diff --git a/.github/workflows/eval.yml b/.github/workflows/eval.yml
index 6d54abbab..25784c9d1 100644
--- a/.github/workflows/eval.yml
+++ b/.github/workflows/eval.yml
@@ -29,7 +29,8 @@ jobs:
       - name: Setup Node.js
         uses: actions/setup-node@v4
         with:
-          node-version: "22"
+          # Roach (see packages/mcp-server-evals/roach.json) needs Node 24.14+.
+          node-version: "24"
 
       # pnpm/action-setup@v4
       - uses: pnpm/action-setup@a7487c7e89a18df4991f7f222e4898a00d66ddda
@@ -62,6 +63,9 @@ jobs:
         env:
           OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
           EMBEDDED_AGENT_PROVIDER: openrouter
+          # Lets Roach record model calls it has no recording for. Fork PRs
+          # get no secrets, so they only replay.
+          ROACH_TOKEN: ${{ secrets.ROACH_TOKEN }}
 
       - name: Create eval status check
         uses: actions/github-script@v7
diff --git a/docs/testing/overview.md b/docs/testing/overview.md
index 4a7891969..fe2c87c1d 100644
--- a/docs/testing/overview.md
+++ b/docs/testing/overview.md
@@ -302,6 +302,22 @@ pnpm eval
 pnpm eval tool-name
 ```
 
+### Eval Recordings
+
+Evals send their OpenRouter calls through [Roach](https://github.com/getsentry/roach),
+which replays a recorded response when it has one. `packages/mcp-server-evals/roach.json`
+names the host and the rule. Recording and replay need Node 24.14 or later.
+
+- `ROACH_TOKEN` lets a run record calls that have no recording. The Eval
+  workflow sets it. Without it, a run only replays, and a call with no
+  recording fails its eval. A replay sends nothing upstream, so any
+  `OPENROUTER_API_KEY` value works for it.
+- `ROACH_MODE=off` sends every call live and records nothing. Use it locally
+  when you change prompts, tools, or fixtures and have no token.
+- `ROACH_MODE=record` sends every call live and replaces the recordings.
+- The Roach setup file comes before `src/setup-env.ts` in `vitest.config.ts`,
+  so MSW does not see Roach's own calls.
+
 ## Test Data Management
 
 ### Using Fixtures
diff --git a/packages/mcp-server-evals/package.json b/packages/mcp-server-evals/package.json
index 778274356..3b4486166 100644
--- a/packages/mcp-server-evals/package.json
+++ b/packages/mcp-server-evals/package.json
@@ -27,5 +27,8 @@
     "vitest": "catalog:",
     "vitest-evals": "catalog:",
     "zod": "catalog:"
+  },
+  "devDependencies": {
+    "@sentry/roach": "catalog:"
   }
 }
diff --git a/packages/mcp-server-evals/roach.json b/packages/mcp-server-evals/roach.json
new file mode 100644
index 000000000..e82db1c94
--- /dev/null
+++ b/packages/mcp-server-evals/roach.json
@@ -0,0 +1,13 @@
+{
+  "tenant": "getsentry/toolkit",
+  "allow": ["https://openrouter.ai"],
+  "rules": [
+    {
+      "name": "model",
+      "match": {
+        "method": "POST",
+        "url": "https://openrouter.ai/api/v1/"
+      }
+    }
+  ]
+}
diff --git a/packages/mcp-server-evals/vitest.config.ts b/packages/mcp-server-evals/vitest.config.ts
index cdbf92da4..c4e5a1906 100644
--- a/packages/mcp-server-evals/vitest.config.ts
+++ b/packages/mcp-server-evals/vitest.config.ts
@@ -10,6 +10,9 @@ export default defineConfig({
       reporter: ["text", "json", "html"],
       include: ["**/*.ts"],
     },
-    setupFiles: ["./src/setup-env.ts"],
+    // Roach records model calls and replays them on later runs. See roach.json.
+    // Its setup file loads before setup-env.ts starts MSW.
+    globalSetup: ["@sentry/roach/vitest"],
+    setupFiles: ["@sentry/roach/vitest/setup", "./src/setup-env.ts"],
   },
 });
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index db7fa68aa..a28c21b1e 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -57,6 +57,9 @@ catalogs:
     '@sentry/react':
       specifier: 11.6.0
       version: 11.6.0
+    '@sentry/roach':
+      specifier: 0.3.0
+      version: 0.3.0
     '@sentry/search-syntax':
       specifier: ^1.0.0
       version: 1.0.0
@@ -955,6 +958,10 @@ importers:
       zod:
         specifier: ^4.6.6
         version: 4.6.7
+    devDependencies:
+      '@sentry/roach':
+        specifier: 'catalog:'
+        version: 0.3.0(vitest@5.0.5)
 
   packages/mcp-server-mocks:
     dependencies:
@@ -4676,6 +4683,15 @@ packages:
     resolution: {integrity: sha512-cJyPYnF6QJA8ADRfAW06YoU7EVIebtWezBOCU9Vag5+W1VLBkXEvH7C4LbeMUiDmrqw92sumq06NgqipIxA7BA==}
     engines: {node: '>=20.19.0 <22.0.0 || >=22.12.0 <23.0.0 || >=23.2.0'}
 
+  '@sentry/roach@0.3.0':
+    resolution: {integrity: sha512-vgYiE3t/Cd5Ruff0KdCNktRNFYoXJTj10RUlcgqqhuMYTBbGlY+e/pU48M02WuueoQWx6zRZMm76jQRJQwp7KA==}
+    engines: {node: '>=24'}
+    peerDependencies:
+      vitest: '>=3'
+    peerDependenciesMeta:
+      vitest:
+        optional: true
+
   '@sentry/search-syntax@1.0.0':
     resolution: {integrity: sha512-R+raAe4PixC+e7Zn6uPH7NSK7nzfHy2MA1kRWDn7dFH8zaU10edFBVPj8qO3TlSb58yts0tejPdy+s1acq80wQ==}
 
@@ -12661,6 +12677,10 @@ snapshots:
       '@sentry/browser-utils': 11.6.0
       '@sentry/core': 11.6.0
 
+  '@sentry/roach@0.3.0(vitest@5.0.5)':
+    optionalDependencies:
+      vitest: 5.0.5(@opentelemetry/api@1.9.1)(@types/node@24.13.5)(@vitest/coverage-v8@5.0.5)(jsdom@28.1.0)(msw@2.10.2(@types/node@24.13.5)(typescript@5.9.3))(vite@8.3.3(@types/node@24.13.5)(esbuild@0.28.2)(jiti@2.7.0)(tsx@4.23.13)(yaml@2.9.0))
+
   '@sentry/search-syntax@1.0.0': {}
 
   '@sentry/server-runtime-injection@11.6.0':
diff --git a/pnpm-workspace.yaml b/pnpm-workspace.yaml
index 0b84f03e9..8686bd37d 100644
--- a/pnpm-workspace.yaml
+++ b/pnpm-workspace.yaml
@@ -57,6 +57,7 @@ catalog:
   "@sentry/core": 11.6.0
   "@sentry/node": 11.6.0
   "@sentry/react": 11.6.0
+  "@sentry/roach": 0.3.0
   "@sentry/search-syntax": ^1.0.0
   "@sentry/vite-plugin": ^4.6.1
   "@tailwindcss/typography": ^0.5.16
@@ -124,6 +125,7 @@ minimumReleaseAgeExclude:
   - "@sentry/react@11.6.0"
   - "@sentry/replay-canvas@11.6.0"
   - "@sentry/replay@11.6.0"
+  - "@sentry/roach@0.3.0"
   - "@sentry/search-syntax@1.0.0"
   - "@sentry/server-runtime-injection@11.6.0"
   - "@sentry/server-utils@11.6.0"