Excel LSP

An LSP for Excel: semantic index + MCP server so AI agents navigate workbooks by symbols, references, and diagnostics — not by reading 50,000 rows.

(LSP-style: the ideas — symbols, references, diagnostics, incremental index — not the LSP wire protocol.)

Measured release evidence: Excel LSP used 3,410 deterministic tool-result tokens versus 222,289 for naive dump—a 65.2× reduction—and answered 12/12 headless runs exactly versus 8/12. Full Codex CLI usage, which includes fixed agent context and schemas, is reported separately in the raw rows.

60-second lineage demo

The demo is assembled only from the numbered desktop-Excel captures. Its evidence manifest records every source hash, the output hash, dimensions, and frame duration; the complete live protocol index records the matching machine-readable assertions.

Related MCP server: MCP Excel Server

Install with your AI agent

Copy this prompt into Codex or another coding agent:

Download and install Excel LSP from https://github.com/Brody900/excel-lsp,
configure it as a local MCP server for this environment, and verify that the
server starts and exposes its tools. Follow the repository's Quickstart,
prefer the git-based uvx command until the PyPI package is available, preserve
unrelated project files, and report exactly what you changed and verified.

The agent can use the commands below directly. Excel LSP runs locally over stdio; installing it does not upload workbook data or start a network daemon.

Quickstart

Excel LSP requires Python 3.11 or newer and runs as a local stdio process. It does not require Microsoft Excel for indexing or editing OOXML files.

Run the MCP server

The verified public-repository install works now:

uvx --from git+https://github.com/Brody900/excel-lsp@main excel-lsp serve

After the PyPI publication is visible, the shorter equivalent is:

uvx excel-lsp serve

Add Excel LSP to Codex

codex mcp add excel-lsp -- uvx --from git+https://github.com/Brody900/excel-lsp@main excel-lsp serve
codex mcp get excel-lsp

Codex will launch the stdio server when it needs it; Excel LSP is not a network daemon. The shorter PyPI registration is codex mcp add excel-lsp -- uvx excel-lsp serve. That syntax was verified with Codex CLI 0.144.5 in an isolated configuration home.

Configure Codex manually

Add the following entry to ~/.codex/config.toml:

[mcp_servers.excel-lsp]
command = "uvx"
args = ["--from", "git+https://github.com/Brody900/excel-lsp@main", "excel-lsp", "serve"]

A copy is available at examples/codex.config.toml.

{
  "mcpServers": {
    "excel-lsp": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/Brody900/excel-lsp@main", "excel-lsp", "serve"]
    }
  }
}

This is a generic MCP-client example, not Codex's native configuration format. It is also committed as examples/mcp.json.

Tools

The contracts below are frozen for v0.1.0 and verified by the completed P7 milestone. Exact generated schemas and one worked example per tool are in the tool reference.

The first 12 tools are read tools. The final two are destructive write tools and carry the corresponding MCP annotations so clients can request confirmation.

Architecture

The workbook remains authoritative; the SQLite sidecar is a disposable derived index. Every transport operation delegates to the same core services.

flowchart LR
    Workbook["Excel OOXML workbook"] --> Loader["Streaming OOXML loader"]
    Loader --> Index["Persistent semantic index<br/>SQLite + R*Tree or interval fallback"]
    Index --> Regions["Regions and symbols"]
    Index --> Graph["Formula dependency graph"]
    Graph --> Diagnostics["Diagnostics"]
    Regions --> MCP["MCP server<br/>14 bounded tools"]
    Graph --> MCP
    Diagnostics --> MCP
    MCP --> Codex["Codex and other MCP clients"]
    MCP --> Editor["Surgical OOXML editor<br/>2 write tools"]
    Editor --> Workbook

See the architecture for implemented boundaries and phase status.

Benchmarks

The verified P8 milestone commits the raw rows, exact/set-semantic answer checks, both headless-Codex repetitions, environment metadata, and scripts that regenerate every chart. The raw results index explains each artifact and the methodology documents isolation, grading, cost guards, and the optional-arm exclusion.

Results

Excel LSP meets S5 on its defined deterministic payload metric: 3,410 tokens versus 222,289, a 65.2× reduction, with equal-or-better headless accuracy. The disclosed 1,000-row archive workload is identical across arms and every original OOXML member stays byte-identical except package declarations and F03's deliberately extended Summary XML; a regression separately proves every pre-existing Summary cell, formula, and cache is unchanged. Mean full Codex usage was 77,310.5 versus 64,909.8 because that secondary measure includes fixed agent context, schemas, and reasoning; it is reported rather than conflated with workbook payload. See the criterion calculation and per-repetition table.

The 50,000-row median is 9.440 seconds cold and 0.066 seconds after a one-sheet change, satisfying S1's strict 10-second and 1-second limits.

Reproduce

Run the deterministic twelve-row replay with:

excel-lsp bench

For a fresh headless run and regenerated timing/charts, follow the commands in the benchmark methodology. Headless runs consume account capacity; use a new JSONL path instead of overwriting the committed evidence.

Comparison

This is a capability comparison, not an overall ranking. Upstream observations are limited to each project's README at a pinned revision accessed 2026-07-29; “not documented” does not mean impossible. See the source notes and exact revisions.

How it works

P2 adds sparse region detection without constructing dense grids. Native Excel ListObjects take precedence; otherwise bounded header, type, style, merge, and density features produce a region and an explicit confidence score. See the architecture.

Verified P3 normalizes copied formulas into R1C1 signatures and groups matching cells into formula blocks. That design lets an agent reason about a large calculated column as one semantic unit while preserving cell-level references and anomalies. See the P3 evidence and index internals.

P4 stores formula range dependencies as rectangles in SQLite R*Tree, with a portable interval-table fallback. Point and range queries then feed bounded precedent, dependent, path, and diagnostic operations without expanding every range into individual edges. See index internals.

Security & scope

The local stdio server makes no runtime network requests and supports realpath-resolved workbook confinement.

The P6 core edit services surgically modify targeted worksheet XML and required calculation metadata. Complete F16/F21 part manifests and a property test prove that every OOXML part not deliberately modified stays byte-identical in the verified implementation. EXCEL_LSP_ROOT provides an optional os.pathsep-separated directory allowlist applied after realpath resolution. Its default is unrestricted local-path access. See SECURITY.md for the current threat model and implementation status.

Evidence

Start with the evidence index. It distinguishes verified artifacts from scope declarations and links the exhaustive README claims-to-artifacts matrix, clean-install report, and registry submission packet.

Not affiliated with Microsoft. Excel is a trademark of Microsoft Corporation.