diff --git a/.cursor/skills/ship-a-change/SKILL.md b/.cursor/skills/ship-a-change/SKILL.md
index 2319274..72b795c 100644
--- a/.cursor/skills/ship-a-change/SKILL.md
+++ b/.cursor/skills/ship-a-change/SKILL.md
@@ -44,7 +44,10 @@ Three details keep the app working from a repository sub-path. Do not undo them:
   straight from the browser on Pages, so the hosted site gets the same tool list as the dev server.
   Keep it that way: a tool that _only_ works behind the optional proxy cannot ship here. The proxy
   in `tools/agent-api.ts` is an extra path for DuckDuckGo search and page reads, mounted by
-  `pnpm dev` and `pnpm proxy`, never by the Pages build.
+  `pnpm dev` and `pnpm proxy`, never by the Pages build. The repository variable `AGENT_API_BASE`
+  can point the build at a hosted one, and both tools fall back to the browser path when it fails —
+  which is the only reason aiming every visitor at one process is defensible. Do not remove that
+  fallback to surface a proxy error.
 
 The Pages concurrency group deliberately does not cancel in-progress runs: cancelling mid-deploy can
 leave the site half-published.
diff --git a/.env.example b/.env.example
index fcbea16..073c320 100644
--- a/.env.example
+++ b/.env.example
@@ -12,7 +12,9 @@
 # providers from the browser. To send DuckDuckGo search and page reads through
 # the optional tool proxy instead, set this to the proxy origin or to
 # `same-origin` (which is what `.env.development` does for `pnpm dev`).
-# Leave it unset on GitHub Pages so the hosted site stays fully static.
+# Empty or unset stays browser-direct. A proxy that fails falls back to the
+# browser path, so this is an optimisation rather than a dependency.
+# For the Pages deploy this comes from the repository variable AGENT_API_BASE.
 
 # VITE_AGENT_API_BASE=same-origin
 # VITE_AGENT_API_BASE=http://localhost:8787
diff --git a/.github/workflows/deploy.yml b/.github/workflows/deploy.yml
index 86e316b..fb51a5a 100644
--- a/.github/workflows/deploy.yml
+++ b/.github/workflows/deploy.yml
@@ -51,6 +51,13 @@ jobs:
 
       - name: Build
         run: pnpm build
+        env:
+          # Optional. Set the repository variable AGENT_API_BASE to a tool proxy
+          # origin and every visitor's DuckDuckGo search and page reads go there
+          # without touching the Tools panel. Unset, the site stays
+          # browser-direct, which is what a fork with no proxy needs. Either way
+          # a proxy that fails falls back to calling the provider from the page.
+          VITE_AGENT_API_BASE: ${{ vars.AGENT_API_BASE }}
 
       - name: Add SPA fallback
         # Pages serves 404.html for unknown paths; making it the app shell keeps
diff --git a/README.md b/README.md
index 82c53c3..9d010de 100644
--- a/README.md
+++ b/README.md
@@ -304,10 +304,16 @@ GitHub Pages cannot host a process, so the published site stays browser-direct.
 - **`pnpm proxy`** — the same handlers on http://localhost:8787, for a static build or the hosted site. Paste that origin into **Tools → Tool proxy URL**, or build with `VITE_AGENT_API_BASE=http://localhost:8787`.
 - **Railway (or any host)** — the `Dockerfile` in the repo root runs only this process. Create a project, connect `devbadya/Jarvis`, set `PROXY_ORIGINS` to `https://devbadya.github.io`, wait until the deploy is live, then **Settings → Networking → Generate domain**. That `https://….up.railway.app` origin is the URL: paste it into **Tools → Tool proxy URL** on the hosted site. There is no URL until a domain exists; the Hobby plan alone does not create one.
 
+To give every visitor that proxy without asking them to paste anything, set the repository variable **`AGENT_API_BASE`** to its origin: `deploy.yml` passes it to the build as `VITE_AGENT_API_BASE`. Leave it unset and the hosted site stays browser-direct, which is what a fork with no proxy of its own needs — an empty value is not a proxy, so a workflow forwarding a variable nobody set cannot aim the build at an `/api` the host does not serve.
+
+**A proxy failure is not a failed turn.** Both tools try the proxy first and fall back to calling the provider from the page, so an outage, a spent budget or an allowlist that has not caught up costs a slower search rather than the answer. That fallback is what makes it safe to point every visitor at one process.
+
 The proxy scrapes DuckDuckGo HTML itself and fetches pages itself. It does not spend the Jina reader budget, and it is not limited to CORS-friendly endpoints. Wikipedia, LangSearch and Jina still leave the tab directly — they already send the headers, and their keys must not travel through this process.
 
 A fetch-on-behalf proxy is still a confused deputy. Every target is resolved and refused if it lands on loopback, link-local or RFC1918, and redirects are re-checked. Do not bind `pnpm proxy` to the public internet without setting `PROXY_ORIGINS` to the pages that may call it (for example `https://devbadya.github.io`). Inference never goes through it.
 
+The allowlist says who may call, not how often, and an allowed page is exactly what a scraper would forge. `pnpm proxy` therefore allows **30 requests a minute per caller** and answers `429` beyond that, counted from the forwarded address rather than the socket, since every edge terminates the connection itself. `PROXY_RATE_LIMIT` changes the number; `0` switches it off. Health checks are exempt.
+
 ### What leaves the browser
 
 Inference does not: prompts, reasoning, and replies never leave the GPU, and neither do [memories](#memory), which are written to IndexedDB in this browser and read back into a prompt that goes no further than the GPU either. Tools are the exception, and always were. A `web_search` call sends the query to the chosen provider, a `read_page` call sends the URL to the reader, a `weather` call sends the place name to Open-Meteo's geocoder and its coordinates to the two forecast services, and a `current_time` call with a place sends the name to the same geocoder.
diff --git a/src/tools/web.test.ts b/src/tools/web.test.ts
index 79e866b..18c0441 100644
--- a/src/tools/web.test.ts
+++ b/src/tools/web.test.ts
@@ -755,6 +755,20 @@ describe('configuredProxyBase', () => {
     )
     expect(configuredProxyBase({ provider: 'duckduckgo', proxyUrl: 'localhost:8787' })).toBeUndefined()
   })
+
+  // A workflow forwarding a repository variable nobody set hands the build an
+  // empty string. Reading that as same-origin would aim every hosted visitor at
+  // an `/api` the static host does not serve.
+  it('treats an empty build-time value as no proxy', () => {
+    vi.stubEnv('VITE_AGENT_API_BASE', '')
+    expect(configuredProxyBase({ provider: 'duckduckgo' })).toBeUndefined()
+
+    vi.stubEnv('VITE_AGENT_API_BASE', 'same-origin')
+    expect(configuredProxyBase({ provider: 'duckduckgo' })).toBe('')
+
+    vi.stubEnv('VITE_AGENT_API_BASE', 'https://proxy.example/')
+    expect(configuredProxyBase({ provider: 'duckduckgo' })).toBe('https://proxy.example')
+  })
 })
 
 describe('searchWeb with a tool proxy', () => {
@@ -779,9 +793,30 @@ describe('searchWeb with a tool proxy', () => {
     expect(lastRequest(fetchMock).body).toMatchObject({ region: 'de-de' })
   })
 
-  it('relays a proxy error rather than wrapping it', async () => {
-    stubFetch(jsonResponse({ error: 'Refusing to fetch a private or loopback address' }, 502))
-    await expect(searchWeb('x', 5, proxied)).rejects.toThrow(/private or loopback/)
+  // A hosted build points every visitor at one proxy, so its bad day must not
+  // become theirs. Browser-direct is the same path this build took before.
+  it('falls back to the reader when the proxy fails', async () => {
+    const fetchMock = stubFetch(
+      jsonResponse({ error: 'Origin not allowed' }, 403),
+      jsonResponse({
+        data: {
+          content: [
+            '1.[WebGPU](https://duckduckgo.com/l/?uddg=https%3A%2F%2Fwebgpu.org%2F&rut=1)',
+            'A GPU API for the web.',
+          ].join('\n'),
+        },
+      }),
+    )
+
+    const results = await searchWeb('webgpu', 3, proxied)
+
+    expect(lastRequest(fetchMock).url.href).toContain('r.jina.ai')
+    expect(results.at(0)).toMatchObject({ title: 'WebGPU', url: 'https://webgpu.org/' })
+  })
+
+  it('reports the reader failure when neither path works', async () => {
+    stubFetch(jsonResponse({ error: 'Too many requests' }, 429), jsonResponse({}, 500), jsonResponse({}, 500))
+    await expect(searchWeb('x', 5, proxied)).rejects.toThrow()
   })
 
   it('does not send Wikipedia through the proxy', async () => {
@@ -803,6 +838,18 @@ describe('readPage with a tool proxy', () => {
     expect(lastRequest(fetchMock).body).toEqual({ url: 'https://example.com/a' })
     expect(page).toEqual({ url: 'https://example.com/a', title: 'A', text: 'Body.' })
   })
+
+  it('falls back to the reader when the proxy fails', async () => {
+    const fetchMock = stubFetch(
+      jsonResponse({ error: 'Too many requests' }, 429),
+      jsonResponse({ data: { url: 'https://example.com/a', title: 'A', content: 'Body.' } }),
+    )
+
+    const page = await readPage('https://example.com/a', proxied)
+
+    expect(lastRequest(fetchMock).url.href).toContain('r.jina.ai')
+    expect(page).toEqual({ url: 'https://example.com/a', title: 'A', text: 'Body.' })
+  })
 })
 
 describe('missingSearchKey', () => {
diff --git a/src/tools/web.ts b/src/tools/web.ts
index 5e38037..2bbdfdd 100644
--- a/src/tools/web.ts
+++ b/src/tools/web.ts
@@ -146,6 +146,10 @@ export function normalizeWebAccess(
  * Where DuckDuckGo search and page reads should go, or `undefined` to stay
  * browser-direct. An empty string means same-origin `/api`, which is what
  * `pnpm dev` serves.
+ *
+ * An empty variable is not a proxy. A workflow that forwards an unset
+ * repository variable produces `''`, and reading that as same-origin would
+ * point every hosted visitor at a `/api` their host does not serve.
  */
 export function configuredProxyBase(config: WebAccessConfig): string | undefined {
   const runtime = config.proxyUrl?.trim()
@@ -161,7 +165,8 @@ export function configuredProxyBase(config: WebAccessConfig): string | undefined
   const env = import.meta.env.VITE_AGENT_API_BASE
   if (typeof env !== 'string') return undefined
   const trimmed = env.trim()
-  if (trimmed === '' || trimmed === 'same-origin') return ''
+  if (trimmed === '') return undefined
+  if (trimmed === 'same-origin') return ''
   return trimmed.replace(/\/$/, '')
 }
 
@@ -712,7 +717,16 @@ export async function searchWeb(
     return searchWikipedia(query, limit)
   }
   const proxy = configuredProxyBase(config)
-  if (proxy !== undefined) return searchViaProxy(proxy, query, limit)
+  if (proxy !== undefined) {
+    try {
+      return await searchViaProxy(proxy, query, limit)
+    } catch {
+      // The proxy is an optimisation, not a dependency. A hosted build points
+      // every visitor at one process, so an outage, a spent budget or an
+      // allowlist that has not caught up must cost a slower search rather than
+      // the answer. Browser-direct is what this build did before the proxy.
+    }
+  }
   return searchDuckDuckGo(query, limit, config)
 }
 
@@ -833,7 +847,16 @@ export async function readPage(rawUrl: string, config: WebAccessConfig): Promise
   }
 
   const proxy = configuredProxyBase(config)
-  if (proxy !== undefined) return fetchViaProxy(proxy, url.toString())
+  if (proxy !== undefined) {
+    try {
+      return await fetchViaProxy(proxy, url.toString())
+    } catch {
+      // Same bargain as search: the reader behind this still works, and a page
+      // the proxy could not fetch is worth one more attempt rather than an
+      // apology. `assertPublicHttpUrl` already refused the private targets, so
+      // nothing the proxy blocks on principle reaches this line.
+    }
+  }
 
   const data = await readWithReader(url.toString(), 'The page reader', config)
   if (!data?.content) throw new Error(`No readable content found at ${url.toString()}`)
diff --git a/src/vite-env.d.ts b/src/vite-env.d.ts
index 5329a56..9a13ea4 100644
--- a/src/vite-env.d.ts
+++ b/src/vite-env.d.ts
@@ -6,8 +6,9 @@ interface ImportMetaEnv {
   /** Path layout under that host, with {model} and {revision} placeholders. */
   readonly VITE_MODEL_PATH_TEMPLATE?: string
   /**
-   * Origin of the optional tool proxy. `same-origin` (or an empty value) means
-   * this page's `/api`. Unset on the Pages build, so tools stay browser-direct.
+   * Origin of the optional tool proxy. `same-origin` means this page's `/api`.
+   * Empty or unset stays browser-direct, which is what a Pages build with no
+   * proxy of its own gets.
    */
   readonly VITE_AGENT_API_BASE?: string
 }
diff --git a/tools/agent-api-listen.ts b/tools/agent-api-listen.ts
index c9b3f88..1bb6425 100644
--- a/tools/agent-api-listen.ts
+++ b/tools/agent-api-listen.ts
@@ -10,7 +10,7 @@
  */
 
 import { createServer, type IncomingMessage, type ServerResponse } from 'node:http'
-import { handleAgentApiRequest } from './agent-api.ts'
+import { createRateLimiter, handleAgentApiRequest } from './agent-api.ts'
 
 const PORT = Number(process.env.PORT) || 8787
 const HOST = process.env.HOST ?? '0.0.0.0'
@@ -19,6 +19,21 @@ const ALLOWLIST = (process.env.PROXY_ORIGINS ?? '')
   .map((origin) => origin.trim())
   .filter(Boolean)
 
+const RATE_LIMIT = Number(process.env.PROXY_RATE_LIMIT ?? 30)
+const RATE_WINDOW_MS = 60_000
+const limiter = createRateLimiter(RATE_LIMIT, RATE_WINDOW_MS)
+
+/**
+ * Railway and every other edge terminate the connection themselves, so the
+ * socket address is the proxy in front of us and would put every visitor in one
+ * bucket. The left-most forwarded address is the caller.
+ */
+function callerKey(req: IncomingMessage): string {
+  const forwarded = req.headers['x-forwarded-for']
+  const first = (Array.isArray(forwarded) ? forwarded[0] : forwarded)?.split(',')[0]?.trim()
+  return first || req.socket.remoteAddress || 'unknown'
+}
+
 function allowedOrigin(req: IncomingMessage): string | undefined {
   const origin = req.headers.origin
   if (ALLOWLIST.length === 0) return origin ?? '*'
@@ -50,6 +65,16 @@ const server = createServer((req, res) => {
       res.end()
       return
     }
+    // Health is exempt: the platform polls it far more often than a person
+    // searches, and a restart loop triggered by our own limit would be absurd.
+    const url = new URL(req.url ?? '/', 'http://localhost')
+    if (url.pathname !== '/api/health' && !limiter.take(callerKey(req))) {
+      res.statusCode = 429
+      res.setHeader('content-type', 'application/json; charset=utf-8')
+      res.setHeader('retry-after', String(Math.ceil(RATE_WINDOW_MS / 1000)))
+      res.end(JSON.stringify({ error: 'Too many requests' }))
+      return
+    }
     const handled = await handleAgentApiRequest(req, res)
     if (!handled) {
       res.statusCode = 404
@@ -70,4 +95,5 @@ server.listen(PORT, HOST, () => {
   console.log('POST /api/search  { query, limit?, region? }')
   console.log('POST /api/fetch   { url }')
   if (ALLOWLIST.length > 0) console.log(`Origins: ${ALLOWLIST.join(', ')}`)
+  console.log(RATE_LIMIT > 0 ? `Rate limit: ${RATE_LIMIT} per minute per caller` : 'Rate limit: off')
 })
diff --git a/tools/agent-api.test.ts b/tools/agent-api.test.ts
index 1269c40..520d854 100644
--- a/tools/agent-api.test.ts
+++ b/tools/agent-api.test.ts
@@ -1,6 +1,7 @@
 import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
 import {
   assertPublicUrl,
+  createRateLimiter,
   decodeEntities,
   extractPage,
   isBlockedHostname,
@@ -227,3 +228,42 @@ describe('routeAgentApi', () => {
     expect(result.payload).toEqual({ error: 'Refusing to fetch a private or loopback address' })
   })
 })
+
+describe('createRateLimiter', () => {
+  it('allows the quota and refuses the next call', () => {
+    const limiter = createRateLimiter(3, 60_000)
+    expect([1, 2, 3].map(() => limiter.take('1.2.3.4', 1000))).toEqual([true, true, true])
+    expect(limiter.take('1.2.3.4', 1000)).toBe(false)
+  })
+
+  it('counts each caller on its own', () => {
+    const limiter = createRateLimiter(1, 60_000)
+    expect(limiter.take('1.2.3.4', 1000)).toBe(true)
+    expect(limiter.take('1.2.3.4', 1000)).toBe(false)
+    expect(limiter.take('5.6.7.8', 1000)).toBe(true)
+  })
+
+  it('lets the window slide rather than resetting on a tick', () => {
+    const limiter = createRateLimiter(2, 60_000)
+    limiter.take('1.2.3.4', 1000)
+    limiter.take('1.2.3.4', 30_000)
+    expect(limiter.take('1.2.3.4', 50_000)).toBe(false)
+    // The first call has aged out by now; the second has not.
+    expect(limiter.take('1.2.3.4', 62_000)).toBe(true)
+    expect(limiter.take('1.2.3.4', 62_000)).toBe(false)
+  })
+
+  // The map is the one thing here that grows with the number of strangers who
+  // find the URL, so callers that stopped must not stay in it.
+  it('forgets a caller that fell silent', () => {
+    const limiter = createRateLimiter(1, 60_000)
+    limiter.take('1.2.3.4', 1000)
+    limiter.take('5.6.7.8', 200_000)
+    expect(limiter.take('1.2.3.4', 200_000)).toBe(true)
+  })
+
+  it('is off when the limit is zero', () => {
+    const limiter = createRateLimiter(0, 60_000)
+    expect([1, 2, 3, 4].every(() => limiter.take('1.2.3.4', 1000))).toBe(true)
+  })
+})
diff --git a/tools/agent-api.ts b/tools/agent-api.ts
index 98cf13e..9855bec 100644
--- a/tools/agent-api.ts
+++ b/tools/agent-api.ts
@@ -346,6 +346,47 @@ function clampLimit(value: unknown): number {
   return Math.min(Math.max(Math.trunc(n) || 5, 1), MAX_SEARCH_LIMIT)
 }
 
+export interface RateLimiter {
+  /** True when this caller may proceed. Counts the call when it does. */
+  take(key: string, now?: number): boolean
+}
+
+/**
+ * A sliding window per caller, in memory.
+ *
+ * The origin allowlist says which page may call; it says nothing about how
+ * often, and an allowed page is exactly what a scraper would forge. The window
+ * is the difference between one visitor searching and a stranger spending a
+ * month of budget in an afternoon. One process holds one window, so a service
+ * on several replicas allows that multiple — still a ceiling, which is the
+ * point.
+ */
+export function createRateLimiter(limit: number, windowMs: number): RateLimiter {
+  const seen = new Map<string, number[]>()
+
+  return {
+    take(key, now = Date.now()) {
+      if (limit <= 0) return true
+      const cutoff = now - windowMs
+      const hits = (seen.get(key) ?? []).filter((at) => at > cutoff)
+
+      // Callers that stopped must not stay in memory: this map is the only
+      // thing here that grows with the number of strangers who found the URL.
+      for (const [other, times] of seen) {
+        if (other !== key && times.every((at) => at <= cutoff)) seen.delete(other)
+      }
+
+      if (hits.length >= limit) {
+        seen.set(key, hits)
+        return false
+      }
+      hits.push(now)
+      seen.set(key, hits)
+      return true
+    },
+  }
+}
+
 export async function routeAgentApi(
   method: string,
   pathname: string,