Standard MCP tools

The standard server exposes task-shaped tools rather than copying every REST endpoint one for one.

Account and usage tools

Feedback tool

Use send_feedback once the issue is clear. Include the situation, what happened and what should have happened. If known, add the related action, AI model and agent harness. Do not include API keys, tokens, signed URLs or full file contents. It costs no credits and does not change a translation.

Account and usage tools are read-only and consume no credits.

Translation tools

The result-link tool defaults to a temporary download URL. For text formats under 2 MB, pass format: 'content' to receive the translated text in the tool response. update_batch changes housekeeping metadata only and is reversible when archiving.

Tools that translate

These tools are the difference between an agent that reads your workspace and one that can do the work. The sequence is fixed (create, plan, approve), and only the third step costs anything.

Interactive review with MCP Apps

After creating or selecting a batch, call show_translation with its batch_id. Compatible hosts can display the file, plan, included/excluded content, measured Standard/Lite prices, workspace balance, progress and result. Other clients receive ordinary structured batch data and can continue using the existing tools.

The view reads and approves through the same authenticated MCP connection. Viewing and refreshing cost no credits. Approval needs batches:create, a current verified plan, an entitled quality tier and sufficient credits. Nonempty custom guidelines require entitlements.custom_guidelines; the maximum is 2,000 characters. The user must accept any listed exclusions before launching from the view.

create_translation (with an instruction), plan_translation and get_batch return plan_revision while a reviewable plan is available. Pass that value as expected_plan_revision to approve_translation: the field is required, and it binds the approval to the plan the user reviewed. If the plan has changed since, plan_changed refuses the launch; read the new scope and price and show them to the user again. REST approval accepts the same field and still works without it, but then skips this check.

The view refreshes active work with a bounded wait and offers a download only after completion. A sample is labelled as a sample. A missing/expired retained file remains an explicit error; refreshing cannot recreate deleted bytes. No UI capability is assumed in a host that does not implement MCP Apps. The AI Glot web app remains available as a fallback.

Sending a local document

The MCP client must be able to read the original file bytes and make an HTTP PUT request outside MCP. Connecting an assistant to AI Glot does not automatically give it access to its attachments or an HTTP client. AI Glot Code Mode cannot perform this transfer itself.

1. Call create_upload_url with the bare filename and exact size_bytes.
2. Send the original bytes to upload_url with its returned method and headers. The Authorization header is a temporary upload-only secret. Never substitute a workspace API key or OAuth token, follow redirects, log it, or show it to the user.
3. After a successful upload, call create_translation with upload_id and optionally instruction, on the same AI Glot connection. Exactly one of upload_id, content and file_url is accepted. Omit filename with upload_id.
4. Review the verified plan and plan.not_included, then check that the user has authorized this plan and credit cost before calling approve_translation.

The upload secret permits one PUT attempt within 10 minutes. Create the translation within one hour. If a PUT response is lost, try create_translation with the same upload_id first. If no complete upload exists, obtain a new session. Repeating creation recovers the same translation and never applies a changed instruction implicitly; use plan_translation to refine it. Recovery by upload_id is available until session cleanup, at least 24 hours after issuance. Keep the batch ID to retrieve the translation from your history afterwards.

If the runtime cannot read the local attachment or send HTTP PUT requests, explain the limitation immediately. For example: “This assistant can connect to AI Glot, but cannot send this local attachment. Upload the document in the AI Glot web app, use the AI Glot CLI, or provide an HTTPS download link that AI Glot can fetch.” Do not claim the document was uploaded, ask the user for their API key, or approve a translation without a verified file and plan.

A document already hosted at an accessible HTTPS download URL can still be sent as file_url. Small text files can be sent verbatim as content. See temporary uploads for the HTTP request and recovery rules.

Glossary tools

edit_glossary_terms is additive by default. replace_all: true deletes every term absent from the supplied map. Both glossary-editing tools are marked destructive because they can remove terms, so supporting clients can require confirmation.

Language resource

Supported languages are exposed as the cacheable MCP resource:

aiglot://languages

It is exposed as a resource because it is static reference data. Read it before creating a language-pair glossary.

Tool results and errors

Every tool declares an input schema and returns structured content plus a readable text representation. The standard server does not declare separate output schemas. Errors are returned as readable tool results with a stable code and documentation URL, allowing the model to correct the request instead of treating a business error as a broken MCP connection.