Localization your agents can run end to end

Give your coding agent the same Glossa your team uses: push files, search strings, work the review queue, run audits and add rules, over MCP, the CLI or REST.

claude · glossa mcp
Work through the Glossa review queue. Split merges whose context differs.
claim_tasks({ "kind": "merge_review", "limit": 10 })
10 tasks · "Open" used by 3 keys: menu.file.open, door.status.open, store.hours.open
menu.file.open is a verb; the other two are adjectives. Splitting it out.
resolve_task({ "task": "8c1e…", "action": "split", "key_ids": ["menu.file.open"], "note": "verb vs adjective" })
Split into a new unit; queued translation for 9 locales.
run_audit({ "locale": "es", "question": "Are we consistent about tú vs usted?" })
Audit queued. Findings arrive as audit_finding and rule_suggestion tasks.
An agent claiming review tasks, splitting a merged string and starting an audit through the MCP server.

What makes it agent-friendly

One tool per op
Every dashboard action is an MCP tool with a description written for an agent and read-only hints, so clients can auto-approve reads.
Small decisions, batched
claim_tasks hands out a batch with everything needed to decide. Parallel agents never collide; claims expire after 30 minutes.
Guardrails in the API
Translations that change placeholders are rejected. AI never overwrites reviewed, approved or imported translations. Every write is logged with its surface.
Readable by machines
llms.txt, llms-full.txt, agents.md, OpenAPI 3.1 and an MCP server card describe the whole API, generated from the same registry.

Connect in one command

Claude Code, with a workspace API key:

claude mcp add --transport http glossa https://glossa-meterapp.vercel.app/api/mcp \  --header "Authorization: Bearer gls_…"

Codex, Cursor, OAuth and the plugin are in the MCP guide. Prefer plain HTTP? Every tool is also POST /api/v1/{op}.

Push your first file

Glossa is free for one app. Sign in, push a translation file, and pull it back in every language.