spreadsheet-kit MCP

agent-spreadsheet

[CI] [Crates.io] [npm] [License]

[agent-spreadsheet]

agent-spreadsheet is the tool interaction service for agent-based spreadsheet usage.

It gives agents a safe, inspectable, token-efficient way to read, analyze, mutate, verify, and operationalize Excel workbooks without falling back to brittle UI automation.

If you want an agent to work with spreadsheets like a real system instead of a screenshot puppet, this is the stack.

What this project is

agent-spreadsheet ships a unified spreadsheet interaction layer across three surfaces:

The WASM build (agent-spreadsheet-wasm) is the SDK's local runtime, not a separate product surface: JS and TypeScript code targets one object model and the execution substrate (server vs embedded engine) is a configuration choice.

In 0.16 these surfaces share the resident Rust document/evaluator runtime and 32-operation registry. Explicit native sessions automatically start a private journal-backed host. SDK/WASM and genuine just-bash sessions retain their history in memory; XLSX export saves a document snapshot, not a restartable session journal. Warm edit/recalculate/read loops avoid XLSX serialization and repeated evaluator ingestion. See runtime guarantees and limits and the SDK/just-bash examples.

Supported workbook modes:

• .xlsx / .xlsm — read + write
• .xls / .xlsb — discovery/read-oriented workflows only

Powered by Formualizer

Default in-process calculation uses Formualizer — a permissively licensed (MIT/Apache-2.0) spreadsheet engine written in Rust: formula parsing, dependency-graph recalculation, 400+ Excel functions, dynamic arrays, and deterministic evaluation built for agents. No Excel COM or headless LibreOffice is required. Optional LibreOffice integration remains available, but Formualizer is the release-critical backend.

That native engine is why this project can offer what most spreadsheet tooling for agents cannot: recalculate the actual workbook, trace which cells changed and why, and prove it — not just read cached values or push blind edits.

Embedding spreadsheet logic in your own product rather than driving workbooks as an agent? Use Formualizer directly (Rust, Python, JS/WASM).

Why agents use agent-spreadsheet

Built for tool use, not just humans

• deterministic JSON contracts
• schema and example discovery from the CLI itself
• explicit pagination and compact output modes
• machine-readable warnings and error envelopes

Safe mutation, not blind mutation

• dry-run first workflows
• stateless output modes and overwrite safety
• event-sourced session editing
• verification surfaces for proving downstream outcomes
• structural impact analysis before risky workbook changes

Spreadsheet-aware, not generic file editing

• region detection
• table and footer-aware append helpers
• template row / row band cloning
• formula-specific replace and diagnostics
• named range CRUD
• recalculation + diff + proof flows

Good agent ergonomics

• nested command groups with legacy alias compatibility
• token-efficient reads
• exact-cell inspection and layout inspection
• workflow helpers for the repetitive parts agents usually get wrong

What is new / what makes this stack different

The current surface is much stronger than a plain “read some cells” tool. Major capabilities now include:

• asp as the primary CLI with agent-spreadsheet preserved as a compatibility alias
• grouped verification via asp verify proof and asp verify diff
• preview-first workflow helpers for:
  • write append
  • write clone-template-row
  • write clone-row-band
• formula-safe batch workflows with parse-policy diagnostics
• cell/layout/export/import inspection surfaces
• named range management (write name define|update|delete)
• formula-only replacement (write formulas replace)
• event-sourced session editing with log, branch, undo/redo, fork, apply, and materialize
• SheetPort manifest lifecycle + execution for contract-driven spreadsheet automation

Install

Shell installer

curl -fsSL https://raw.githubusercontent.com/PSU3D0/agent-spreadsheet/main/install.sh | sh

The installer downloads a prebuilt CLI to ~/.local/bin and creates the asp command. Pin a release with ASP_VERSION=0.16.0, set ASP_INSTALL_DIR to choose another destination, or pass --mcp to install the MCP server too:

curl -fsSL https://raw.githubusercontent.com/PSU3D0/agent-spreadsheet/main/install.sh | sh -s -- --mcp

npm

npm i -g agent-spreadsheet

This installs both asp (the primary command) and agent-spreadsheet (the compatibility alias) from a prebuilt native binary. No Rust toolchain is required.

cargo-binstall

cargo binstall agent-spreadsheet

This installs the prebuilt CLI in seconds.

Cargo

cargo install agent-spreadsheet --features recalc --bin asp --bin agent-spreadsheet

This builds the CLI from source. Formualizer (the native Rust recalc engine) is included by default.

mise

mise use -g "ubi:PSU3D0/agent-spreadsheet[exe=asp]"
# Homebrewbrew install psu3d0/tap/agent-spreadsheet

MCP server

cargo install agent-spreadsheet-mcp

Docker

# Read-only / slimdocker pull ghcr.io/psu3d0/agent-spreadsheet-mcp:latest
# Write + recalc + screenshotsdocker pull ghcr.io/psu3d0/agent-spreadsheet-mcp:latest-full

JavaScript SDK

npm i agent-spreadsheet-sdk

Prebuilt binaries

Download raw binaries and archives from GitHub Releases.

Published native assets include:

• Linux x86_64
• Linux arm64
• macOS x86_64
• macOS arm64
• Windows native binaries are not shipped for 0.16; native release validation targets Linux and macOS

Start here: the core workflows

1) Orient the workbook before reading cells

# What sheets are here?asp read sheets data.xlsx
# What regions/tables/parameter blocks does this sheet contain?asp read overview data.xlsx "Model"
# What named items are available?asp read names data.xlsx
# Read a structured region as a tableasp read table data.xlsx --sheet "Model"

2) Inspect exactly what an agent needs

# Raw values for exact rangesasp read values data.xlsx Model A1:C20
# Detail-view for targeted cells (value / formula / cached / style triage)asp read cells data.xlsx Model B2 D10:F12
# Layout-aware rendering for a bounded rangeasp read layout data.xlsx Model --range A1:H30 --render both
# Export a bounded range to csv or grid jsonasp read export data.xlsx Model A1:H30 --format csv --output model.csv

3) Do a safe stateless edit → recalc → proof → diff loop

asp workbook copy data.xlsx /tmp/draft.xlsxasp write cells /tmp/draft.xlsx Inputs "B2=500" "C2==B2*1.1"asp workbook recalculate /tmp/draft.xlsxasp verify proof data.xlsx /tmp/draft.xlsx --targets Summary!B2,Summary!B3 --named-rangesasp verify diff data.xlsx /tmp/draft.xlsx --details --limit 50

A representative label-mode lookup:

asp analyze find-value data.xlsx "Net Income" --mode label --label-direction below

4) Preview structural risk before mutating the workbook

asp analyze ref-impact data.xlsx --ops @structure_ops.json --show-formula-delta

This is intentionally read-only. It surfaces shifted spans, absolute-reference warnings, token counts, and optional before/after formula samples.

5) Use workflow helpers instead of reinventing row logic

# Stateless batch writesasp write batch transform data.xlsx --ops @ops.json --dry-runasp write batch style data.xlsx --ops @style_ops.json --dry-run
# Append rows into a detected region or table, respecting footer rows when presentasp write append data.xlsx --sheet Revenue --table-name RevenueTable --from-csv rows.csv --header --dry-run
# Clone one template row with preview-first planningasp write clone-template-row data.xlsx --sheet Inputs --source-row 8 --after 8 --count 3 --dry-run
# Clone a contiguous row band repeatedlyasp write clone-row-band data.xlsx --sheet Forecast --source-rows 12:16 --after 16 --repeat 4 --dry-run

6) Use a stateful session when the edit story gets complex

asp session start --base data.xlsx --workspace .asp session op --session <id> --ops @edit.json --workspace .asp session apply --session <id> <staged_id> --workspace .asp session materialize --session <id> --output result.xlsx --workspace .

And when you need proper history and branching:

asp session log --session <id> --workspace .asp session fork --session <id> scenario-a --workspace .asp session undo --session <id> --workspace .asp session redo --session <id> --workspace .asp session checkout --session <id> <op_id> --workspace .

7) Turn workbook interfaces into contracts with SheetPort

# Discover candidate ports from workbook structureasp sheetport manifest candidates model.xlsx
# Validate or normalize a manifestasp sheetport manifest validate manifest.yamlasp sheetport manifest normalize manifest.yaml
# Bind-check a workbook against a manifestasp sheetport bind-check model.xlsx manifest.yaml
# Execute the manifest with JSON inputsasp sheetport run model.xlsx manifest.yaml --inputs @inputs.json

CLI overview

The primary CLI is asp.

agent-spreadsheet remains available as a compatibility alias, so both of these are valid:

asp read sheets data.xlsxagent-spreadsheet read sheets data.xlsx

Preferred command groups

• asp read ...
• asp analyze ...
• asp write ...
• asp workbook ...
• asp verify ...
• asp session ...
• asp sheetport ...

Legacy aliases

Legacy flat commands are still normalized to the new nested surface where practical. That makes migration easier for older prompts, docs, and automation.

Discoverability built into the CLI

When an agent is unsure of payload shape, it can ask the tool directly:

asp operations                  # CLI-supported runtime subsetasp registry --all              # complete host-independent registry + schemasasp schema read_cellsasp example read_cellsasp schema write batch transformasp example write batch transformasp schema session op transform.write_matrixasp example session op transform.write_matrix

Canonical machine calls use the same registry and dispatcher as other surfaces. asp schema <canonical-op> and asp example <canonical-op> project the native adapter contract: resource_id and, for verification, baseline_resource_id are omitted from required JSON because --bind and --baseline inject ephemeral resources. The host-independent asp registry --all remains unchanged.

asp op read_cells --bind data.xlsx --json '{"sheet_name":"Sheet1","selection":{"kind":"range","ranges":["A1:C10"]}}'asp op verify_workbook --baseline base.xlsx --bind current.xlsx --json '{}'echo '{"action":"schema"}' | asp op sheetport_manifest

--bind reads the current workbook, while --baseline supplies the second workbook only for verify_workbook. Canonical mutable CLI calls require exactly one persistence target: --output <path> writes a new file or --in-place atomically replaces the bound file; pure preview persists nothing and accepts neither. Durable fork, checkpoint, stage, and history operations are intentionally absent from stateless CLI discovery. This is a core design principle: the surface should explain itself to the agent.

Command families

read — extraction and inspection

Why these matter for agents

Agents rarely need “the whole spreadsheet.” They need:

• the right region
• the right page
• the right cells
• just enough layout to understand intent

That is why the read surface combines region detection, structured reads, detail inspection, and explicit continuation.

analyze — search, diagnostics, and impact understanding

Why this matters

Headless spreadsheet automation wins when it can explain consequences, not just execute mutations. ref-impact, formula-trace, and grouped diagnostics are all part of that story.

write — safe mutations and workflow helpers

Safety model

Most mutating commands support a strict mode matrix:

• --dry-run
• --in-place
• --output <PATH>

This matters for agents because it allows:

• dry-run planning
• non-destructive execution
• explicit overwrite control

Formula maintenance

Formula mutation is now a first-class surface:

asp write formulas replace data.xlsx Sheet1 --find '$64' --replace '$65' --dry-runasp write formulas replace data.xlsx Sheet1 --find 'Sheet1!' --replace 'Sheet2!' --range A1:Z100 --output fixed.xlsx

Named range maintenance

asp write name define data.xlsx RevenueInput 'Inputs!$B$2'asp write name update data.xlsx RevenueInput 'Inputs!$B$2:$B$4' --in-placeasp write name delete data.xlsx RevenueInput --in-place

workbook — file-level flows

verify — proof, not vibes

Why verification matters

Most spreadsheet automation tools stop at “the edit applied.”

agent-spreadsheet goes further:

• did the target cells change the way we expected?
• did the workbook introduce new errors?
• which changes were direct edits vs recalculation fallout?
• what changed overall, grouped in a way an agent can reason about?

This verification layer is a big part of why this project is a serious agent substrate rather than a utility script.

session — event-sourced stateful editing

The session surface is for workflows that are too complex for a single stateless write.

What sessions give you

• persistent editing state
• staged dry-run operations
• compare-and-swap apply semantics
• logs and replayability
• branch/switch/fork flows
• undo / redo / checkout
• explicit materialization back to a workbook file

Canonical loop

asp session start --base model.xlsx --workspace .asp session op --session <id> --ops @ops.json --workspace .asp session apply --session <id> <staged_id> --workspace .asp session materialize --session <id> --output result.xlsx --workspace .

History and branching

asp session log --session <id> --workspace .asp session branches --session <id> --workspace .asp session fork --session <id> experiment-b --workspace .asp session switch --session <id> experiment-b --workspace .asp session undo --session <id> --workspace .asp session redo --session <id> --workspace .asp session checkout --session <id> <op_id> --workspace .

Use sessions when you want repeatability, auditability, and multi-step safety.

sheetport — spreadsheet interfaces as executable contracts

SheetPort is the workflow surface for turning workbook inputs/outputs into explicit machine contracts.

Manifest lifecycle

asp sheetport manifest candidates model.xlsxasp sheetport manifest schemaasp sheetport manifest validate manifest.yamlasp sheetport manifest normalize manifest.yaml

Bind-check + run

asp sheetport bind-check model.xlsx manifest.yamlasp sheetport run model.xlsx manifest.yaml --inputs @inputs.json --freeze-volatile

Use this when you want a workbook to behave less like an opaque file and more like a declared service interface.

Output contracts for agents

Canonical vs compact shapes

All commands default to JSON. Many also support:

--shape canonical--shape compact

Policy:

• canonical keeps the full stable schema
• compact removes wrapper noise where the contract allows it while preserving continuation fields and command-specific semantics

Shape policy:

• Canonical (default): preserve the full response schema.
• range-values: returns a stable values: [...] envelope in both canonical and compact modes.
• range-values default encoding: dense JSON (dense.encoding = "dense_v1") with dictionary + run-length row_runs.
• range-values --include-formulas: includes sparse formula coordinates in dense mode (dense.formulas), or a matrix in explicit json format.
• read-table and sheet-page: compact preserves the active branch and continuation fields (next_offset, next_start_row).
• formula-trace compact: omits per-layer highlights while preserving layers and next_cursor.

Deterministic pagination loops

# sheet-page continuationasp read page data.xlsx Sheet1 --format compact --page-size 200asp read page data.xlsx Sheet1 --format compact --page-size 200 --start-row 201
# read-table continuationasp read table data.xlsx --sheet "Sheet1" --table-format values --limit 200 --offset 0asp read table data.xlsx --sheet "Sheet1" --table-format values --limit 200 --offset 200

sheet-page machine contract

• Inspect top-level format before reading payload fields.
• format=full: read top-level rows plus optional header_row and next_start_row.
• format=compact: read compact.headers, compact.header_row, compact.rows plus optional next_start_row.
• format=values_only: read values_only.rows plus optional next_start_row.
• Continuation is always driven by top-level next_start_row when present.
• Global --shape compact preserves the active sheet-page branch; it does not flatten sheet-page payloads.

Machine continuation example:

1. Request page 1 without --start-row.
2. If next_start_row is present, call sheet-page again with --start-row <next_start_row>.
3. Stop when next_start_row is omitted.

Self-describing payloads

When the agent is unsure what to send, ask for a schema or example:

asp schema write batch rulesasp example write batch rulesasp schema session op structure.insert_rowsasp example session op structure.insert_rows

Batch payload examples

All batch payloads use a top-level envelope object. Most commands require {"ops":[...]}; column-size-batch prefers {"sheet_name":"...","ops":[...]} and also accepts per-op sheet_name inside {"ops":[...]}.

transform-batch payloads (@transform_ops.json)

• Minimal: {"ops":[{"kind":"fill_range","sheet_name":"Sheet1","target":{"kind":"range","range":"B2:B4"},"value":"0"}]}
• Advanced: {"ops":[{"kind":"replace_in_range","sheet_name":"Sheet1","target":{"kind":"region","region_id":1},"find":"N/A","replace":"","match_mode":"contains","case_sensitive":false,"include_formulas":true}]}

style-batch payloads (@style_ops.json)

• Minimal: {"ops":[{"sheet_name":"Sheet1","target":{"kind":"range","range":"B2:B2"},"patch":{"font":{"bold":true}}}]}
• Advanced: {"ops":[{"sheet_name":"Sheet1","target":{"kind":"cells","cells":["B2","B3"]},"patch":{"number_format":"$#,##0.00","alignment":{"horizontal":"right"}},"op_mode":"merge"}]}

write batch formula-pattern payloads (@formula_ops.json)

• Minimal: {"ops":[{"sheet_name":"Sheet1","target_range":"C2:C4","anchor_cell":"C2","base_formula":"B2*2"}]}
• Advanced: {"ops":[{"sheet_name":"Sheet1","target_range":"C2:E4","anchor_cell":"C2","base_formula":"B2*2","fill_direction":"both","relative_mode":"excel"}]}
• relative_mode valid values: excel, abs_cols, abs_rows

structure-batch payloads (@structure_ops.json)

• Minimal: {"ops":[{"kind":"rename_sheet","old_name":"Summary","new_name":"Dashboard"}]}
• Advanced: {"ops":[{"kind":"copy_range","sheet_name":"Sheet1","dest_sheet_name":"Summary","src_range":"A1:C4","dest_anchor":"A1","include_styles":true,"include_formulas":true}]}

column-size-batch payloads (@column_size_ops.json)

• Minimal (preferred): {"sheet_name":"Sheet1","ops":[{"range":"A:A","size":{"kind":"width","width_chars":12.0}}]}
• Advanced (preferred): {"sheet_name":"Sheet1","ops":[{"target":{"kind":"columns","range":"A:C"},"size":{"kind":"auto","min_width_chars":8.0,"max_width_chars":24.0}}]}
• Also accepted (harmonized shape): {"ops":[{"sheet_name":"Sheet1","range":"A:A","size":{"kind":"width","width_chars":12.0}}]}

sheet-layout-batch payloads (@layout_ops.json)

• Minimal: {"ops":[{"kind":"freeze_panes","sheet_name":"Sheet1","freeze_rows":1,"freeze_cols":1}]}
• Advanced: {"ops":[{"kind":"set_page_setup","sheet_name":"Sheet1","orientation":"landscape","fit_to_width":1,"fit_to_height":1}]}

rules-batch payloads (@rules_ops.json)

• Minimal: {"ops":[{"kind":"set_data_validation","sheet_name":"Sheet1","target_range":"B2:B4","validation":{"kind":"list","formula1":"\"A,B,C\""}}]}
• Advanced: {"ops":[{"kind":"set_conditional_format","sheet_name":"Sheet1","target_range":"C2:C10","rule":{"kind":"expression","formula":"C2>100"},"style":{"fill_color":"#FFF2CC","bold":true}}]}

write batch formula-pattern clears cached results for touched formula cells; run workbook recalculate to refresh computed values.

Formula parse policy

Formula-aware commands support:

--formula-parse-policy fail|warn|off

• fail — abort
• warn — continue and attach grouped diagnostics
• off — skip silently

This lets agents choose between strictness and progress depending on the workflow.

CLI reference excerpts

• read values <file> <sheet> <range> [range...] [--format dense\|json\|values\|csv] [--include-formulas]
• read cells <file> <sheet> <target> [target...] [--include-empty]
• read page <file> <sheet> --format <full|compact|values_only> [--start-row ROW] [--page-size N]
• workbook create <path> [--sheets Inputs,Calc,...] [--overwrite]
• analyze find-value <file> <query> [--sheet S] [--mode value\|label] [--label-direction right\|below\|any]
• write batch transform <file> --ops @ops.json (--dry-run\|--in-place\|--output PATH)

Formula write-path provenance (write_path_provenance)

Formula-writing commands emit optional provenance metadata for troubleshooting:

• written_via: write path (edit, transform_batch, apply_formula_pattern)
• formula_targets: sheet/cell or sheet/range targets touched by formula writes

Debug compare workflow:

1. Apply the same formula target via two paths.
2. Compare write_path_provenance.written_via and formula_targets in responses.
3. Use inspect-cells plus recalculate to compare resulting behavior.

Financial presentation starter defaults

• Keep label columns (often column A) explicitly sized (roughly 24–36 chars) to prevent clipping.
• Apply consistent number formats by semantic type:
  • Currency: "$"#,##0.00_);[Red]("$"#,##0.00)
  • Percent: 0.0%
  • Integer/count: #,##0
• Apply sheet-layout-batch freeze panes after header layout stabilizes.

JSON output is compact by default; use --quiet to suppress warnings. Global --output-format csv is currently unsupported; use command-specific CSV options like read table --table-format csv.

MCP server quickstart

The MCP surface is the stateful server version of agent-spreadsheet.

Use it when you want:

• workbook caching across calls
• fork lifecycle instead of stateless file replacement
• multi-turn agent workflows
• screenshots and richer server-side orchestration

Claude Code / Claude Desktop

Add to ~/.claude.json or project .mcp.json:

{  "mcpServers": {    "spreadsheet": {      "command": "agent-spreadsheet-mcp",      "args": ["--workspace-root", "/path/to/workbooks", "--transport", "stdio"]    }  }}

Docker

{  "mcpServers": {    "spreadsheet": {      "command": "docker",      "args": [        "run", "-i", "--rm",        "-v", "/path/to/workbooks:/data",        "ghcr.io/psu3d0/agent-spreadsheet-mcp:latest-full",        "--transport", "stdio"      ]    }  }}

:latest is the read-only slim image (write/fork/recalc tools disabled); :latest-full includes the write tools and recalculation (LibreOffice-backed).

HTTP mode

agent-spreadsheet-mcp --workspace-root /path/to/workbooks# -> http://127.0.0.1:8079  (POST /mcp, and the canonical route under /v1)

The same process also serves a plain canonical HTTP route for programmatic clients that do not speak MCP:

curl -sS -X POST http://127.0.0.1:8079/v1/op/list_workbooks -d '{}'

Canonical errors return the canonical error envelope with the status mapped from the error code (400 invalid request, 404 unknown operation / resource, 409 revision conflict, 500 operation failed, 501 capability unavailable). Like /mcp, /v1 has no authentication; the loopback default bind is the security boundary. See docs/architecture/canonical-http-route.md.

Configuration

Every setting is available as a CLI flag (agent-spreadsheet-mcp --help), an environment variable, or a config file key (--config file.yaml). CLI takes precedence over the config file.

Setting any of the timeout/limit variables (TOOL_TIMEOUT_MS, MAX_RESPONSE_BYTES, MAX_PAYLOAD_BYTES, MAX_CELLS, MAX_ITEMS) to 0 disables that limit.

MCP tool surface

The default MCP router is generated from the canonical operation registry. A write-capable baseline exposes 27 operations: 17 discovery/read/analysis operations and 10 write/lifecycle operations. Capability-backed deployments can add screenshot_sheet, sheetport_manifest, execute_sheetport, and inspect_vba, for up to 31. Nineteen canonical operations are read-only.

Each tool uses the registry descriptor's closed input schema and returns the canonical schema_version / operation / resource_id / revision_id / data envelope. A revision_id identifies the complete observable resource state, not only workbook cell-content bytes: recalculation, evaluation/provenance state, or lifecycle metadata can advance it without a content edit. Clients must use the revision from the latest response envelope for compare-and-swap, cursor, and freshness decisions rather than computing or retaining a content hash. Static MCP annotations report the descriptor's worst-case risk; descriptions identify action-specific risk. close_workbook is excluded because cache eviction is runtime administration.

For the complete operation list and contracts, see Canonical Operation Surface. To add the legacy 0.13 names for one compatibility window, set SPREADSHEET_MCP_SLIM_SURFACE=false; shared names use their legacy schema and envelope in compatibility mode and are registered only once.

JS SDK

agent-spreadsheet-sdk is the app-facing integration layer — the one surface JS/TS code should target.

It provides generated canonical convenience methods, legacy input aliases, unchanged canonical envelopes, and typed capability errors. The complete 31-operation manifest supplies schemas and methods, while availability always comes from the selected live backend.

Backends are a configuration choice, not separate APIs:

• MCP backend — connect to a running agent-spreadsheet-mcp server (shared state, forks, multi-client)
• Embedded WASM backend — the engine runs in-process on workbook bytes, no server or filesystem required (browser, Node, serverless). The agent-spreadsheet-wasm crate in this repo is that backend's build artifact; it is an internal dependency of the SDK, not a package you consume directly.

Install:

npm i agent-spreadsheet-sdk

Backend status: MCP backend is stable and negotiates operations through live tool discovery (or an explicit supportedOperations list); the embedded WASM backend derives support from its generated binding's operations() descriptors and is tested against the real wasm-bindgen Node package in CI.

Recalc backends

Formula recalculation is pluggable.

Feature notes:

• recalc-formualizer is enabled by default
• recalc-libreoffice is available for LibreOffice-backed builds
• read and many write flows still work without recalc; only recalculate itself requires a backend

Docker images

Published at ghcr.io/psu3d0/agent-spreadsheet-mcp:

Examples:

# Read-onlydocker run -v /path/to/workbooks:/data -p 8079:8079 ghcr.io/psu3d0/agent-spreadsheet-mcp:latest
# Write + recalcdocker run -v /path/to/workbooks:/data -p 8079:8079 ghcr.io/psu3d0/agent-spreadsheet-mcp:latest-full

Workspace layout

agent-spreadsheet/├── crates/│   ├── agent-spreadsheet/        # shared engine + asp / agent-spreadsheet CLI│   ├── agent-spreadsheet-mcp/        # MCP server adapter│   └── agent-spreadsheet-wasm/   # experimental WASM-facing wrapper├── npm/│   ├── agent-spreadsheet/      # npm CLI wrapper│   └── agent-spreadsheet-sdk/    # JS SDK├── docs/                       # architecture and design docs├── benchmarks/                 # scenario budget regression harnesses└── .github/workflows/          # CI, release, docker builds

Package roles

Architecture notes

[Architecture Overview]

Core ideas:

• one semantic core shared across CLI, MCP, session, and WASM-facing work
• region detection for structural awareness
• token-efficient defaults so agents do not over-read spreadsheets
• verification as a first-class feature rather than an afterthought
• workflow helpers for the common mutations that spreadsheet agents repeatedly struggle with

Token-efficient workflow reference:

[Token Efficiency Workflow]

Recommended progression:

1. discover workbook + sheets
2. detect regions / table-like structures
3. inspect only the exact region or cells needed
4. mutate with dry-run or session staging
5. recalculate if needed
6. verify proof and review grouped diffs

Development

# Build everythingcargo build --release
# Run formatting, lint, and testscargo fmt --all -- --checkcargo clippy --workspace --all-targets -- -D warningscargo test --workspace

Local MCP iteration:

WORKSPACE_ROOT=/path/to/workbooks ./scripts/local-docker-mcp.sh

Or point your MCP client directly at the local binary:

{  "mcpServers": {    "spreadsheet": {      "command": "./target/release/agent-spreadsheet-mcp",      "args": ["--workspace-root", "/path/to/workbooks", "--transport", "stdio"]    }  }}

Read more

• CLI package README: npm/agent-spreadsheet
• Core crate README: crates/agent-spreadsheet
• MCP crate README: crates/agent-spreadsheet-mcp
• JS SDK README: npm/agent-spreadsheet-sdk
• WASM wrapper README: crates/agent-spreadsheet-wasm
• Packaging/versioning notes: docs/PACKAGING.md
• Heuristics and region detection: docs/HEURISTICS.md
• Recalc architecture: docs/RECALC.md

License

Apache-2.0