Honest pricing

Pennies per file.
Not $27 per month per seat.

locAGI runs on your own AI keys (OpenAI, Anthropic, Google, DeepSeek, Mistral — or a free local Ollama model). You pay the AI vendor at wholesale rates. We charge nothing for the workbench itself. The numbers below are real.

Side-by-side — the same 1,500-word document

locAGI Phrase Personal memoQ Translator Pro Trados Studio Smartcat
Cost to translate one 1,500-word doc with AI ~$0.02 – $0.40 $27/mo flat (5k AI words/mo) €0 (no AI; bring plugin) €0 (no AI; plugin extra) Metered, ~$0.06/word
Up-front license Free (MIT) Subscription only €620 one-time €695 – €995 Free tier exists
AI baked in (no plugin install) Yes Add-on tier No (manual MT) Plugin required Yes
Self-hosted (your firewall, your machine) Yes No Desktop only (Windows) Desktop only (Windows) No (cloud only)
Linux / macOS native Yes (browser) Yes (browser) Wine workaround Windows only Yes (browser)
Real-time per-call cost visibility Yes — chip on every screen Billed monthly n/a n/a Wallet only
Per-model price comparison before you commit Yes No No No No
Open exports (XLIFF, TMX, TBX, MQXLIFF, SDLXLIFF) Yes Yes Yes Yes Yes
AI translation QA (MQM-aligned) Yes Add-on Manual Plugin / add-on Yes
Glossary / termbase injection into AI prompt Yes Limited Yes Yes Yes
Multilingual project view Yes Yes Yes Yes Yes
Your data leaves your machine? Only to the AI vendor you chose Yes (Phrase cloud) No (desktop) No (desktop) Yes (Smartcat cloud)

Pricing data collected 2026-05-20 from each vendor's public site — check theirs for current rates. locAGI's "pennies per file" figure assumes the user picked a mid-tier model (DeepSeek V4-Pro or Gemini 2.5 Flash) and ran AI Improve + Critic on every segment. Cheaper if you pick Gemini Flash; more if you pick GPT-5 or Claude Opus.

What the AI actually costs — live demo

Below is a realistic spend for translating one 1,500-word document through locAGI's Studio workflow (AI Improve on every segment, Critic pass, exporting to MQXLIFF). Numbers are projected from the catalog pricing we surface in the analysis modal.

1,500-word document · ~90 segments · Improve + Critic

Input tokens · ~9,500 tokens
Output tokens · ~9,500 tokens
DeepSeek V4-Pro   $0.30/M in · $0.85/M out$0.011
Gemini 2.5 Flash   $0.075/M in · $0.30/M out$0.004
GPT-4o-mini   $0.15/M in · $0.60/M out$0.007
Claude Sonnet 4.5   $3/M in · $15/M out$0.171
GPT-5   $5/M in · $20/M out$0.238
Ollama (your own GPU)   localFREE
Cheapest paid option (Gemini 2.5 Flash)$0.004

What we don’t pretend to have yet

Honest about the gaps

  • No formal reviewer/approver roles beyond admin (Trados Studio has finer-grained workflow gates).
  • No enterprise SSO yet (planned). Standard sessions + API keys today.
  • TM cleanup wizard is on the roadmap, not yet shipped.
  • Phrase has a wider hosted-MT marketplace; we route to whichever model you connect, but we don’t resell.
  • The desktop CAT tools (memoQ, Trados) are more battle-tested for very large XLIFF files (50k+ segments). We’ve tested up to ~10k.

Frequently asked questions

What does “BYOK” mean for my bill?

You connect your own OpenAI / Anthropic / Google / DeepSeek / Mistral / xAI / OpenRouter / Ollama-Cloud key in Settings. locAGI sends the API call straight to that vendor; your card on file with them is the one charged. We don’t add a markup, see your bill, or have a billing relationship with you at all.

Can I run it fully offline?

Yes — connect a local Ollama server, pick a local model, and the AI calls never leave your machine. Free in dollar terms (you pay in GPU watts). Great fit for regulated content where data residency matters.

What if my AI vendor raises prices?

You swap to a cheaper one in Settings — no migration, no lost work. The per-model cost projection on the analysis modal lets you see the impact of the change before you commit. (This is the wedge: competitors lock you into their chosen MT vendor.)

Is the source code open?

Yes — MIT licensed. You can self-host on a $4/mo VPS, fork it, run it air-gapped inside a SCIF, or paste an LLM response on top of someone else’s deployment. Source on GitHub.

How accurate is the per-model cost projection?

Within roughly ±20%. We use vendor-published per-million- token rates and a ~3.5 chars/token estimate. Reasoning models can spend extra hidden tokens (chain-of-thought) that don’t show in the input/output split — rare but worth knowing. Open the Studio analysis modal for the live numbers on a real project.

Try locAGI Studio →