Best AI Agent Skills for Translators in 2026
A. Frans
Published April 29, 2026
Table of Contents
- 01The eight skills, by job
- 021. docx-word, the file format every translator hates
- 032. pdf-tools, the worse format every translator hates more
- 043. xlsx-spreadsheet, for the bilingual glossary
- 054. firecrawl-skill, clean scrape for web content
- 065. deep-research, terminology lookup with sources
- 076. writing-style-skill, keep client voice consistent
- 087. brand-guidelines, long-memory client preferences
- 098. mcp-language-server, bilingual review pass
- 10Stack for solo freelancers
- 11Stack for small agencies (3-15 translators)
- 12Security note on community skills
- 13What's missing in the ecosystem
- 14FAQ
# Best AI Agent Skills for Translators in 2026
Translators have been quiet beneficiaries of the AI agent skill movement. Claude Code skills don't replace CAT tools like Trados or memoQ. They wrap around them, handling the file munging, glossary lookups, and quality checks that used to eat half a day per project.
This is the working stack I see freelance translators and small LSPs (language service providers) install in 2026.
The eight skills, by job
| Skill | What it does | Source | Install |
|---|---|---|---|
| docx-word | Parses .docx files preserving formatting | Anthropic | /plugin install docx-word |
| pdf-tools | Extract + reflow PDF text for translation | Anthropic | /plugin install pdf-tools |
| xlsx-spreadsheet | Bilingual glossary management in .xlsx | Anthropic | /plugin install xlsx-spreadsheet |
| firecrawl-skill | Scrape source web pages cleanly | Mendable | /plugin install firecrawl-skill |
| deep-research | Domain terminology research | Community | /plugin install deep-research |
| writing-style-skill | Style-guide enforcement | Community | /plugin install writing-style-skill |
| brand-guidelines | Client tone/voice memory | Community | /plugin install brand-guidelines |
| mcp-language-server | LSP-style bilingual review | Community | /plugin install mcp-language-server |
1. docx-word, the file format every translator hates
Eighty percent of incoming translation work arrives as .docx. Eighty percent of CAT tool segmentation errors trace back to .docx parsing. The docx-word skill from Anthropic handles paragraph styles, tracked changes, comments, and embedded tables without losing formatting metadata.
The right workflow: import the .docx with this skill, extract the segments to a JSON working file, translate in your CAT tool of choice, then merge back. The merge step is historically painful but works here because the skill keeps a positional index of every text node.
Install: claude /plugin install docx-word. Verified by Anthropic.
2. pdf-tools, the worse format every translator hates more
Scanned PDFs need OCR before translation can begin. Native PDFs (text-based) need careful extraction so paragraph breaks don't fracture mid-sentence. The pdf-tools skill does both, with optional Tesseract integration for OCR if you have it installed locally.
Where it shines: legal contracts, immigration documents, medical reports. The PDFs translators charge premium rates for because the file format is hostile.
Where it struggles: PDFs with two-column layouts where columns interleave on extraction. You'll still need a manual cleanup pass for those.
GitHub: github.com/anthropics/skills/tree/main/document-skills/pdf-tools.
3. xlsx-spreadsheet, for the bilingual glossary
Glossaries live in spreadsheets. Always have. The xlsx-spreadsheet skill lets Claude read your existing termbase, suggest new entries from the source text, and check translation consistency against approved terminology, all without leaving the editor.
Useful pattern: dump the source segment, ask Claude to surface glossary terms that match, then validate the translation reuses them. For technical translation (medical devices, IT, legal) this is the difference between a passing review and a failed one.
GitHub: Anthropic-published. Free.
4. firecrawl-skill, clean scrape for web content
Websites translate poorly when you copy-paste them. Navigation, footers, cookie banners all leak in. Firecrawl pulls just the article body in markdown, ready to translate.
Use case: translation of client landing pages, marketing blog posts, product documentation hosted in their CMS. Get a clean text file out, translate normally, hand back markdown or HTML for the client's CMS team.
Pricing model: Firecrawl itself has a free tier (500 pages/month) and a paid plan ($16/month for 3K pages). The Claude skill is free; the underlying API is the cost.
5. deep-research, terminology lookup with sources
Specialist translation needs verified terminology. The deep-research skill spawns a sub-agent that searches academic sources, official glossaries (IATE, UN Termium), and authoritative documents to confirm a term choice with citations.
In practice: you flag a term you're unsure about. The skill returns 3-5 candidate translations with the publication that uses each. Pick one, log it in your glossary, move on. Saves the 20-minute Google rabbit hole.
6. writing-style-skill, keep client voice consistent
Translation isn't just word-swap. Each client wants their tone preserved: formal Spanish for legal contracts, conversational Spanish for tourism brochures. The writing-style-skill stores per-client style profiles and validates output against them.
The right setup: build a style profile per client by feeding the skill 5-10 of their existing translated documents. Future translations get scored against that profile and flagged when tone drifts.
GitHub: community-maintained. Audit before installing (covered below).
7. brand-guidelines, long-memory client preferences
Different from style-skill. Where writing-style-skill checks tone, brand-guidelines stores hard rules: "always translate 'environment' as 'medio ambiente' not 'entorno'", "use European Spanish, not Latin American", "abbreviate company name to ACME after first mention".
These are the rules clients ask about every project and freelancers forget every project. Claude reads them at the start of each session.
8. mcp-language-server, bilingual review pass
Borrowed from the IDE world. The mcp-language-server skill cross-references your translation against the source on a sentence-by-sentence basis, flagging missing segments, mistranslated numbers, and untranslated proper nouns.
Not a replacement for human QA. A replacement for the rough first review pass that's mostly mechanical.
Stack for solo freelancers
If you're a one-person operation translating 50K-200K words a month:
- docx-word + pdf-tools (file handling)
- xlsx-spreadsheet (glossary)
- writing-style-skill (one client, one profile)
- brand-guidelines (rules per client)
Skip firecrawl unless you regularly translate web content. Skip deep-research unless you're in specialized verticals.
Total install time: under 30 minutes. Total cost: free except Firecrawl if you add it.
Stack for small agencies (3-15 translators)
Add to the freelancer stack:
- deep-research for terminology validation across the team
- mcp-language-server for the project manager's pre-delivery QA
- A shared glossary repository (Git or Dropbox) so everyone references the same xlsx files
The shared glossary part isn't a skill. It's a workflow choice that makes the skills useful across the team.
For broader tooling beyond skills, see [our full list for translators](/best-ai-tools-for/translators).
Security note on community skills
Three of the eight skills here are community-maintained: writing-style-skill, brand-guidelines, mcp-language-server. Community doesn't mean unsafe, but it does mean you should:
1. Read the SKILL.md before installing. It tells you what tools and permissions the skill uses. 2. Check the GitHub repo for an active maintainer (commits in the last 90 days). 3. Audit any skill that accesses files outside the project directory. 4. Pin to a specific version in production rather than auto-updating.
Anthropic-published skills (docx-word, pdf-tools, xlsx-spreadsheet) get the same audit as Anthropic's other code; they're safer by default but the same review hygiene applies.
For deeper guidance, see our skill security review series.
What's missing in the ecosystem
A few gaps I'd love filled by mid-2026.
CAT tool MCP servers. Trados and memoQ both expose APIs but no Claude skill wraps them. Translators have to manually export/import, which defeats half the productivity gain.
Termbase sync. MultiTerm, MultiTerm Online, and TermBase eXchange (TBX) format support is missing. We're stuck doing xlsx round-trips.
Quality scoring. No skill scores a translation against MQM or DQF metrics yet. Manual review still.
If you're building skills, these are the openings.
FAQ
Can these skills replace a CAT tool? No, and they're not trying to. CAT tools handle translation memory, segmentation, and project management at scale. Claude skills handle the messy edges around them.
Do I need Claude Code, or does the API work? Claude Code is where skills live today. The Anthropic API doesn't natively load skills, though the Agent SDK can simulate the same patterns programmatically.
Will using AI translation skills get me banned by my LSP? Depends on the LSP and the contract. Some prohibit AI in translation outputs, others require disclosure. Read your MSA. Skills that handle file munging or QA are usually fine; skills that generate translation drafts may not be.
What about Indonesian-English translation specifically? The skills above all work with Indonesian. Claude's Indonesian fluency is strong as of 2026. Quality varies more on niche domains (legal, medical) where deep-research helps validate terminology.
Are there free alternatives to Trados or memoQ that pair with these skills? OmegaT and Smartcat (free tier) both work with the file outputs from docx-word and pdf-tools. Less polished than Trados but pairs well with the skills above for solo translators on a tight budget.
Share this article
⚙Related Tools
📄Related Articles
Get More AI Tool Guides
New comparisons and guides every week. Join thousands of professionals staying ahead of the AI curve.