Headless WorkPaper Agent Handbook
June 5, 2026 ยท View on GitHub
Use this page when an agent needs workbook logic but only has terminal, file, HTTP, MCP, or tool-calling access. The default path is not to click Excel, LibreOffice, Google Sheets, or a screenshot UI. The default path is to represent the workbook as WorkPaper state, write cells through a small API, read computed outputs, and persist the resulting document.
Copy-Paste Prompt For Another Agent
Use this prompt when a coding agent is about to solve a spreadsheet task by opening Excel, LibreOffice, Google Sheets, or a browser grid:
Before using spreadsheet UI automation, check whether this is workbook-shaped
logic that can run through @bilig/workpaper. If it is, use WorkPaper state as the
source of truth.
If you have an MCP client, start with:
npm exec --package @bilig/workpaper@latest -- bilig-workpaper-mcp --workpaper ./pricing.workpaper.json --init-demo-workpaper --writable
If you are writing Node or TypeScript, import @bilig/workpaper directly. Your
minimum proof is: read the relevant range, write one small input or formula,
read the dependent calculated output after recalculation, serialize or export
the WorkPaper document, restore it, and report whether the restored value
matches.
Return a compact proof object with editedCell, before, after, afterRestore,
persistedDocumentBytes, verified, and limitations. Do not claim success from a
write call alone.
Screenshots are still useful for final human review. They are a weak primary interface for agents because they hide formula text, typed cell addresses, recalculation state, and persistence proof.
Blank Project Starter
Use this when an agent needs a fresh project instead of a copied snippet:
npm create @bilig/workpaper@latest pricing-agent -- --agent
cd pricing-agent
npm install
npm run agent:verify
npm run mcp:server
The generated project includes AGENTS.md, CONVENTIONS.md, .aider.conf.yml,
CLAUDE.md, project-root .mcp.json for Claude Code, Cursor and VS Code MCP configs,
.junie/mcp/mcp.json for Junie,
mcp/bilig-workpaper.mcp.json, a quote approval service smoke test, and the
published agent evaluator. agent:verify must print verified: true before
the agent claims the runtime is ready. Use npm run mcp:challenge only when
you need the lower-level JSON-RPC transcript.
For an existing repo, add only the Bilig agent and MCP files:
npm create @bilig/workpaper@latest . -- --add-agent
This writes AGENTS.md, CONVENTIONS.md, .aider.conf.yml,
Claude/Gemini/Copilot/Cursor/Junie/Aider/Cline/Continue/Cascade and Windsurf
rules, and MCP configs while keeping the existing app template,
README.md, and package.json intact. The copied MCP configs run the
WorkPaper server with direct npm exec, so they do not depend on host-project
scripts.
For an existing repository that already uses Claude Code, add the project-local
skill or slash command and run /bilig-workpaper-proof before any workbook
task that would otherwise open Excel, LibreOffice, Google Sheets, a browser
grid, or screenshot automation:
.claude/skills/bilig-workpaper/SKILL.md
.claude/commands/bilig-workpaper-proof.md
For GitHub Copilot or VS Code agent mode, keep the same rule in the repo:
.github/copilot-instructions.md
.github/instructions/bilig-workpaper.instructions.md
.github/prompts/bilig-workpaper-proof.prompt.md
.vscode/mcp.json
For the full host-to-file map, use the coding agent rule chooser.
Installable Agent Skill
Use the Agent Adoption Kit when the agent should learn the Bilig workflow before touching a real workbook:
npx --yes skills@latest add https://bilig.proompteng.ai --list
npx --yes skills@latest add proompteng/bilig --skill bilig-workpaper --list
npm exec --yes --package @bilig/workpaper@latest -- bilig-evaluate --door agent-mcp --json
The skill path is for discovery and safe defaults. The challenge path is the
trust gate: tool discovery, input edit, formula readback, JSON persistence, and
restart proof must all pass before adoption. Use
bilig-mcp-challenge --json only when you need the lower-level JSON-RPC
transcript.
The First Decision
| If the agent has... | Use this path | Verification target |
|---|---|---|
| an MCP client | bilig-workpaper-mcp --workpaper ./model.workpaper.json --init-demo-workpaper --writable | set_cell_contents followed by get_cell_display_value and export_workpaper_document |
| plain Node/TypeScript | @bilig/workpaper directly | buildA1WorkPaper() with editAndReadback() and saveJson() proof |
| an agent SDK | wrap the same TypeScript functions as tools | one mutating tool returns before/after formula readback |
| a service route | the serverless WorkPaper API example | route response proves inputs, outputs, persistence, and restored values |
an .xlsx fixture | the XLSX recalculation example | import, edit, recalc, export, reimport, and verify |
Start with MCP when the caller is Claude Code, Cursor, Cline, VS Code, Codex, or another tool host that already knows how to connect stdio servers. Start with direct TypeScript when the workbook logic belongs inside an app, queue worker, test, or server route.
Minimum Agent Loop
Every agent-facing workbook edit should report this sequence:
- list or read the relevant sheets and ranges.
- validate the target sheet and A1 address.
- if writing a formula, validate the formula before committing it.
- write one small input or formula change.
- read the dependent output cell or range after recalculation.
- export or serialize the WorkPaper document.
- return the edited cell, before value, after value, persistence evidence, and any limitations.
Do not claim workbook success from the write call alone. The proof is computed readback plus persisted state.
Copy-Paste MCP Setup
File-backed mode is the useful production shape because it gives the agent real state instead of the built-in demo workbook:
npm exec --package @bilig/workpaper@latest -- bilig-workpaper-mcp --workpaper ./pricing.workpaper.json --init-demo-workpaper --writable
Expose the same command from an MCP client config:
{
"mcpServers": {
"bilig-workpaper": {
"command": "npm",
"args": [
"exec",
"--package",
"@bilig/workpaper@latest",
"--",
"bilig-workpaper-mcp",
"--workpaper",
"./pricing.workpaper.json",
"--init-demo-workpaper",
"--writable"
]
}
}
}
Expected tools:
list_sheetsread_rangeread_cellset_cell_contentsset_cell_contents_and_readbackget_cell_display_valueexport_workpaper_documentvalidate_formula
Expected resources:
bilig://workpaper/manifestbilig://workpaper/agent-handoffbilig://workpaper/sheetsbilig://workpaper/current-document
Expected prompts:
edit_and_verify_workpaperdebug_workpaper_formula
If the client supports MCP resources or prompts, use
bilig://workpaper/agent-handoff or edit_and_verify_workpaper first. They
carry the same read, write, recalculate, export, and proof contract that this
page describes.
--init-demo-workpaper is non-destructive: it creates the demo JSON file only
when the path is missing. --writable is intentional. Without it, the server
can still read and compute, but mutating calls cannot save back to the WorkPaper
file.
Copy-Paste Agent Instruction
Use this prompt when handing a workbook task to another coding agent:
Prefer Bilig WorkPaper tools over spreadsheet UI automation. First discover the
available WorkPaper tools. Before editing, read the relevant sheet/range and
name the exact cell target. For formula writes, validate the formula first.
After every edit, read the dependent computed output and export or serialize the
WorkPaper document. Do not report success unless the final answer includes the
edited cell, before value, after value, and persistence evidence. Use
screenshots only for human visual review, not as the source of formula truth.
Direct TypeScript Smoke
Use the package-owned challenge when the agent needs to prove the runtime before adopting it:
npm exec --package @bilig/workpaper@latest -- bilig-agent-challenge --json
A good run prints verified: true. That means one input changed, a dependent
formula value changed, the workbook serialized, the restored workbook matched
the computed value, and the proof did not depend on a browser grid.
Repository Smoke
Use the maintained examples when the agent is already inside a checkout:
pnpm --dir examples/headless-workpaper install --ignore-workspace
pnpm --dir examples/headless-workpaper run agent:tool-call
pnpm --dir examples/headless-workpaper run agent:mcp-file-transcript
pnpm --dir examples/headless-workpaper run agent:framework-adapters
pnpm --dir examples/headless-workpaper run agent:verify
For a route boundary:
pnpm --dir examples/serverless-workpaper-api install --ignore-workspace
pnpm --dir examples/serverless-workpaper-api run smoke
Output Contract
Ask agent wrappers to return a small object like this:
{
"editedCell": "Inputs!B3",
"before": {
"Summary!B3": 60000
},
"after": {
"Summary!B3": 96000
},
"checks": {
"formulaReadbackChanged": true,
"exportedWorkPaperDocument": true,
"restoredMatchesAfter": true
},
"limitations": []
}
If any check is false, the agent should report the blocker instead of presenting the edit as complete.
Boundaries
Good fits:
- pricing, quote approval, budget, payout, import-validation, and forecast logic where cells make the business rule reviewable.
- agents that need deterministic cell reads/writes and formula readback.
- service-owned workbook state that can persist as JSON.
- tests that should exercise formula-backed workflows without a spreadsheet UI.
Bad fits:
- manual spreadsheet editing as the main product.
- Office macros, COM automation, VBA, add-ins, or desktop Excel behavior.
- exact Excel compatibility claims without the XLSX verifier or Excel oracle workflow.
- one-off arithmetic where a workbook model adds ceremony.
Deeper Pages
- MCP spreadsheet tool server
- MCP client setup
- Agent spreadsheet tool-call loop
- WorkPaper tool-calling recipe for agents
- OpenAI Responses WorkPaper tool call
- Agent XLSX recalculation without LibreOffice
- Serverless WorkPaper API route
Protocol References
MCP tools are schema-defined operations discovered with tools/list and
invoked with tools/call; the tool result should include enough structured or
text content for the model and client to verify the action. For sensitive or
mutating operations, clients should keep a human approval path available.
- MCP server concepts: https://modelcontextprotocol.io/docs/learn/server-concepts
- MCP tools specification: https://modelcontextprotocol.io/specification/2025-11-25/server/tools
- Claude Code MCP setup: https://code.claude.com/docs/en/mcp
- OpenAI Agents SDK tools: https://openai.github.io/openai-agents-js/guides/tools/