diff --git a/apps/maple-agent/docs/i18n.md b/apps/maple-agent/docs/i18n.md
new file mode 100644
index 000000000..124f3f4f8
--- /dev/null
+++ b/apps/maple-agent/docs/i18n.md
@@ -0,0 +1,113 @@
+# Agent localization (i18n)
+
+Design and plan for localizing the Maple Agent GPUI application. Status:
+proposal for review; no code has been extracted yet.
+
+## Current state
+
+The app is English-only with no localization infrastructure:
+
+- No i18n dependencies (`fluent`, `gettext`, `rust-i18n`, or similar) anywhere
+  in `apps/maple-agent`.
+- No language or locale preference in `app/src/settings.rs`.
+- User-facing strings are hardcoded English literals inline in the GPUI UI
+  modules, roughly 377 `text(...)`/string call sites across `app/src/ui/` and
+  its subdirectories (`chat/`, `settings/`, `remote/`, `markdown/`,
+  `store_gate/`, ...). Some are not user-facing (markdown code-language
+  labels, debug strings), so the real translatable surface is somewhat
+  smaller.
+
+## Decision: Fluent
+
+Localization will use the Mozilla Fluent stack (`fluent` +
+`fluent-templates`) rather than `rust-i18n` or gettext.
+
+The deciding factor is scope: we target roughly twenty languages, not two or
+three. At that scale Fluent's strengths are decisive:
+
+- **Correct plural and grammar rules.** Fluent applies CLDR plural
+  categories per language. Languages in our target set with three or more
+  plural categories (for example Russian, Polish, Arabic) cannot be handled
+  by `rust-i18n`'s two-form `_one`/`_other` convention without hand-rolled
+  logic per message. Retrofitting grammar later means reworking every
+  pluralized message, which is the expensive direction.
+- **Translator workflow.** `.ftl` files carry comments and per-message
+  metadata, are edited without touching Rust, and have an existing tooling
+  ecosystem (checkers, editor support) built for localization teams.
+- **Runtime locale switching is first-class** rather than a bolted-on
+  global.
+
+Trade-offs we accept:
+
+- More integration code than `rust-i18n`: we own bundle loading, locale
+  negotiation, and fallback chains (section 3 makes this a single small
+  module).
+- No compile-time key checking in either candidate; both fail soft at
+  runtime, so we add the catalog completeness test in section 5.
+- Slightly higher dependency and build cost.
+
+Gettext was rejected for the `.po` build-time workflow overhead and weaker
+fit for GPUI's inline string construction.
+
+## Architecture
+
+### Catalogs
+
+- `apps/maple-agent/app/locales/{locale}/app.ftl`, one file per locale with
+  identical key sets. `en` is the fallback and the source of truth.
+- Start with `en.ftl` only. Other locales fill in incrementally and fall
+  back to `en` at runtime; the app never ships blank strings.
+- Keys are namespaced per module: `login-title`, `settings-privacy-header`,
+  `remote-pairing-cancelled`, ...
+
+### Lookup layer
+
+Widgets do not touch `FluentBundle` directly. A new `app/src/ui/i18n.rs`
+provides:
+
+- `tr("key") -> SharedString` and `tr_args("key", ...) -> SharedString` for
+  parameterized messages, so call sites stay one-liners inside GPUI layout
+  expressions.
+- A swap-once bundle handle so changing locale is cheap and GPUI re-renders
+  naturally (text resolves at render time).
+- Fallback: active locale → `en` → the key itself, with a log line on every
+  fallback so misses surface in development instead of showing blank UI.
+
+### Locale resolution
+
+- Default from the OS locale at startup.
+- Overridable in Settings, stored next to existing preferences in
+  `app/src/settings.rs`.
+
+## Extraction plan
+
+Incremental, per module; the app stays shippable after every step.
+
+1. **Infrastructure PR (this plan's first code change):** `i18n.rs` lookup
+   layer, `en.ftl` skeleton, catalog completeness test, and one module
+   converted end to end as the proof.
+2. **Module-by-module extraction,** roughly in order: `settings/`,
+   `login.rs`, `chat/`, `remote/`, `store_gate/`, remaining `ui/` modules.
+   Each PR moves its module's literals into `en.ftl` and nothing else.
+3. **Locale enablement:** add target locales as translations arrive; the
+   fallback chain means partial catalogs are safe to ship.
+
+## Guardrails
+
+- **Catalog completeness test:** parse `en.ftl`, assert every other locale's
+  key set matches (missing key fails, extra key warns).
+- **Fluent syntax check** in CI so a broken placeholder in a translation
+  fails the lane rather than a user's screen.
+- **Hardcoded-string gate:** a test or CI check flagging new user-facing
+  string literals in `app/src/ui/` to prevent drift. Imperfect by design;
+  it catches the common case.
+
+## Open questions
+
+- **Tray and native menus** (`app/src/tray.rs`, `app/src/ui/menus.rs`) go
+  through platform APIs rather than GPUI text. Do they count as in-scope for
+  the first pass?
+- **Non-GPUI surfaces:** `hosted_oauth_complete.html` is a standalone HTML
+  page; it needs its own mechanism or an explicit out-of-scope decision.
+- **Catalog location:** app crate (assumed here) versus shared resources, if
+  other components ever consume the same catalogs.