@mehmoodqureshi/chrome-mcp 0.6.3 → 0.6.5

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md CHANGED
@@ -4,36 +4,73 @@
4
4
  [![npm](https://img.shields.io/npm/v/%40mehmoodqureshi%2Fchrome-mcp?label=npm)](https://www.npmjs.com/package/@mehmoodqureshi/chrome-mcp)
5
5
  [![license](https://img.shields.io/npm/l/%40mehmoodqureshi%2Fchrome-mcp?label=license)](LICENSE)
6
6
 
7
- Drive a **real Chrome browser** from Claude (or any MCP host). One pluggable
8
- `Executor` interface, two backends:
9
-
10
- - **Extension (primary):** an MV3 extension drives your real Chrome real
11
- logins, real cookies — via `chrome.scripting`/`chrome.tabs`. The CLI runs a
12
- localhost WebSocket server; the extension dials in.
13
- - **CDP fallback:** when no extension is paired, the CLI launches/attaches a
14
- Playwright-driven Chromium for scripted/headless use.
7
+ **Let Claude use the Chrome you are already logged into.** Not a fresh
8
+ automated browser that greets every site as a stranger — *your* Chrome, with
9
+ your sessions, your cookies, your 2FA already done. If you can see a page in
10
+ your browser, your agent can read it, without logging in again and without
11
+ pasting credentials anywhere.
12
+
13
+ Most browser MCP servers launch their own Chromium and hand your agent a
14
+ signed-out window. chrome-mcp does the opposite: an MV3 extension dials into a
15
+ localhost WebSocket server and drives the browser you already have open, through
16
+ `chrome.scripting`/`chrome.tabs`. Works with Claude Code, Claude Desktop, and any
17
+ other MCP host.
15
18
 
16
19
  Distributed as an `npx` CLI (the MCP server) plus a load-unpacked extension.
17
20
 
21
+ > **This build is extension-only.** It never launches or attaches a Chromium of
22
+ > its own, so **the extension is required, not optional** — without it, no tool
23
+ > can run. The CDP flags (`--cdp-fallback`, `--no-cdp-fallback`, `--cdp-endpoint`,
24
+ > `--prefer`) are still accepted for back-compat but are **ignored**.
25
+
18
26
  > **Full design:** [`docs/BLUEPRINT.md`](docs/BLUEPRINT.md) — architecture, wire
19
27
  > protocol, the complete tool surface, the extension manifest, the security
20
28
  > model, and the phased build plan.
21
29
 
22
30
  ## Quickstart
23
31
 
24
- **1. Register the MCP server** with your host (e.g. Claude Desktop / Code):
32
+ **1. Register the MCP server** with your host.
33
+
34
+ <details open>
35
+ <summary><b>Claude Code (terminal)</b> — one command, no config file to find</summary>
36
+
37
+ ```bash
38
+ claude mcp add chrome-mcp -s user -- \
39
+ npx -y @mehmoodqureshi/chrome-mcp \
40
+ --allow-domain example.com --enable-mutations --persist-token
41
+ ```
42
+
43
+ Everything **before** `--` belongs to Claude Code; everything **after** it is this
44
+ server's command and flags. Keep the `--` or `--allow-domain` gets read as a
45
+ Claude Code option.
46
+
47
+ `-s user` registers it for every project on your machine. Use `-s local` (the
48
+ default) for just the current project, or `-s project` to write a `.mcp.json`
49
+ your team can commit.
50
+
51
+ Check it came up with `claude mcp list`. After upgrading the server, reconnect it
52
+ with `/mcp` inside a session — no restart needed.
53
+
54
+ </details>
55
+
56
+ <details>
57
+ <summary><b>Claude Desktop</b> and other MCP hosts — JSON config</summary>
25
58
 
26
59
  ```jsonc
27
60
  {
28
61
  "mcpServers": {
29
62
  "chrome-mcp": {
30
63
  "command": "npx",
31
- "args": ["-y", "@mehmoodqureshi/chrome-mcp", "--allow-domain", "example.com", "--enable-mutations"]
64
+ "args": ["-y", "@mehmoodqureshi/chrome-mcp",
65
+ "--allow-domain", "example.com", "--enable-mutations",
66
+ "--persist-token"]
32
67
  }
33
68
  }
34
69
  }
35
70
  ```
36
71
 
72
+ </details>
73
+
37
74
  By default everything is **deny-all** (no domains, no eval, no mutations). Grant
38
75
  exactly what you need with `--allow-domain <glob>` (repeatable), `--enable-mutations`,
39
76
  `--enable-downloads`, `--enable-uploads`, `--unsafe-enable-eval`, or `--unsafe-all-domains`.
@@ -44,38 +81,82 @@ exactly what you need with `--allow-domain <glob>` (repeatable), `--enable-mutat
44
81
  > with `--uploads-dir <path>` to restrict uploads to files inside that directory
45
82
  > (`..` traversal is blocked) — strongly recommended for unattended use.
46
83
 
47
- **Drive only your real Chrome (recommended for the extension).** Add
48
- `--no-cdp-fallback` so the server never launches a separate Chromium, and
49
- `--persist-token` so the pairing token survives restarts — **pair once, never
50
- again**:
84
+ **Pair once, never again.** Both examples above include `--persist-token`, which
85
+ is what makes the pairing survive a restart drop it if you'd rather have the
86
+ stricter default described next.
87
+
88
+ Without `--persist-token` a fresh token is minted every boot (the secure
89
+ default), which means re-pairing the extension on each restart. With it, the
90
+ token is stored 0600 at `~/.chrome-mcp/token` and reused; the extension's
91
+ keepalive auto-reconnects with no manual step. `CHROME_MCP_TOKEN` pins the token
92
+ explicitly (and is never written to disk).
93
+
94
+ **2. Load the extension** — **required**; the server can drive nothing without it.
95
+
96
+ `extension-dist/` ships prebuilt inside the npm package, so there is nothing to
97
+ compile. Install globally to get a stable path to it:
98
+
99
+ ```bash
100
+ npm install -g @mehmoodqureshi/chrome-mcp
101
+ npm root -g # → <root>; the extension is at <root>/@mehmoodqureshi/chrome-mcp/extension-dist
102
+ ```
103
+
104
+ Then `chrome://extensions` → enable **Developer mode** → **Load unpacked** →
105
+ select that `extension-dist/` directory. (Working from a git clone instead? Run
106
+ `npm install && npm run build:ext` first — `extension-dist/` is gitignored.)
107
+
108
+ **3. Pair it:** run `npx chrome-mcp --print-pairing` to write the handshake and
109
+ print its path, open the extension's **Options** page, and paste the `port` +
110
+ `token` from `~/.chrome-mcp/handshake.json`.
111
+
112
+ ### Running more than one session
113
+
114
+ The extension dials exactly **one** bridge port, so only one chrome-mcp can drive
115
+ your Chrome at a time — but every MCP host session (each Claude tab/window)
116
+ spawns its own server. With a pinned `--port`, the newest session **takes the
117
+ port over**: it reads the owning pid from `handshake.json`, confirms that process
118
+ really is a chrome-mcp, and stops it. Newest tab wins; the older session's browser
119
+ tools go quiet until it reconnects. Nothing that isn't a verified chrome-mcp is
120
+ ever touched — a port held by some other program is reported, never killed.
121
+
122
+ Two servers can only run side by side if each has its own port **and** its own
123
+ paired extension — i.e. a separate Chrome profile running its own copy of the
124
+ extension, pointed at the other port (`--port 9223`). A single Chrome pairs to one
125
+ server at a time, so a second server with no extension of its own can drive
126
+ nothing.
127
+
128
+ One server can, however, serve **several browsers at once**: connections are
129
+ routed by profile key (`--profile <name>`, matching the profile set in the
130
+ extension's Options), so each paired Chrome gets its own routing slot.
131
+
132
+ Without `--port`, each server binds an ephemeral port (no conflict ever), but the
133
+ port changes every boot — so you'd re-pair the extension each time. Pin `--port`
134
+ plus `--persist-token` for a pair-once setup.
135
+
136
+ ### Windows
137
+
138
+ WSL2 is **not** required — native Windows works. One config change is, though:
139
+ on Windows `npx` is `npx.cmd`, a batch shim, and MCP hosts spawn the server
140
+ without a shell, which cannot execute a `.cmd`. So `"command": "npx"` fails to
141
+ start. Wrap it in `cmd /c`:
51
142
 
52
143
  ```jsonc
53
144
  {
54
145
  "mcpServers": {
55
146
  "chrome-mcp": {
56
- "command": "npx",
57
- "args": ["-y", "@mehmoodqureshi/chrome-mcp",
147
+ "command": "cmd",
148
+ "args": ["/c", "npx", "-y", "@mehmoodqureshi/chrome-mcp",
58
149
  "--allow-domain", "example.com", "--enable-mutations",
59
- "--no-cdp-fallback", "--persist-token"]
150
+ "--persist-token"]
60
151
  }
61
152
  }
62
153
  }
63
154
  ```
64
155
 
65
- Without `--persist-token` a fresh token is minted every boot (the secure
66
- default), which means re-pairing the extension on each restart. With it, the
67
- token is stored 0600 at `~/.chrome-mcp/token` and reused; the extension's
68
- keepalive auto-reconnects with no manual step. `CHROME_MCP_TOKEN` pins the token
69
- explicitly (and is never written to disk).
70
-
71
- **2. Load the extension** (to drive your *real* Chrome): build it, then
72
- `chrome://extensions` → enable Developer mode → **Load unpacked** → select
73
- `extension-dist/`.
156
+ Or from Claude Code: `claude mcp add chrome-mcp -- cmd /c npx -y @mehmoodqureshi/chrome-mcp --allow-domain example.com`
74
157
 
75
- **3. Pair it:** run `npx chrome-mcp --print-pairing` to write the handshake and
76
- print its path, open the extension's **Options** page, and paste the `port` +
77
- `token` from `~/.chrome-mcp/handshake.json`. (Without the extension, the CLI
78
- falls back to a Playwright-driven Chromium automatically.)
158
+ Everything else is the same load `extension-dist/` from `npm root -g` and pair
159
+ as above.
79
160
 
80
161
  The tools cover tabs, navigation, interaction (`click`/`type`/`press`/`hover`/
81
162
  `scroll`/`select_option`), reads (`get_text`/`get_html`/`screenshot`/`eval`/`wait_for`),
@@ -183,7 +264,11 @@ chrome-mcp --unsafe-all-domains # loud footgun
183
264
  ```
184
265
 
185
266
  The per-boot 256-bit token in `~/.chrome-mcp/handshake.json` (mode 0600) is the
186
- only trust boundary; it is never written to stdout/stderr.
267
+ only trust boundary; it is never written to stdout/stderr. On POSIX the mode is
268
+ re-verified after every write and the server **fails closed** if the file ends up
269
+ group/other-accessible. Windows has no such bits — `chmod` there only toggles the
270
+ read-only attribute — so the check is skipped and the token's confidentiality
271
+ rests on the per-user ACL of `%USERPROFILE%\.chrome-mcp`.
187
272
 
188
273
  ## Develop
189
274
 
@@ -27,6 +27,18 @@ export declare const BRIDGE_HOST: "127.0.0.1";
27
27
  /** WebSocket close codes we use deliberately. */
28
28
  export declare const CLOSE_UNAUTHORIZED: 4401;
29
29
  export declare const CLOSE_SUPERSEDED: 4000;
30
+ /**
31
+ * Capabilities an extension advertises in `hello`. Additive and optional, so an
32
+ * older extension (which sends none) keeps the conservative behaviour — never
33
+ * gate a capability behind a version-string comparison.
34
+ *
35
+ * `tab-url`: this extension (a) enforces the policy mirror FAIL-CLOSED — it
36
+ * refuses every command until a policy has arrived — and (b) reports the target
37
+ * tab's post-command URL as `ResultFrame.tabUrl`. Together those let the server
38
+ * gate from the reported URL instead of paying a `tabs_list` round-trip before
39
+ * every call. Without it, the server falls back to fetching the URL itself.
40
+ */
41
+ export declare const WIRE_CAP_TAB_URL: "tab-url";
30
42
  /**
31
43
  * Every method that may travel on the wire = the MCP primitives 1:1, plus
32
44
  * `download_file` (privileged, executor-owned) and `ping_probe` (a short-deadline
@@ -52,6 +64,9 @@ export interface HelloFrame extends BaseFrame {
52
64
  * NOT a security boundary (the token is) — it selects which connection slot the
53
65
  * server routes commands to, so several browsers can stay paired at once. */
54
66
  profile?: string;
67
+ /** Optional capability advertisements (see `WIRE_CAP_TAB_URL`). An extension
68
+ * that sends none gets the conservative path, so old builds stay correct. */
69
+ caps?: string[];
55
70
  }
56
71
  /**
57
72
  * The wire-serializable subset of the server's policy, delivered in `welcome` so
@@ -92,6 +107,13 @@ export interface ResultFrame extends BaseFrame {
92
107
  ok: true;
93
108
  /** For `screenshot`: { dataBase64, mimeType, width, height, truncated }. */
94
109
  data: unknown;
110
+ /**
111
+ * The target tab's URL as observed AFTER the command ran — the extension has
112
+ * it locally, so sending it costs nothing and saves the server a round-trip
113
+ * on the next policy gate. Omitted when it can't be resolved (e.g. the tab was
114
+ * closed). Only sent by extensions advertising `WIRE_CAP_TAB_URL`.
115
+ */
116
+ tabUrl?: string;
95
117
  }
96
118
  export interface ErrorFrame extends BaseFrame {
97
119
  type: 'error';
@@ -14,7 +14,7 @@
14
14
  * wire method beyond the primitives.
15
15
  */
16
16
  Object.defineProperty(exports, "__esModule", { value: true });
17
- exports.WIRE_METHODS = exports.CLOSE_SUPERSEDED = exports.CLOSE_UNAUTHORIZED = exports.BRIDGE_HOST = exports.DEFAULT_WS_PORT = exports.PROTOCOL_VERSION = void 0;
17
+ exports.WIRE_METHODS = exports.WIRE_CAP_TAB_URL = exports.CLOSE_SUPERSEDED = exports.CLOSE_UNAUTHORIZED = exports.BRIDGE_HOST = exports.DEFAULT_WS_PORT = exports.PROTOCOL_VERSION = void 0;
18
18
  /** Bumped on any breaking change to the frames below. */
19
19
  exports.PROTOCOL_VERSION = 1;
20
20
  /**
@@ -29,6 +29,18 @@ exports.BRIDGE_HOST = '127.0.0.1';
29
29
  /** WebSocket close codes we use deliberately. */
30
30
  exports.CLOSE_UNAUTHORIZED = 4401;
31
31
  exports.CLOSE_SUPERSEDED = 4000;
32
+ /**
33
+ * Capabilities an extension advertises in `hello`. Additive and optional, so an
34
+ * older extension (which sends none) keeps the conservative behaviour — never
35
+ * gate a capability behind a version-string comparison.
36
+ *
37
+ * `tab-url`: this extension (a) enforces the policy mirror FAIL-CLOSED — it
38
+ * refuses every command until a policy has arrived — and (b) reports the target
39
+ * tab's post-command URL as `ResultFrame.tabUrl`. Together those let the server
40
+ * gate from the reported URL instead of paying a `tabs_list` round-trip before
41
+ * every call. Without it, the server falls back to fetching the URL itself.
42
+ */
43
+ exports.WIRE_CAP_TAB_URL = 'tab-url';
32
44
  /** Runtime list of every WireMethod, for boot-time drift assertions on both ends. */
33
45
  exports.WIRE_METHODS = [
34
46
  'tabs_list',
@@ -7,6 +7,7 @@
7
7
  * that want a stable token so the extension never has to re-pair.
8
8
  * - Written atomically (tmp + rename) to `handshake.json` at mode 0600; the
9
9
  * mode is re-verified after write and we FAIL CLOSED if it can't be set.
10
+ * POSIX only — Windows has no group/other bits to check (see assertPrivate).
10
11
  * - Compared by hashing both sides to SHA-256 and `timingSafeEqual`-ing the
11
12
  * digests — no length precondition, no length leak.
12
13
  * - The token is NEVER written to stdout/stderr or any log (a test asserts it).
@@ -8,6 +8,7 @@
8
8
  * that want a stable token so the extension never has to re-pair.
9
9
  * - Written atomically (tmp + rename) to `handshake.json` at mode 0600; the
10
10
  * mode is re-verified after write and we FAIL CLOSED if it can't be set.
11
+ * POSIX only — Windows has no group/other bits to check (see assertPrivate).
11
12
  * - Compared by hashing both sides to SHA-256 and `timingSafeEqual`-ing the
12
13
  * digests — no length precondition, no length leak.
13
14
  * - The token is NEVER written to stdout/stderr or any log (a test asserts it).
@@ -32,6 +33,23 @@ const datadir_1 = require("./datadir");
32
33
  function generateToken() {
33
34
  return (0, node_crypto_1.randomBytes)(32).toString('base64url');
34
35
  }
36
+ /**
37
+ * Fail closed if `path` is group/other-accessible.
38
+ *
39
+ * No-op on Windows: NTFS ACLs don't map onto the POSIX mode bits, `chmodSync`
40
+ * there only toggles the read-only attribute, and `statSync` reports a synthetic
41
+ * 0o666 for any writable file. Enforcing the check would therefore throw on
42
+ * every well-formed file. The token's confidentiality on Windows rests on the
43
+ * per-user ACL of the profile directory holding it.
44
+ */
45
+ function assertPrivate(path, what) {
46
+ if (process.platform === 'win32')
47
+ return;
48
+ const mode = (0, node_fs_1.statSync)(path).mode & 0o777;
49
+ if ((mode & 0o077) !== 0) {
50
+ throw new Error(`${what} ${path} is group/other-accessible (mode ${mode.toString(8)}); refusing to expose the token`);
51
+ }
52
+ }
35
53
  /** Path of the optional persisted-token file (only used when persistence is on). */
36
54
  function tokenPath(dir) {
37
55
  return (0, node_path_1.join)(dir, 'token');
@@ -50,10 +68,7 @@ function readPersistedToken(dir) {
50
68
  catch {
51
69
  return null;
52
70
  }
53
- const mode = (0, node_fs_1.statSync)(path).mode & 0o777;
54
- if ((mode & 0o077) !== 0) {
55
- throw new Error(`persisted token ${path} is group/other-accessible (mode ${mode.toString(8)}); refusing to reuse it`);
56
- }
71
+ assertPrivate(path, 'persisted token');
57
72
  const token = raw.trim();
58
73
  return token.length > 0 ? token : null;
59
74
  }
@@ -65,10 +80,7 @@ function writePersistedToken(dir, token) {
65
80
  (0, node_fs_1.chmodSync)(tmp, 0o600);
66
81
  (0, node_fs_1.renameSync)(tmp, path);
67
82
  (0, node_fs_1.chmodSync)(path, 0o600);
68
- const mode = (0, node_fs_1.statSync)(path).mode & 0o777;
69
- if ((mode & 0o077) !== 0) {
70
- throw new Error(`persisted token ${path} is group/other-accessible (mode ${mode.toString(8)}); refusing to expose it`);
71
- }
83
+ assertPrivate(path, 'persisted token');
72
84
  return path;
73
85
  }
74
86
  /**
@@ -118,10 +130,7 @@ function writeHandshake(dir, fields) {
118
130
  (0, node_fs_1.chmodSync)(tmp, 0o600);
119
131
  (0, node_fs_1.renameSync)(tmp, path);
120
132
  (0, node_fs_1.chmodSync)(path, 0o600);
121
- const mode = (0, node_fs_1.statSync)(path).mode & 0o777;
122
- if ((mode & 0o077) !== 0) {
123
- throw new Error(`handshake file ${path} is group/other-accessible (mode ${mode.toString(8)}); refusing to expose the token`);
124
- }
133
+ assertPrivate(path, 'handshake file');
125
134
  return path;
126
135
  }
127
136
  function readHandshake(dir) {
@@ -17,6 +17,8 @@ export interface ConnectionDeps {
17
17
  extId: string;
18
18
  sessionId: string;
19
19
  heartbeatMs: number;
20
+ /** Capabilities from `hello` (see WIRE_CAP_TAB_URL). Old builds send none. */
21
+ caps?: string[];
20
22
  onEvent?: (event: WireEvent, data: Record<string, unknown>) => void;
21
23
  onClose?: (code: number) => void;
22
24
  onLog?: (message: string) => void;
@@ -30,6 +32,10 @@ export declare class ExtensionConnection {
30
32
  private closed;
31
33
  private heartbeat;
32
34
  private missedPongs;
35
+ /** Whether this extension reports `tabUrl` and gates fail-closed. */
36
+ private readonly reportsTabUrl;
37
+ /** Last URL the ACTIVE tab reported, with the wall-clock it arrived. */
38
+ private activeUrl;
33
39
  private readonly onEvent?;
34
40
  private readonly onClose?;
35
41
  private readonly onLog?;
@@ -43,6 +49,18 @@ export declare class ExtensionConnection {
43
49
  isOpen(): boolean;
44
50
  private handleMessage;
45
51
  private settle;
52
+ /**
53
+ * Cache the URL a result rode home with — but ONLY when it describes the active
54
+ * tab (no explicit tabId) and actually resolved. A blank `tabUrl` means the
55
+ * extension couldn't read it (closed tab, restricted page), which is a reason
56
+ * to forget what we knew, never to keep believing it.
57
+ */
58
+ private rememberActiveUrl;
59
+ /**
60
+ * The active tab's last reported URL if it is younger than `maxAgeMs`, else
61
+ * null — the caller then resolves it the slow way. Never returns a guess.
62
+ */
63
+ lastActiveUrl(maxAgeMs: number): string | null;
46
64
  private handleClose;
47
65
  private startHeartbeat;
48
66
  }
@@ -36,6 +36,8 @@ function mapWireErrorCode(code) {
36
36
  };
37
37
  return known[code] ?? 'TARGET_GONE';
38
38
  }
39
+ /** Commands that can change WHICH tab is active, invalidating a cached URL. */
40
+ const ACTIVE_TAB_CHANGERS = new Set(['tab_select', 'tab_new', 'tab_close']);
39
41
  class ExtensionConnection {
40
42
  extId;
41
43
  sessionId;
@@ -45,6 +47,10 @@ class ExtensionConnection {
45
47
  closed = false;
46
48
  heartbeat = null;
47
49
  missedPongs = 0;
50
+ /** Whether this extension reports `tabUrl` and gates fail-closed. */
51
+ reportsTabUrl;
52
+ /** Last URL the ACTIVE tab reported, with the wall-clock it arrived. */
53
+ activeUrl = null;
48
54
  onEvent;
49
55
  onClose;
50
56
  onLog;
@@ -52,6 +58,7 @@ class ExtensionConnection {
52
58
  this.ws = deps.ws;
53
59
  this.extId = deps.extId;
54
60
  this.sessionId = deps.sessionId;
61
+ this.reportsTabUrl = deps.caps?.includes(protocol_1.WIRE_CAP_TAB_URL) ?? false;
55
62
  this.onEvent = deps.onEvent;
56
63
  this.onClose = deps.onClose;
57
64
  this.onLog = deps.onLog;
@@ -71,6 +78,11 @@ class ExtensionConnection {
71
78
  }
72
79
  const id = String(++this.seq);
73
80
  const timeoutMs = opts?.timeoutMs ?? defaultTimeoutFor(method);
81
+ // Anything that reshuffles tabs makes the cached URL a claim about a tab that
82
+ // may no longer be the active one. Drop it before the command, not after, so
83
+ // a failure mid-flight can't leave a stale entry behind.
84
+ if (ACTIVE_TAB_CHANGERS.has(method))
85
+ this.activeUrl = null;
74
86
  const frame = {
75
87
  type: 'command',
76
88
  v: protocol_1.PROTOCOL_VERSION,
@@ -86,7 +98,7 @@ class ExtensionConnection {
86
98
  reject(new types_1.ExecutorError('TIMEOUT', `"${method}" timed out after ${timeoutMs}ms`));
87
99
  }, timeoutMs);
88
100
  timer.unref?.();
89
- this.pending.set(id, { resolve, reject, timer, method });
101
+ this.pending.set(id, { resolve, reject, timer, method, activeTab: opts?.tabId === undefined });
90
102
  try {
91
103
  this.ws.send(JSON.stringify(frame));
92
104
  }
@@ -148,12 +160,38 @@ class ExtensionConnection {
148
160
  clearTimeout(p.timer);
149
161
  this.pending.delete(id);
150
162
  if (frame.type === 'result') {
163
+ this.rememberActiveUrl(p, frame);
151
164
  p.resolve(frame.data);
152
165
  }
153
166
  else {
167
+ // A failed command tells us nothing reliable about where the tab ended up.
168
+ if (p.activeTab)
169
+ this.activeUrl = null;
154
170
  p.reject(new types_1.ExecutorError(mapWireErrorCode(frame.error.code), frame.error.message));
155
171
  }
156
172
  }
173
+ /**
174
+ * Cache the URL a result rode home with — but ONLY when it describes the active
175
+ * tab (no explicit tabId) and actually resolved. A blank `tabUrl` means the
176
+ * extension couldn't read it (closed tab, restricted page), which is a reason
177
+ * to forget what we knew, never to keep believing it.
178
+ */
179
+ rememberActiveUrl(p, frame) {
180
+ if (!this.reportsTabUrl || !p.activeTab)
181
+ return;
182
+ if (ACTIVE_TAB_CHANGERS.has(p.method))
183
+ return; // already invalidated; re-caching would race
184
+ this.activeUrl = frame.tabUrl ? { url: frame.tabUrl, at: Date.now() } : null;
185
+ }
186
+ /**
187
+ * The active tab's last reported URL if it is younger than `maxAgeMs`, else
188
+ * null — the caller then resolves it the slow way. Never returns a guess.
189
+ */
190
+ lastActiveUrl(maxAgeMs) {
191
+ if (!this.activeUrl)
192
+ return null;
193
+ return Date.now() - this.activeUrl.at <= maxAgeMs ? this.activeUrl.url : null;
194
+ }
157
195
  handleClose(code) {
158
196
  if (this.closed)
159
197
  return;
@@ -0,0 +1,32 @@
1
+ /**
2
+ * src/bridge/evict.ts — take a pinned port back from a stale chrome-mcp.
3
+ *
4
+ * The extension dials exactly ONE bridge port, so only one server can own a
5
+ * given Chrome at a time. But every MCP host session spawns its own chrome-mcp
6
+ * child, so a second session (another Claude tab/window) racing for a pinned
7
+ * `--port` would otherwise just fail with EADDRINUSE and force the user to kill
8
+ * the old process by hand. Newest-session-wins: we do that kill for them.
9
+ *
10
+ * The safety bar is high, because a pid can be recycled onto an unrelated
11
+ * process. We evict ONLY when every check agrees:
12
+ * - the handshake names that port (so it describes *this* listener), and
13
+ * - the pid is not us, and
14
+ * - the pid is alive, and
15
+ * - its command line mentions chrome-mcp.
16
+ * If any check fails we leave the process alone and let the caller surface the
17
+ * plain-English port-busy error instead. Refusing to evict is always safe; a
18
+ * wrong kill is not.
19
+ */
20
+ /** Does this command line belong to a chrome-mcp server? */
21
+ export declare function looksLikeChromeMcp(cmdline: string): boolean;
22
+ export interface EvictDeps {
23
+ /** Injectable for tests; defaults to the real probes. */
24
+ isAlive?: (pid: number) => boolean;
25
+ commandLine?: (pid: number) => string | null;
26
+ kill?: (pid: number, signal: NodeJS.Signals) => void;
27
+ }
28
+ /**
29
+ * Try to free `port` by terminating the chrome-mcp that owns it.
30
+ * Returns true only if an owner was found, killed, and confirmed gone.
31
+ */
32
+ export declare function evictPortOwner(dataDir: string, port: number, log: (message: string) => void, deps?: EvictDeps): Promise<boolean>;
@@ -0,0 +1,135 @@
1
+ "use strict";
2
+ /**
3
+ * src/bridge/evict.ts — take a pinned port back from a stale chrome-mcp.
4
+ *
5
+ * The extension dials exactly ONE bridge port, so only one server can own a
6
+ * given Chrome at a time. But every MCP host session spawns its own chrome-mcp
7
+ * child, so a second session (another Claude tab/window) racing for a pinned
8
+ * `--port` would otherwise just fail with EADDRINUSE and force the user to kill
9
+ * the old process by hand. Newest-session-wins: we do that kill for them.
10
+ *
11
+ * The safety bar is high, because a pid can be recycled onto an unrelated
12
+ * process. We evict ONLY when every check agrees:
13
+ * - the handshake names that port (so it describes *this* listener), and
14
+ * - the pid is not us, and
15
+ * - the pid is alive, and
16
+ * - its command line mentions chrome-mcp.
17
+ * If any check fails we leave the process alone and let the caller surface the
18
+ * plain-English port-busy error instead. Refusing to evict is always safe; a
19
+ * wrong kill is not.
20
+ */
21
+ Object.defineProperty(exports, "__esModule", { value: true });
22
+ exports.looksLikeChromeMcp = looksLikeChromeMcp;
23
+ exports.evictPortOwner = evictPortOwner;
24
+ const node_child_process_1 = require("node:child_process");
25
+ const auth_1 = require("./auth");
26
+ /** How long to let a SIGTERM'd owner exit before escalating to SIGKILL. */
27
+ const EXIT_WAIT_MS = 3_000;
28
+ const POLL_MS = 100;
29
+ const delay = (ms) => new Promise((r) => setTimeout(r, ms));
30
+ /** Whether `pid` exists. Signal 0 checks liveness without delivering anything. */
31
+ function isAlive(pid) {
32
+ try {
33
+ process.kill(pid, 0);
34
+ return true;
35
+ }
36
+ catch (err) {
37
+ // EPERM means it exists but belongs to another user — alive, just not ours.
38
+ return err?.code === 'EPERM';
39
+ }
40
+ }
41
+ /**
42
+ * Full command line of `pid`, or null if it can't be read.
43
+ *
44
+ * Windows needs PowerShell/CIM: `tasklist` reports only the image name
45
+ * ("node.exe"), which can't distinguish chrome-mcp from any other Node process
46
+ * — far too weak a signal to kill on.
47
+ */
48
+ function commandLine(pid) {
49
+ const [cmd, args] = process.platform === 'win32'
50
+ ? [
51
+ 'powershell.exe',
52
+ [
53
+ '-NoProfile',
54
+ '-Command',
55
+ `(Get-CimInstance Win32_Process -Filter "ProcessId=${pid}").CommandLine`,
56
+ ],
57
+ ]
58
+ : ['ps', ['-p', String(pid), '-o', 'command=']];
59
+ try {
60
+ const out = (0, node_child_process_1.execFileSync)(cmd, args, {
61
+ encoding: 'utf8',
62
+ stdio: ['ignore', 'pipe', 'ignore'],
63
+ timeout: 5_000,
64
+ });
65
+ return out.trim() || null;
66
+ }
67
+ catch {
68
+ return null;
69
+ }
70
+ }
71
+ /** Does this command line belong to a chrome-mcp server? */
72
+ function looksLikeChromeMcp(cmdline) {
73
+ return /chrome-mcp/i.test(cmdline);
74
+ }
75
+ /**
76
+ * Try to free `port` by terminating the chrome-mcp that owns it.
77
+ * Returns true only if an owner was found, killed, and confirmed gone.
78
+ */
79
+ async function evictPortOwner(dataDir, port, log, deps = {}) {
80
+ const alive = deps.isAlive ?? isAlive;
81
+ const cmdOf = deps.commandLine ?? commandLine;
82
+ const kill = deps.kill ?? ((pid, sig) => process.kill(pid, sig));
83
+ let pid;
84
+ let hsPort;
85
+ try {
86
+ const hs = (0, auth_1.readHandshake)(dataDir);
87
+ pid = Number(hs?.pid);
88
+ hsPort = Number(hs?.port);
89
+ }
90
+ catch {
91
+ return false; // no handshake (clean shutdown) or unreadable — nothing to evict
92
+ }
93
+ if (!Number.isInteger(pid) || pid <= 0)
94
+ return false;
95
+ // The handshake describes a different listener; it says nothing about ours.
96
+ if (hsPort !== port)
97
+ return false;
98
+ if (pid === process.pid)
99
+ return false;
100
+ if (!alive(pid))
101
+ return false;
102
+ const cmdline = cmdOf(pid);
103
+ if (!cmdline || !looksLikeChromeMcp(cmdline)) {
104
+ log(`port ${port} is held by pid ${pid}, which does not look like chrome-mcp — ` +
105
+ `leaving it alone. Stop it yourself, or use a different --port.`);
106
+ return false;
107
+ }
108
+ log(`port ${port} is held by chrome-mcp pid ${pid} (another session) — taking the port over`);
109
+ try {
110
+ kill(pid, 'SIGTERM');
111
+ }
112
+ catch {
113
+ return false; // vanished or not ours to signal
114
+ }
115
+ const deadline = Date.now() + EXIT_WAIT_MS;
116
+ while (Date.now() < deadline) {
117
+ if (!alive(pid)) {
118
+ log(`previous owner (pid ${pid}) exited; reclaiming port ${port}`);
119
+ return true;
120
+ }
121
+ await delay(POLL_MS);
122
+ }
123
+ // Wouldn't go quietly.
124
+ try {
125
+ kill(pid, 'SIGKILL');
126
+ }
127
+ catch {
128
+ /* already gone */
129
+ }
130
+ const gone = !alive(pid);
131
+ if (gone)
132
+ log(`previous owner (pid ${pid}) force-killed; reclaiming port ${port}`);
133
+ return gone;
134
+ }
135
+ //# sourceMappingURL=evict.js.map
@@ -26,6 +26,9 @@ export interface BridgeOptions {
26
26
  port?: number;
27
27
  host?: string;
28
28
  heartbeatMs?: number;
29
+ /** Data dir holding the handshake. Enables reclaiming a pinned port from a
30
+ * stale chrome-mcp (see ./evict). Omit to disable eviction entirely. */
31
+ dataDir?: string;
29
32
  /** Diagnostics — MUST never receive the token (a test asserts this). */
30
33
  onLog?: (message: string) => void;
31
34
  onDisplacement?: (info: DisplacementInfo) => void;
@@ -62,6 +65,12 @@ export declare class BridgeServer {
62
65
  timeoutMs?: number;
63
66
  profile?: string;
64
67
  }): Promise<unknown>;
68
+ /**
69
+ * The active tab's URL for `profile` as last reported by the extension, if it
70
+ * is younger than `maxAgeMs`. Null means "ask properly" — an extension too old
71
+ * to report URLs, a tab shuffle since, or simply nothing recent enough.
72
+ */
73
+ lastActiveUrl(profile: string | undefined, maxAgeMs: number): string | null;
65
74
  private noPairMessage;
66
75
  status(): {
67
76
  extensionConnected: boolean;
@@ -20,6 +20,7 @@ const policy_1 = require("../../shared/policy");
20
20
  const types_1 = require("../executor/types");
21
21
  const connection_1 = require("./connection");
22
22
  const auth_1 = require("./auth");
23
+ const evict_1 = require("./evict");
23
24
  const config_1 = require("../config");
24
25
  /** The routing label for a hello with no/blank profile — the back-compat default. */
25
26
  const DEFAULT_PROFILE = 'default';
@@ -46,11 +47,12 @@ const delay = (ms) => new Promise((r) => setTimeout(r, ms));
46
47
  /** A friendly, actionable message for the rare case the port stays busy past PORT_WAIT_MS. */
47
48
  function portBusyMessage(host, port) {
48
49
  return (`Couldn't start: another program is already using ${host}:${port}.\n` +
49
- `This is almost always a previous chrome-mcp that didn't shut down. To fix it:\n` +
50
- ` 1. Fully quit/restart your MCP host (e.g. Claude Code), or\n` +
51
- ` 2. Stop the leftover process, then reconnect:\n` +
50
+ `A stale chrome-mcp is reclaimed automatically, so this is some OTHER program.\n` +
51
+ `To fix it:\n` +
52
+ ` 1. Stop whatever owns the port, then reconnect:\n` +
52
53
  ` macOS/Linux: lsof -nP -iTCP:${port} -sTCP:LISTEN then kill <PID>\n` +
53
- ` 3. Or run chrome-mcp with a different port: --port <number>`);
54
+ ` Windows: netstat -ano | findstr :${port} then taskkill /PID <PID> /F\n` +
55
+ ` 2. Or run chrome-mcp with a different port: --port <number>`);
54
56
  }
55
57
  class BridgeServer {
56
58
  opts;
@@ -76,6 +78,7 @@ class BridgeServer {
76
78
  // wait-and-retry for a few seconds so the old process can release it; only if
77
79
  // it never frees up do we surface a plain-English, actionable error.
78
80
  const deadline = Date.now() + PORT_WAIT_MS;
81
+ let evicted = false;
79
82
  for (let attempt = 1;; attempt++) {
80
83
  try {
81
84
  const wss = await this.listenOnce(host, port);
@@ -87,11 +90,20 @@ class BridgeServer {
87
90
  }
88
91
  catch (err) {
89
92
  const inUse = err?.code === 'EADDRINUSE';
90
- if (!inUse || port === 0 || Date.now() >= deadline) {
91
- if (inUse)
92
- throw new Error(portBusyMessage(host, port));
93
- throw err;
93
+ if (!inUse || port === 0)
94
+ throw inUse ? new Error(portBusyMessage(host, port)) : err;
95
+ // A live chrome-mcp from another session will never release the port on
96
+ // its own, so waiting it out is futile — take it over (once). Only a
97
+ // verified chrome-mcp is ever killed; anything else falls through to the
98
+ // wait-and-retry below, which covers our own just-replaced instance.
99
+ if (!evicted && this.opts.dataDir) {
100
+ evicted = true;
101
+ if (await (0, evict_1.evictPortOwner)(this.opts.dataDir, port, (m) => this.log(m))) {
102
+ continue;
103
+ }
94
104
  }
105
+ if (Date.now() >= deadline)
106
+ throw new Error(portBusyMessage(host, port));
95
107
  if (attempt === 1)
96
108
  this.log(`port ${host}:${port} busy — waiting for the previous instance to release it…`);
97
109
  await delay(PORT_RETRY_MS);
@@ -166,6 +178,15 @@ class BridgeServer {
166
178
  }
167
179
  return conn.sendCommand(method, params, opts);
168
180
  }
181
+ /**
182
+ * The active tab's URL for `profile` as last reported by the extension, if it
183
+ * is younger than `maxAgeMs`. Null means "ask properly" — an extension too old
184
+ * to report URLs, a tab shuffle since, or simply nothing recent enough.
185
+ */
186
+ lastActiveUrl(profile, maxAgeMs) {
187
+ const conn = this.conns.get(routeKey(profile));
188
+ return conn?.isOpen() ? conn.lastActiveUrl(maxAgeMs) : null;
189
+ }
169
190
  noPairMessage(profile) {
170
191
  return (`No browser is paired for profile "${profile}". In that Chrome's chrome-mcp ` +
171
192
  `extension Options, set Port ${this.boundPort}, paste the token, set Profile to ` +
@@ -223,7 +244,7 @@ class BridgeServer {
223
244
  authed = true;
224
245
  clearTimeout(helloTimer);
225
246
  ws.off('message', onMessage);
226
- this.promote(ws, frame.ext ?? { id: 'unknown', version: '0', chrome: '0' }, routeKey(frame.profile));
247
+ this.promote(ws, frame.ext ?? { id: 'unknown', version: '0', chrome: '0' }, routeKey(frame.profile), Array.isArray(frame.caps) ? frame.caps : undefined);
227
248
  };
228
249
  ws.on('message', onMessage);
229
250
  ws.on('error', () => {
@@ -240,7 +261,7 @@ class BridgeServer {
240
261
  /* ignore */
241
262
  }
242
263
  }
243
- promote(ws, ext, profile) {
264
+ promote(ws, ext, profile, caps) {
244
265
  const sessionId = (0, node_crypto_1.randomUUID)();
245
266
  // Supersede only the SAME profile's connection (a re-pair). Other profiles
246
267
  // keep their live connections, so several browsers stay paired at once.
@@ -263,6 +284,7 @@ class BridgeServer {
263
284
  extId: ext.id,
264
285
  sessionId,
265
286
  heartbeatMs: this.heartbeatMs,
287
+ caps,
266
288
  onEvent: this.opts.onEvent,
267
289
  onLog: (m) => this.log(m),
268
290
  onClose: () => {
package/dist/src/cli.js CHANGED
@@ -159,6 +159,7 @@ async function main() {
159
159
  // Wire-serializable policy subset (no local uploadsDir) so the extension mirrors the gate.
160
160
  policy: { allowDomains, allowEval, allowDownloads, allowUploads, allowAllTabs, enableMutations },
161
161
  port: cfg.wsPort,
162
+ dataDir,
162
163
  onLog: (m) => (0, server_2.logErr)(m),
163
164
  onDisplacement: (d) => (0, server_2.logErr)(`SECURITY: extension connection displaced (different id: ${d.differentId})`),
164
165
  });
@@ -65,8 +65,14 @@ function clearProfileLocks(profileDir) {
65
65
  }
66
66
  }
67
67
  function isChromiumProcess(pid) {
68
+ // Windows has no `ps`; tasklist is the equivalent and is a real .exe, so it
69
+ // needs no shell. Its CSV row leads with the image name (e.g. "chrome.exe"),
70
+ // and the no-match case prints an INFO line that fails the same test.
71
+ const [cmd, args] = process.platform === 'win32'
72
+ ? ['tasklist', ['/FI', `PID eq ${pid}`, '/FO', 'CSV', '/NH']]
73
+ : ['ps', ['-p', String(pid), '-o', 'command=']];
68
74
  try {
69
- const out = (0, node_child_process_1.execFileSync)('ps', ['-p', String(pid), '-o', 'command='], {
75
+ const out = (0, node_child_process_1.execFileSync)(cmd, args, {
70
76
  encoding: 'utf8',
71
77
  stdio: ['ignore', 'pipe', 'ignore'],
72
78
  });
@@ -21,6 +21,9 @@ export declare class ExtensionExecutor implements Executor {
21
21
  ensureReady(): Promise<void>;
22
22
  ping(deadlineMs?: number): Promise<boolean>;
23
23
  dispose(): Promise<void>;
24
+ /** The active tab's URL as reported by the last command on this profile, if it
25
+ * is fresh enough to gate against. See `ACTIVE_URL_TTL_MS`. */
26
+ cachedActiveUrl(): string | null;
24
27
  tabsList(): Promise<TabInfo[]>;
25
28
  tabSelect(tabId: TabId): Promise<TabInfo>;
26
29
  tabNew(url?: string, opts?: {
@@ -12,6 +12,16 @@ Object.defineProperty(exports, "__esModule", { value: true });
12
12
  exports.ExtensionExecutor = void 0;
13
13
  const types_1 = require("./types");
14
14
  const workspace_1 = require("../bridge/workspace");
15
+ /**
16
+ * How long a reported active-tab URL stays usable for the policy gate.
17
+ *
18
+ * Deliberately short. It exists to cover back-to-back calls (a `batch`, or an
19
+ * agent's read → click → read), where the tab demonstrably has not changed
20
+ * between them. Past that, pay the round-trip. Note the extension re-gates every
21
+ * command against the tab's live URL regardless, so this window trades a little
22
+ * pre-check precision for half the traffic — never enforcement itself.
23
+ */
24
+ const ACTIVE_URL_TTL_MS = 1_000;
15
25
  /** Flatten a Target into the params a wire command carries. */
16
26
  function targetParams(t) {
17
27
  if (!t)
@@ -64,6 +74,11 @@ class ExtensionExecutor {
64
74
  async dispose() {
65
75
  // Never close the user's Chrome.
66
76
  }
77
+ /** The active tab's URL as reported by the last command on this profile, if it
78
+ * is fresh enough to gate against. See `ACTIVE_URL_TTL_MS`. */
79
+ cachedActiveUrl() {
80
+ return this.bridge.lastActiveUrl(this.activeProfile(), ACTIVE_URL_TTL_MS);
81
+ }
67
82
  // -- tabs ---------------------------------------------------------------
68
83
  async tabsList() {
69
84
  return (await this.send('tabs_list', {}));
@@ -13,14 +13,32 @@ export interface StubOptions {
13
13
  activeUrl?: string;
14
14
  /** When true, `eval` resolves `{ok:false}` to mimic a page-side throw. */
15
15
  evalThrows?: boolean;
16
+ /** When true, `tabsList` rejects — mimics a transient bridge failure. */
17
+ tabsListThrows?: boolean;
18
+ /** When true, `tabsList` resolves empty — mimics a browser reporting no tabs. */
19
+ noTabs?: boolean;
20
+ /** When true, the (single) tab reports an empty URL — mimics a chrome:// page
21
+ * or a site the extension has no host access to. */
22
+ blankTabUrl?: boolean;
23
+ /** A URL the backend claims to already know, as the extension reports on every
24
+ * result frame. Set it to assert the gate uses it INSTEAD of calling tabsList. */
25
+ cachedUrl?: string;
16
26
  }
17
27
  export declare class StubExecutor implements Executor {
18
28
  readonly backend: BackendKind;
19
29
  private url;
20
30
  private readonly evalThrows;
31
+ private readonly tabsListThrows;
32
+ private readonly noTabs;
33
+ private readonly blankTabUrl;
34
+ private readonly cached;
35
+ /** How many times the gate actually asked for the tab list — the round-trip
36
+ * counter the caching path exists to keep at zero. */
37
+ tabsListCalls: number;
21
38
  private ready;
22
39
  constructor(opts?: StubOptions);
23
40
  private tab;
41
+ cachedActiveUrl(): string | null;
24
42
  status(): ExecutorStatus;
25
43
  ensureReady(): Promise<void>;
26
44
  ping(): Promise<boolean>;
@@ -10,6 +10,7 @@
10
10
  */
11
11
  Object.defineProperty(exports, "__esModule", { value: true });
12
12
  exports.StubExecutor = void 0;
13
+ const types_1 = require("./types");
13
14
  /** 1×1 transparent PNG, base64 — a valid image block for screenshot tests. */
14
15
  const TINY_PNG = 'iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNkYPhfDwAChwGA60e6kgAAAABJRU5ErkJggg==';
15
16
  const ok = { ok: true };
@@ -17,13 +18,33 @@ class StubExecutor {
17
18
  backend = 'extension';
18
19
  url;
19
20
  evalThrows;
21
+ tabsListThrows;
22
+ noTabs;
23
+ blankTabUrl;
24
+ cached;
25
+ /** How many times the gate actually asked for the tab list — the round-trip
26
+ * counter the caching path exists to keep at zero. */
27
+ tabsListCalls = 0;
20
28
  ready = false;
21
29
  constructor(opts = {}) {
22
30
  this.url = opts.activeUrl ?? 'about:blank';
23
31
  this.evalThrows = opts.evalThrows ?? false;
32
+ this.tabsListThrows = opts.tabsListThrows ?? false;
33
+ this.noTabs = opts.noTabs ?? false;
34
+ this.blankTabUrl = opts.blankTabUrl ?? false;
35
+ this.cached = opts.cachedUrl ?? null;
24
36
  }
25
37
  tab() {
26
- return { tabId: 'extension:stub:1', url: this.url, title: 'Stub Page', active: true, index: 0 };
38
+ return {
39
+ tabId: 'extension:stub:1',
40
+ url: this.blankTabUrl ? '' : this.url,
41
+ title: 'Stub Page',
42
+ active: true,
43
+ index: 0,
44
+ };
45
+ }
46
+ cachedActiveUrl() {
47
+ return this.cached;
27
48
  }
28
49
  status() {
29
50
  return {
@@ -44,7 +65,10 @@ class StubExecutor {
44
65
  this.ready = false;
45
66
  }
46
67
  async tabsList() {
47
- return [this.tab()];
68
+ this.tabsListCalls++;
69
+ if (this.tabsListThrows)
70
+ throw new types_1.ExecutorError('EXTENSION_DISCONNECTED', 'stub bridge is down');
71
+ return this.noTabs ? [] : [this.tab()];
48
72
  }
49
73
  async tabSelect(tabId) {
50
74
  return { ...this.tab(), tabId };
@@ -155,6 +155,14 @@ export interface Executor {
155
155
  ping(deadlineMs?: number): Promise<boolean>;
156
156
  /** Close ONLY if we own the browser; never the user's Chrome. */
157
157
  dispose(): Promise<void>;
158
+ /**
159
+ * The active tab's URL if the backend ALREADY knows it — the extension reports
160
+ * it on every result frame, so a gate that runs right after a command needs no
161
+ * round-trip. Null means "not known recently enough", and the caller must then
162
+ * resolve it properly; it never returns a guess. Optional: backends that can't
163
+ * report cheaply simply omit it.
164
+ */
165
+ cachedActiveUrl?(): string | null;
158
166
  tabsList(): Promise<TabInfo[]>;
159
167
  tabSelect(tabId: TabId): Promise<TabInfo>;
160
168
  /** Open a tab. `active` (default true) focuses it; pass false to open in the background. */
@@ -81,19 +81,60 @@ exports.TOOL_DEFINITIONS = [
81
81
  },
82
82
  },
83
83
  ];
84
- /** Resolve the URL the policy should be evaluated against (the active tab). */
84
+ const GATE_CONTEXT = 'cannot resolve the active tab URL for the policy gate';
85
+ /**
86
+ * Resolve the URL the policy should be evaluated against (the active tab).
87
+ *
88
+ * NEVER substitutes a placeholder URL. If the tab list can't be read, the real
89
+ * origin is unknown, and evaluating the policy against a fabricated URL would
90
+ * silently allow or deny against the wrong origin with no signal to the caller.
91
+ * So a `tabsList` failure (or a browser reporting no tabs at all) propagates —
92
+ * the dispatch firewall renders it as a structured error carrying the code.
93
+ *
94
+ * Prefers a URL the backend already reported over asking again: the extension
95
+ * rides the tab's landing URL home on every result frame, which is what keeps a
96
+ * gated call to ONE round-trip instead of two.
97
+ */
85
98
  async function activeUrl(ex) {
99
+ const known = ex.cachedActiveUrl?.();
100
+ if (known)
101
+ return known;
102
+ let tabs;
86
103
  try {
87
- const tabs = await ex.tabsList();
88
- return tabs.find((t) => t.active)?.url ?? tabs[0]?.url ?? 'about:blank';
104
+ tabs = await ex.tabsList();
89
105
  }
90
- catch {
91
- return 'about:blank';
106
+ catch (err) {
107
+ // Keep the underlying code (TIMEOUT / EXTENSION_DISCONNECTED / …) so the
108
+ // caller can tell a transient bridge failure from a policy decision — the
109
+ // rendered message is prefixed with it, since only the text crosses MCP.
110
+ if (err instanceof types_1.ExecutorError)
111
+ throw new types_1.ExecutorError(err.code, `${GATE_CONTEXT}: ${err.message}`);
112
+ throw err; // an internal fault, not a browser one — don't relabel it
92
113
  }
114
+ const active = tabs.find((t) => t.active) ?? tabs[0];
115
+ if (!active)
116
+ throw new types_1.ExecutorError('TAB_NOT_FOUND', `${GATE_CONTEXT}: the browser reports no open tabs`);
117
+ // An empty URL is Chrome declining to reveal one (a chrome:// page, or a tab
118
+ // the extension has no host access to) — NOT an origin. Gating on '' would
119
+ // produce a baffling "blocked on " denial that reads like a policy decision.
120
+ if (!active.url) {
121
+ throw new types_1.ExecutorError('TAB_NOT_FOUND', `${GATE_CONTEXT}: the active tab (id ${active.tabId}) reports no URL. Chrome hides it for ` +
122
+ `internal pages (chrome://, the Web Store) and until the extension has access to that site — ` +
123
+ `switch to a normal page, or open the target site in a new tab.`);
124
+ }
125
+ return active.url;
93
126
  }
94
- /** Policy chokepoint. `urlOverride` is the destination for navigation. */
127
+ /**
128
+ * Policy chokepoint. `urlOverride` is the destination for navigation.
129
+ *
130
+ * Only resolves the active URL for methods whose verdict actually depends on one
131
+ * (`isUrlGated`). Tab management and the capability gates — eval, downloads,
132
+ * uploads, mutations — are decided without any URL, so making them wait on the
133
+ * tab list bought nothing and, worse, made `tab_new` fail exactly when the tab
134
+ * list was unreadable: the one call that could dig you out.
135
+ */
95
136
  async function gate(ctx, method, urlOverride) {
96
- const url = urlOverride ?? (await activeUrl(ctx.ex));
137
+ const url = urlOverride ?? ((0, policy_1.isUrlGated)(method) ? await activeUrl(ctx.ex) : '');
97
138
  (0, policy_1.assertUrlAllowed)(url, method, ctx.policy);
98
139
  }
99
140
  const tabId = (args) => (0, validators_1.optionalString)(args, 'tabId');
@@ -349,7 +390,12 @@ exports.TOOL_HANDLERS = {
349
390
  // Dispatch (never-throw firewall)
350
391
  // ---------------------------------------------------------------------------
351
392
  function errMessage(err) {
352
- if (err instanceof validators_1.McpToolError || err instanceof types_1.ExecutorError)
393
+ // Only the text crosses the MCP boundary, so the code has to travel inside it —
394
+ // otherwise a caller cannot tell EXTENSION_DISCONNECTED (retry in a moment)
395
+ // from POLICY_DENIED (retrying will never help).
396
+ if (err instanceof types_1.ExecutorError)
397
+ return `[${err.code}] ${err.message}`;
398
+ if (err instanceof validators_1.McpToolError)
353
399
  return err.message;
354
400
  if (err instanceof Error)
355
401
  return `internal error: ${err.message}`;
@@ -2,6 +2,7 @@
2
2
  (() => {
3
3
  // shared/protocol.ts
4
4
  var PROTOCOL_VERSION = 1;
5
+ var WIRE_CAP_TAB_URL = "tab-url";
5
6
  var WIRE_METHODS = [
6
7
  "tabs_list",
7
8
  "tab_select",
@@ -63,7 +64,11 @@
63
64
  v: PROTOCOL_VERSION,
64
65
  token,
65
66
  ext: { id: chrome.runtime.id, version: chrome.runtime.getManifest().version, chrome: chromeVersion() },
66
- profile: profile && profile.trim() ? profile.trim() : void 0
67
+ profile: profile && profile.trim() ? profile.trim() : void 0,
68
+ // This build gates fail-closed and reports tab URLs on results, so the
69
+ // server may skip its pre-flight tabs_list. An older build omits this and
70
+ // the server keeps fetching the URL itself.
71
+ caps: [WIRE_CAP_TAB_URL]
67
72
  };
68
73
  ws2.send(JSON.stringify(hello));
69
74
  };
@@ -494,6 +499,9 @@
494
499
  const u = cmd.params.url;
495
500
  return typeof u === "string" ? u : "";
496
501
  }
502
+ return observedTabUrl(cmd);
503
+ }
504
+ async function observedTabUrl(cmd) {
497
505
  try {
498
506
  const tabId = await targetTab(cmd);
499
507
  const t = await chrome.tabs.get(tabId);
@@ -1124,6 +1132,12 @@
1124
1132
  async dispatch(cmd) {
1125
1133
  try {
1126
1134
  const policy = this.deps.getPolicy();
1135
+ if (!policy && cmd.method !== "ping_probe") {
1136
+ throw new CmdError(
1137
+ "POLICY_DENIED",
1138
+ "no policy has arrived from the chrome-mcp server yet, so this extension is refusing every command"
1139
+ );
1140
+ }
1127
1141
  if (policy) {
1128
1142
  const url = isUrlGated(cmd.method) ? await urlForCommand(cmd) : "";
1129
1143
  const verdict = evaluatePolicy(url, cmd.method, policy);
@@ -1131,6 +1145,8 @@
1131
1145
  }
1132
1146
  const data = await this.deps.exec.run(cmd);
1133
1147
  const frame = { type: "result", v: PROTOCOL_VERSION, id: cmd.id, ok: true, data };
1148
+ const tabUrl = await observedTabUrl(cmd);
1149
+ if (tabUrl) frame.tabUrl = tabUrl;
1134
1150
  this.deps.send(frame);
1135
1151
  } catch (err) {
1136
1152
  const code = err instanceof CmdError ? err.code : "CDP_ERROR";
package/package.json CHANGED
@@ -1,7 +1,7 @@
1
1
  {
2
2
  "name": "@mehmoodqureshi/chrome-mcp",
3
- "version": "0.6.3",
4
- "description": "Drive a real Chrome browser over MCP. A stdio MCP server (CLI) plus an MV3 extension, behind one pluggable Executor (extension via chrome.scripting, or a Playwright CDP fallback).",
3
+ "version": "0.6.5",
4
+ "description": "Drive your real Chrome browser over MCP — real logins, real cookies. A stdio MCP server (CLI) plus an MV3 extension, driving Chrome via chrome.scripting/chrome.tabs. Multi-tab batch automation, accessibility snapshots, deny-all security by default.",
5
5
  "author": "Mehmood Ur Rehman Qureshi",
6
6
  "license": "MIT",
7
7
  "homepage": "https://github.com/Mehmoodqureshi/chrome-mcp#readme",
@@ -32,12 +32,20 @@
32
32
  "keywords": [
33
33
  "mcp",
34
34
  "model-context-protocol",
35
+ "mcp-server",
35
36
  "chrome",
36
37
  "chrome-extension",
37
38
  "browser-automation",
38
- "cdp",
39
- "playwright",
40
- "claude"
39
+ "browser-mcp",
40
+ "real-browser",
41
+ "logged-in",
42
+ "session",
43
+ "cookies",
44
+ "authenticated",
45
+ "web-scraping",
46
+ "ai-agent",
47
+ "claude",
48
+ "claude-code"
41
49
  ],
42
50
  "scripts": {
43
51
  "build": "tsc -p tsconfig.json",
@@ -39,7 +39,13 @@ function main() {
39
39
  }
40
40
  try {
41
41
  const { execFileSync } = require('node:child_process');
42
- execFileSync('npx', ['playwright', 'install', 'chromium'], { stdio: 'inherit' });
42
+ const { dirname, join } = require('node:path');
43
+ // Drive Playwright's CLI through node rather than the `npx`/`playwright` bin
44
+ // shim: on Windows those are .cmd files, which execFileSync cannot spawn
45
+ // without a shell. `cli.js` is playwright's own bin target; resolving it via
46
+ // package.json avoids the exports map, which exposes no './cli' subpath.
47
+ const cli = join(dirname(require.resolve('playwright/package.json')), 'cli.js');
48
+ execFileSync(process.execPath, [cli, 'install', 'chromium'], { stdio: 'inherit' });
43
49
  } catch (err) {
44
50
  process.stdout.write(
45
51
  `[postinstall] Chromium install skipped (non-fatal): ${err && err.message ? err.message : err}\n`,