DocProof

Your AI translates. DocProof checks supported content before writing a new Word file.

中文介绍 · Download Alpha · Quick start 中文 · Validation

Local, bring-your-own-AI translation guardrails. MIT licensed. No DocProof account or API key required for the companion/MCP workflow. Your chosen AI's access, limits and charges still apply.

Alpha, not an error-free translator. In authored regression fixtures, 20/20 correct preservation candidates passed unchanged and 14/14 seeded rule errors were flagged. All 6 tested semantic-only errors were missed. Three valid localized equivalents were also flagged by the strict policy. No live-model comparison or independent semantic evaluation has established that DocProof improves overall translation accuracy. Complete report.

Start without coding

Download the Windows x64 ZIP under the release's Assets, extract the whole archive, and double-click Start-DocProof.cmd. Node is included. The Windows interface is currently Chinese. GitHub's automatic “Source code” ZIP is for developers, not this portable app.

1. Select/drop a .docx, choose a language and your existing AI.

2. Click “复制任务并打开 AI” (copy task & open AI), paste/send in your AI chat, then copy the reply and click “从剪贴板导入回复” (import reply). Repeat for the next batch or requested repair.

3. Review source and translation side by side, acknowledge any untouched exclusions, and save a new Word file. Open it in Word and check meaning and layout.

Progress is saved locally between restarts. JSON is in an optional details panel; a single fenced JSON answer with ordinary surrounding explanation is accepted. Ambiguous/multiple payloads are refused. Unchanged names can be explicitly approved; missing numbers cannot be bypassed that way.

The companion does not auto-send messages, watch your clipboard, read browser cookies, or attach itself to arbitrary AI websites. Copy/import is deliberate. This is a Windows desktop app, not a mobile app.

Related MCP server: Project Files Read-only MCP

Let your AI use the tools directly

Use a dedicated folder of document copies, not your entire home directory. Ask your AI:

Use DocProof to translate manual.docx into Chinese. First disclose unsupported content. Preserve numbers, units and equipment IDs. Repair only failed segments, then save manual.zh.docx for my review.

Tools: begin_document, get_job, submit_batch, finalize_document, forget_job. The stdio server makes no model calls itself and opens no web port. MCP jobs are in memory; desktop resume does not apply to MCP host restarts. If a host rejects the extension, use the companion/manual configuration—do not disable security controls.

Source development (Node.js 22+):

npm ci --ignore-scripts
node mcp-server.mjs --root /absolute/path/to/document-copies

Codex CLI, using actual absolute paths:

codex mcp add docproof -- node /absolute/path/to/docproof/mcp-server.mjs --root /absolute/path/to/document-copies

Official Codex configuration · MCPB format · Cursor install links

Protected—and not protected

• Bounded batches, job IDs, targeted retries, missing/duplicate segment detection.

• Strict recognized numbers/units/identifiers, protected terms and glossary constraints. Valid localization/conversion may be flagged; the tool does not silently rewrite candidates.

• Source hash, strict XML parsing in the guarded writer, structural masking and output readback. Untouched package-entry payloads remain byte-identical; not a promise of byte-identical ZIP archives or pagination.

• New output and receipt only; refuse to overwrite existing files. The source is not modified by the workflow.

• Simple uniformly styled paragraphs/cells in supported parts. Mixed styles, fields, links, revisions, text boxes, content controls and other complex content are excluded and disclosed. Images are preserved, not translated.

Not covered: semantic equivalence, factual truth, within-paragraph omissions without protected evidence, negation/order/entity correctness, full OOXML schema validation, actual Word rendering, or complete translation of excluded content. Generation has an experimental draft evidence checker, not an end-to-end Word generation UI.

Limits: .docx ≤10 MB, ≤300 editable paragraphs, ≤40,000 source characters, ≤3,000 characters per paragraph; at most three valid-format attempts per batch. Start with the synthetic sample, not important documents. Do not use unattended for high-stakes delivery.

Verify and contribute

npm test
npm run eval

Windows UI check:

powershell.exe -NoProfile -STA -ExecutionPolicy Bypass -File desktop/Start-DocProof.ps1 -SelfTest

Builds: npm run build:portable (Windows x64, bundles current Node and license); npm run build:mcpb (bundles dependencies, host supplies Node 22+). Explicit public-file allowlists exclude user data. The optional independent DeepSeek API adapter is not the default UI route; networking has only mocked tests.

Legacy audit: node cli.mjs audit source.docx target.docx --html report.html --json report.json. Its XML inventory is heuristic, distinct from the guarded writer. Its PASS label is not semantic approval.

Validation · Privacy/uninstall · Contributing · Security · MIT