MCP evaluator:

npm exec --yes --package @bilig/workpaper@latest -- bilig-evaluate --door agent-mcp --json

That evaluator starts the published WorkPaper tool server, discovers tools, edits an input, reads the dependent formula, exports JSON, restarts, restores, and returns verified: true. Use llms-install.md when a host wants one install file, and use the tool-host proof matrix when the host matters more than the package boundary.

The published package also carries AGENTS.md and SKILL.md so hosts inspecting node_modules/@bilig/workpaper can find the same proof locally. Cloned checkouts keep host-specific config indexed in agent rule chooser: CLAUDE.md, .claude/skills/bilig-workpaper/SKILL.md, .claude/commands/bilig-workpaper-proof.md, .cursor/rules/bilig-workpaper.mdc, .devin/rules/bilig-workpaper.md, .windsurf/rules/bilig-workpaper.md, .clinerules/bilig-workpaper.md, .continue/rules/bilig-workpaper.md, .zed/settings.json, opencode.jsonc, and .opencode/agents/bilig-workpaper.md. The public manifest is docs/.well-known/agent.json.

npx --yes skills@latest add https://bilig.proompteng.ai --list
npx --yes skills@latest add proompteng/bilig --skill bilig-workpaper --list

Integration Recipes After The Proof

Run one evaluator first. Then use the recipe that matches the platform boundary:

• Open WebUI WorkPaper MCP: https://proompteng.github.io/bilig/open-webui-workpaper-mcp.html.
• OpenAI Agents SDK with direct tools, stdio MCP, or MCPServerStreamableHttp: https://proompteng.github.io/bilig/openai-agents-sdk-workpaper-tool.html.
• ChatGPT Apps WorkPaper MCP: https://proompteng.github.io/bilig/chatgpt-apps-workpaper-mcp.html.
• Hugging Face Gradio MCP Space: https://proompteng.github.io/bilig/huggingface-workpaper-space.html.
• n8n self-hosted workflows can use @bilig/n8n-nodes-workpaper; Pipedream formula readback is documented at https://proompteng.github.io/bilig/pipedream-workpaper-formula-readback.html and sourced from docs/pipedream-workpaper-formula-readback.md.
• Directus Persisted Calculated Fields: https://proompteng.github.io/bilig/directus-workpaper-flow-operation.html.
• Trigger.dev Durable Formula Tasks and Dagster Formula Assets are covered in the workflow recipe docs.
• ELT validation recipes: airbyte-workpaper-validation and meltano-workpaper-utility.

Choose An Evaluation Path

If you are evaluating...	Start here	What should be true before you adopt
Node service formulas	Node service WorkPaper evaluator	A starter writes one input, recalculates, persists JSON, restores, and prints verified: true.
MCP tool contract	MCP workbook evaluator	MCP tool discovery, input edit, formula readback, persistence, and restart proof all pass.
Integration proof chooser	Tool-host proof matrix, MCP spreadsheet tool server, and Vercel AI SDK formula readback	The integration path starts with the smallest verified proof and avoids write-only or UI-only claims.
Runtime intent adapters	Workbook runtime intent API and runtime intent example	A model prepares transport-neutral plan data, strict runtime proof, command receipts, and check evidence.
Basic fit	Why use Bilig?	The problem is workbook-shaped business logic that needs API readback and persistence.
Published npm package	90-second Node quickstart	@bilig/workpaper edits one input, recalculates, persists JSON, restores, and prints verified: true.
Backend service shape	Quote approval WorkPaper API	A realistic route-style workflow returns formula readback and restoredMatchesAfter: true.
MCP clients and host integrations	WorkPaper host handbook, MCP spreadsheet tool server, Gemini CLI extension, and Claude Desktop MCPB bundle	The host installs a tool path, follows the handoff guide, then proves write/readback/persist.
Technical WorkPaper review	WorkPaper maintainer proof note	One compact page has the npm check, benchmark caveat, known limits, and open questions.
Trust and performance	npm provenance and benchmark evidence	npm shows SLSA provenance, and benchmark claims match the checked artifact.
Imported files	Workbook Compatibility Report, file formula recalculation, and ExcelJS formula recalculation	The file boundary is inspected before a service, CI job, or workflow trusts imported formulas.
Almost a fit	implementation gap discussion	Name the formula, import/export, persistence, framework, MCP, package, or benchmark gap.
Formula or import bug	formula bug clinic and submit a workbook fixture	Share one reduced public case that can become a fixture.

Reduced workbook already in hand? If the blocker is an import, formula, or persistence gap, generate the fixture report:

npm exec --package @bilig/workpaper@latest -- bilig-formula-clinic ./reduced.xlsx --cells "Summary!B7,Inputs!B2"

Handing a workbook task to an MCP client or host integration? Start with the host handoff guide before opening Excel, LibreOffice, Google Sheets, or a screenshot UI. That section keeps the host handoff prompt for clients that require copy-paste instructions. To prove the package-owned MCP loop without cloning the repo:

npm exec --yes --package @bilig/workpaper@latest -- bilig-evaluate --door workpaper-service --json
npm exec --yes --package @bilig/workpaper@latest -- bilig-evaluate --door agent-mcp --json
npm exec --package @bilig/workpaper@latest -- bilig-mcp-challenge --json

Tool hosts that support skill manifests can start from skill.md or the well-known index at docs/.well-known/agent-skills/index.json. Claude Desktop users can install the released MCPB bundle directly: https://github.com/proompteng/bilig/releases/latest/download/bilig-workpaper.mcpb. For host-specific project files, use the agent rule chooser.

Try It In 90 Seconds

This uses the published npm package. It builds a workbook, changes one input, reads the calculated value, saves JSON, restores the workbook, and prints the same value again.

npm create @bilig/workpaper@latest pricing-workpaper
cd pricing-workpaper
npm install
npm run smoke

Expected output includes these fields:

{
  "before": {
    "summary": {
      "decision": "review"
    },
    "inputCells": {
      "units": "Inputs!B2",
      "listPrice": "Inputs!B3"
    }
  },
  "edit": {
    "before": {
      "decision": "review"
    },
    "after": {
      "decision": "approved"
    },
    "restored": {
      "decision": "approved"
    },
    "checks": {
      "decisionChanged": true,
      "formulasPersisted": true,
      "restoredMatchesAfter": true,
      "serializedBytes": 1242
    }
  },
  "verified": true
}

The generated starter uses the same WorkPaper fields as the public mirror at https://proompteng.github.io/bilig/npm-eval.ts and examples/headless-workpaper/npm-eval.ts. The exact byte count can change between package versions; verified: true, decisionChanged, formulasPersisted, and restoredMatchesAfter are the checks.

For a route-shaped quote approval API today, run the maintained example:

git clone --depth 1 https://github.com/proompteng/bilig.git
cd bilig
pnpm --dir examples/serverless-workpaper-api install --ignore-workspace
pnpm --dir examples/serverless-workpaper-api run smoke

For a generated project from a blank directory, run npm create @bilig/workpaper@latest pricing-workpaper through the @bilig/create-workpaper package. The package source lives in packages/create-workpaper, and the publish gate is documented in create a Bilig WorkPaper starter. For an MCP-enabled project with host integration files, MCP client configs, and an agent:verify script, run npm create @bilig/workpaper@latest pricing-agent -- --agent. Representative host files include AGENTS.md, CLAUDE.md, GEMINI.md, .claude/skills/bilig-workpaper/SKILL.md, .cursor/rules/bilig-workpaper.mdc, .trae/mcp.json, and .zed/settings.json. For an existing repo, run npm create @bilig/workpaper@latest . -- --add-agent; it adds Bilig MCP and host instructions without replacing your app template or editing package.json. If a host policy already exists, it writes BILIG_WORKPAPER_INSTALL.md with the skipped paths and a short handoff block.

If that proof almost matches a service or integration workflow you maintain, the useful next step is a concrete gap report in Discussions: formula coverage, service persistence, MCP setup, agent writeback, import/export boundary, or benchmark coverage.

TypeScript API Shape

Most integrations are just this: build a workbook, write an input, read the calculated value, and save the workbook state. When a workflow writes more than one input, use editManyAndReadback() so the edits are applied atomically and the proof compares typed readback values, persisted restore output, and formula diagnostics.

import { buildA1WorkPaper } from '@bilig/workpaper'

const book = buildA1WorkPaper({
  Inputs: [
    ['Metric', 'Value'],
    ['Customers', 20],
    ['Average revenue', 1200],
  ],
  Summary: [
    ['Metric', 'Value'],
    ['Revenue', '=Inputs!B2*Inputs!B3'],
  ],
})

const proof = book.editAndReadback('Inputs!B2', 32, {
  readbackRange: 'Summary!B2',
})

console.log({
  editedCell: proof.editedCell,
  revenue: proof.afterReadback.displayValues[0]?.[0],
  persistedDocumentBytes: proof.persistedDocumentBytes,
  verified: proof.verified,
})

book.dispose()

The lower-level WorkPaper runtime is still exported for engine integrations, but the A1 facade is the default service and agent path. Use book.set(), book.setMany(), book.readMany(), book.display(), and book.saveJson() when a full readback proof is not needed. Use book.editManyAndReadback() when several inputs should be committed and proven as one atomic workbook edit.

When To Reach For It

Use @bilig/workpaper when:

• a Node service owns a workbook-shaped calculation;
• an agent needs tools such as readRange and setInputCell, with computed before/after values instead of screenshots;
• tests need deterministic spreadsheet state and formula readback;
• a workflow needs to save the edited workbook as JSON and restore it later.

Use something else when you need a visual spreadsheet grid, Office macros, desktop Excel automation, or a one-off arithmetic helper. Do not treat embedded XLSX stored formula results as truth; use the Excel oracle workflow when accuracy matters.

Package Boundary

Current checked npm metadata for @bilig/workpaper@latest:

• Published package: 57.7 kB unpacked, 49 package entries.
• Boundary: the public package owns WorkPaper starters, evaluators, MCP command wrappers, formula clinic reports, JSON persistence, and restored readback.
• Runtime: Node >=22.0.0; Node 22 compatibility is covered by the runtime package workflow.

Published Package Trust

@bilig/workpaper is published with npm registry signatures and SLSA provenance attestations. Verify the package version you are about to adopt:

npm view @bilig/workpaper version dist.attestations dist.signatures --json

After installing, npm can verify the current dependency tree:

npm audit signatures

The current package trust path is documented in npm provenance and package trust. Repository security posture is tracked by OpenSSF Scorecard and uploaded to GitHub code scanning on every main update.

Deeper Evaluation Paths

After the first proof in Start Here, use the deeper guide that matches the next job.

1. Run the 90-second npm eval in a blank project.
2. Run the flagship serverless WorkPaper API example: npm run quote-approval-api.
3. If the workflow starts with a saved workbook file, run the XLSX formula recalculation in Node: npm start.
4. If a tool host needs workbook tools, start with the headless WorkPaper host handbook, then use the MCP server guide when the caller is an MCP client.
5. If a real workbook almost works, start with the formula bug clinic. Include the exact cell, expected value, actual value, and command output. If the fixture is already reduced, submit the structured fixture form so the blocker can become a test, example, or corpus case instead of private feedback. https://github.com/proompteng/bilig/issues/new?template=workbook_fixture.yml. If you are still reducing the case, discuss the shape first: https://github.com/proompteng/bilig/discussions/414.

The rest of the docs are an index, not a prerequisite.

For comparison and integration details, use the plain-language fit guide, screenshot automation boundary, Google Sheets API boundary, Google Sheets QUERY/SORTN in Node, workbook automation examples, the formula workbooks proof page, the Node spreadsheet formula engine guide, server-side spreadsheet automation, framework adapters, formula bug clinic, workbook fixture submissions, OpenAI Agents SDK tools, Browser Use formula tool, Google ADK MCP tools, OpenHands MCP setup, OpenCode MCP setup, Microsoft Agent Framework MCP tools, Goose MCP recipe, tool-host proof matrix, MCP spreadsheet formula server for tool hosts, Vercel AI SDK formula readback, AI SDK and LangChain tools, CrewAI adapter, the WorkPaper host handbook, the MCP server guide, spreadsheet MCP server comparison, MCP directory status, MCP client setup, Gemini CLI extension, FastMCP Python client, Claude Desktop MCPB bundle, npm provenance and package trust, JavaScript library comparison, Node spreadsheet formula engine guide, server-side spreadsheet automation, saved-workbook formula recalculation, XLSX formula support answers, SheetJS/ExcelJS boundary, ExcelJS formula result boundary, Microsoft Graph Excel boundary, and engine comparison.

Useful deeper examples: invoice totals, budget variance alerts, fulfillment capacity plan, quote approval threshold, subscription MRR forecast, agent framework adapters, MCP tool server shape, XLSX formula recalculation in Node, and serverless quote approval. Run npm run quote-approval-api, npm run agent:openai-agents-sdk, npm run agent:framework-adapters, npm run agent:mcp-tools, npm run agent:mcp-transcript, npm run agent:mcp-file-transcript, npm run agent:mcp-stdio, or npm exec --package @bilig/workpaper -- bilig-workpaper-mcp when that is the path you are evaluating.

Saved workbook diagnostics stay available when a file is the integration boundary:

npm exec --yes --package @bilig/xlsx-formula-recalc@latest -- bilig-evaluate --door workbook-compatibility --json
npm exec --yes --package @bilig/xlsx-formula-recalc@latest -- workbook-compatibility-report workbook.xlsx --json
npm exec --package @bilig/xlsx-formula-recalc@latest -- xlsx-recalc --demo --json
npm exec --package @bilig/sheetjs-formula-recalc@latest -- sheetjs-recalc --demo --json

The serverless example also includes npm run next-route-handler, npm run next-server-action, npm run next-server-action-formdata, npm run hono-route, npm run framework-adapters, and npm run persistence-adapters for framework-specific boundary checks.

The MCP server is also listed in the official registry: https://registry.modelcontextprotocol.io/v0.1/servers?search=io.github.proompteng%2Fbilig-workpaper. Clients that support Streamable HTTP MCP can also smoke-test the stateless hosted endpoint at https://bilig.proompteng.ai/mcp; use the local stdio server when the agent needs to persist a project WorkPaper JSON file.

Examples You Can Run

The runnable examples are TypeScript files. Some source imports end in .js because Node ESM resolves compiled package output that way; the files you edit and run are still .ts.

From a cloned checkout:

pnpm --dir examples/headless-workpaper install --ignore-workspace
pnpm --dir examples/headless-workpaper run start
pnpm --dir examples/headless-workpaper run json-records
pnpm --dir examples/headless-workpaper run csv-shaped
pnpm --dir examples/headless-workpaper run invoice-totals
pnpm --dir examples/headless-workpaper run budget-variance
pnpm --dir examples/headless-workpaper run fulfillment-capacity
pnpm --dir examples/headless-workpaper run quote-approval
pnpm --dir examples/headless-workpaper run subscription-mrr
pnpm --dir examples/headless-workpaper run persistence

The most useful entry points:

• JSON records input
• CSV shaped input
• invoice totals
• budget variance alerts
• fulfillment capacity plan
• quote approval threshold
• subscription MRR forecast
• SheetJS, xlsx-populate, and ExcelJS recalculation bridge

For tool integrations:

pnpm --dir examples/headless-workpaper run agent:verify
pnpm --dir examples/headless-workpaper run agent:tool-call
pnpm --dir examples/headless-workpaper run agent:openai-agents-sdk
pnpm --dir examples/headless-workpaper run agent:openai-agents-sdk-mcp
pnpm --dir examples/headless-workpaper run agent:openai-agents-sdk-hosted-mcp
pnpm --dir examples/headless-workpaper run agent:openai-responses
pnpm --dir examples/headless-workpaper run agent:ai-sdk-generate-text
pnpm --dir examples/headless-workpaper run agent:ai-sdk-stream-text
pnpm --dir examples/headless-workpaper run agent:framework-adapters
pnpm --dir examples/serverless-workpaper-api run hono-route
pnpm --dir examples/mastra-workpaper-tool run smoke
pnpm --dir examples/langgraph-workpaper-tool-state run smoke
pnpm --dir examples/langchain-mcp-workpaper-toolnode run smoke
pnpm --dir examples/headless-workpaper run agent:mcp-tools
pnpm --dir examples/headless-workpaper run agent:mcp-file-transcript
pnpm --dir examples/headless-workpaper run agent:mcp-xlsx-risk-preflight
pnpm --dir examples/headless-workpaper run agent:mcp-stdio

The AI SDK example uses ai-sdk-generate-text-tool-smoke.ts. The OpenAI Agents SDK guide is docs/openai-agents-sdk-workpaper-tool.md. It includes direct tool() wrapping, private MCPServerStdio discovery, and hosted stateless MCPServerStreamableHttp discovery through the WorkPaper MCP tool loop. The ChatGPT Apps Developer Mode setup is docs/chatgpt-apps-workpaper-mcp.md. It shows the public /mcp endpoint as a data/tool-only remote MCP app and keeps custom Apps SDK component UI as future scope. The OpenAI Responses guide is docs/openai-responses-workpaper-tool-call.md. The agent framework guide is docs/vercel-ai-sdk-langchain-spreadsheet-tool.md. The Mastra guide includes a real @mastra/core createTool() smoke: docs/mastra-workpaper-spreadsheet-tool.md. The LangGraph.js ToolNode proof is docs/langgraph-workpaper-toolnode-spreadsheet.md. It includes a no-key @langchain/mcp-adapters smoke that discovers the published WorkPaper MCP stdio tools and executes them through ToolNode.

The package also ships the MCP stdio binary:

npm exec --package @bilig/workpaper@latest -- bilig-agent-challenge --json
npm exec --package @bilig/workpaper@latest -- bilig-formula-clinic ./reduced.xlsx --cells "Summary!B7,Inputs!B2"
npm exec --package @bilig/workpaper@latest -- bilig-mcp-challenge --json
npm exec --package @bilig/workpaper@latest -- bilig-workpaper-mcp
npm exec --package @bilig/workpaper@latest -- bilig-workpaper-mcp --workpaper ./pricing.workpaper.json --init-demo-workpaper --writable
npm exec --package @bilig/workpaper@latest -- bilig-workpaper-mcp --from-xlsx ./pricing.xlsx
npm exec --package @bilig/workpaper@latest -- bilig-workpaper-mcp --from-xlsx ./pricing.xlsx --workpaper ./.bilig/pricing.workpaper.json --writable
pnpm --dir examples/headless-workpaper run agent:mcp-xlsx-risk-preflight
docker build --target bilig-workpaper-mcp -t bilig-workpaper-mcp:local .

bilig-agent-challenge prints the same edit, formula readback, WorkPaper JSON export, restore, and verified: true proof object used by the agent workbook challenge page.

bilig-mcp-challenge proves the file-backed MCP path end to end: initialize JSON-RPC, list tools/resources/prompts, edit Inputs!B3, read recalculated Summary!B3, export the WorkPaper JSON, restart from disk, and return verified: true.

bilig-formula-clinic imports a reduced XLSX locally, samples formulas, reads requested cells through WorkPaper, and prints a Markdown issue body. It does not upload workbook contents.

Without --workpaper, the binary starts the built-in demo workbook. With --workpaper, it loads your persisted WorkPaper JSON and exposes list_sheets, read_range, read_cell, set_cell_contents, set_cell_contents_and_readback, get_cell_display_value, export_workpaper_document, and validate_formula; --writable persists set_cell_contents or set_cell_contents_and_readback edits back to the same file. If you already have an XLSX, --from-xlsx ./pricing.xlsx imports it into an in-memory WorkPaper server for readback, throwaway edits, and analyze_workbook_risk without writing a sidecar. Add --workpaper ... --writable only when the agent needs persisted file state. It also exposes MCP resources and prompts for bilig://workpaper/agent-handoff, bilig://workpaper/current-document, edit_and_verify_workpaper, and debug_workpaper_formula, so capable clients can discover the workflow before calling tools. For a maintained real-XLSX transcript, run pnpm --dir examples/headless-workpaper run agent:mcp-xlsx-risk-preflight. It calls analyze_workbook_risk, edits Inputs!B3, verifies Summary!B3 changes from 60000 to 96000, exports WorkPaper JSON, and keeps excelParity: "not_proven". The Docker target is for MCP directory scanners: it seeds a demo WorkPaper JSON inside the image and starts the file-backed --writable tool surface so tools/list, resources/list, and prompts/list return the general WorkPaper agent surface without cloning this monorepo. For remote MCP clients, the app runtime exposes https://bilig.proompteng.ai/mcp as a stateless JSON-only Streamable HTTP endpoint for tool discovery and write/readback smoke tests.

It is published in the official MCP Registry as io.github.proompteng/bilig-workpaper: https://registry.modelcontextprotocol.io/v0.1/servers?search=io.github.proompteng%2Fbilig-workpaper. It is also live on Glama with Try in Browser, A-grade tool pages, and the file-backed WorkPaper tools: https://glama.ai/mcp/servers/proompteng/bilig.

Proof You Can Reproduce

• The 90-second TypeScript check above edits one input, restores the saved JSON document, and verifies the dependent formula result.
• For a service evaluator path, run the quote approval WorkPaper API proof. It starts from an empty Node directory, downloads one maintained TypeScript route smoke, writes quote inputs, recalculates an approval decision, persists JSON, and verifies restored readback.
• For a shorter evaluation page, read formula workbooks for Node services and tool integrations. It compresses the WorkPaper boundary, MCP file-backed mode, benchmark caveat, and alternative-tool guidance into one evaluation path.
• For a compact review note, use the WorkPaper maintainer proof note. It keeps the empty npm-project command, verified: true output, benchmark caveat, known limits, and open questions together.
• For saved-file integration, run the XLSX formula recalculation example: examples/xlsx-recalculation-node. It imports a generated XLSX pricing workbook, edits input cells, reads the recalculated approval decision, exports XLSX, reimports it, and verifies the formulas survived the round trip. The public decision page is XLSX formula recalculation in Node.js.
• Run pnpm workpaper:bench:competitive:check. The checked-in artifact shows 100/100 comparable WorkPaper mean wins and 100/100 mean+p95 wins; the current worst p95 row is sheet-rename-dependencies at 0.792x.
• The benchmark card is generated from that artifact: docs/assets/workpaper-benchmark-card.png.
• Read the compatibility limits before importing real Excel workbooks.
• Use the production adoption checklist before promoting a WorkPaper-backed workflow beyond evaluation.
• For XLSX accuracy audits, use the Excel oracle harness. It separates import success, timeouts, cached workbook values, and fresh Microsoft Excel recalculation results.
• The WorkPaper MCP server is listed in the official MCP Registry and on Glama. The directory status page keeps the npm command, remote endpoint, static MCP server card, and directory evidence in one place.
• Public feedback threads: workflow questions, service examples, persistence adapters, JavaScript spreadsheet library guide, OpenAI Responses tool calls, and benchmark critique.

If you are evaluating Bilig runtime packages for production and want release notifications, watch releases: https://github.com/proompteng/bilig/subscription.

XLSX Accuracy Policy

Cached formula values embedded in .xlsx files are cache diagnostics, not an accuracy verdict. A Bilig correctness bug should only be claimed when the expected value came from a fresh Excel recalculation oracle.

OUT=.cache/excel-oracle-evaluation
pnpm workpaper:xlsx-oracle -- prepare-oracle /path/to/xlsx-corpus "$OUT"
pnpm workpaper:xlsx-oracle -- evaluate-cache /path/to/xlsx-corpus "$OUT"
pnpm workpaper:xlsx-oracle -- evaluate-oracle /path/to/xlsx-corpus "$OUT/recalculated" "$OUT"
pnpm workpaper:xlsx-oracle -- summarize "$OUT"

evaluate-cache writes cache-diagnostic.json and stays non-authoritative. evaluate-oracle writes excel-oracle-report.json, and summarize writes summary.md. If Excel automation is unavailable, cells are classified as missing_excel_oracle instead of being promoted to bugs.

What Is In This Repo

• packages/workpaper: public WorkPaper package, starters, evaluator binaries, and MCP wrappers.
• packages/headless: lower-level WorkPaper runtime that backs the public package.
• packages/excel-import: saved-workbook import/export boundary.
• packages/formula: formula parser, binder, compiler, and evaluator.
• packages/core: workbook engine, snapshots, mutation flow, and scheduler.
• packages/grid and apps/web: browser spreadsheet shell.
• apps/bilig: fullstack monolith runtime, API surface, and static asset server.
• packages/renderer: React workbook renderer.
• packages/protocol, packages/binary-protocol, packages/agent-api, and packages/worker-transport: protocol and integration boundaries.
• packages/wasm-kernel: AssemblyScript/WASM numeric fast path.
• packages/benchmarks: benchmark harness and performance contracts.

For XLSX import/export from TypeScript:

import { WorkPaper } from '@bilig/workpaper'
import { exportXlsx, importXlsx } from '@bilig/workpaper/xlsx'

Use WorkPaper.buildFromSnapshot(imported.snapshot) after import and workbook.exportSnapshot() before exportXlsx().

Local Development

Use Node 24+, Bun, and [email protected].

pnpm install
pnpm dev:web
pnpm dev:web-local
pnpm dev:sync

For a full local preflight:

pnpm lint
pnpm typecheck
pnpm test
pnpm test:browser
pnpm run ci

Generated sources and public evidence are checked:

pnpm protocol:check
pnpm formula-inventory:check
pnpm workspace-resolution:check
pnpm workpaper:bench:competitive:check
pnpm docs:discovery:check

For Coding Agents

Start with the public package boundary unless the task is explicitly engine work.

1. Read packages/workpaper/README.md before touching public WorkPaper behavior.
2. Read docs/AGENTS.md, docs/skill.md, or docs/llms-full.txt when building an agent-facing integration from outside the repo.
3. Use public exports from @bilig/workpaper; do not reach into src/ or dist/ when writing consumer examples.
4. Keep examples TypeScript-first.
5. Do not call embedded XLSX stored formula results an accuracy oracle.
6. Add focused tests before changing formulas, persistence, range bounds, config rebuilds, events, row/column moves, or sheet lifecycle.
7. Run the focused package tests first, then broaden to pnpm run ci.

Contributing

Read CONTRIBUTING.md before opening a PR. If this is your first patch, start with the new contributor guide and then claim a scoped starter issue.

Good first patches usually fit one of these shapes:

• formula fixtures with clear expected behavior;
• small WorkPaper examples that prove a real service or agent workflow;
• focused correctness fixes with regression tests;
• grid accessibility and keyboard-behavior improvements;
• docs that turn an existing architecture note into a runnable command.

The shortest public on-ramp is the starter issues queue. It keeps code/test picks, example tasks, adapters, and focused docs work in one current list, with small acceptance commands for first patches.

If this is your first contribution to bilig, use the first-timers-only filter.

Security And Support

Read SECURITY.md before sharing vulnerability details, private workbook data, tokens, credentials, or exploit reproductions. Security reports should use GitHub private vulnerability reporting when available, or [email protected] when the private flow is not visible.

Use SUPPORT.md for the fastest public support path. Good reports include the package version, Node version, OS, exact formula or workbook input, expected value, actual value, and the smallest command or script that reproduces the issue.

CI

Forgejo Actions is the primary CI surface via .forgejo/workflows/forgejo-ci.yml. GitHub Actions mirrors the verification contract in .github/workflows/ci.yml.

The strict gate includes frozen lockfile install, full pnpm run ci, artifact budget checks, browser smoke, and tracked-file cleanliness checks.

License

MIT.