proxy-mcp

proxy-mcp is an MCP server that runs an explicit HTTP/HTTPS MITM proxy (L7). It captures requests/responses, lets you modify traffic in-flight (headers/bodies/mock/forward/drop), supports upstream proxy chaining, and records TLS fingerprints for connections to the proxy (JA3/JA4) plus optional upstream server JA3S. Ships "interceptors" to route the cloakbrowser stealth browser, CLI tools, and Docker containers through the proxy, plus Playwright-driven browser automation with locator-based click, typing, scroll, and ARIA snapshots.

72 tools + 7 resources + 3 resource templates. Built on mockttp and cloakbrowser.

Important

proxy-mcp is no longer published to npmjs. npm's publishing system has become too annoying to be worth it — expiring tokens that silently break releases, and a mandatory passkey/2FA enrollment maze just to change a package setting. Distribution now happens directly from this repository:

npx -y "github:yfe404/proxy-mcp#semver:^3"

Versions ≤ 3.3.2 remain on npmjs but will never be updated — npx -y proxy-mcp@latest silently stays stale. Update your MCP config to the GitHub form above. See Setup.

Setup

Quick install (Claude Code)

claude mcp add proxy-mcp -- npx -y "github:yfe404/proxy-mcp#semver:^3"

This installs proxy-mcp as an MCP server using stdio transport, straight from GitHub — no registry involved. The #semver:^3 range resolves against this repo's version tags, so new releases are picked up automatically.

Note: the first install compiles from source (a prepare build), so it takes a few seconds longer than a registry tarball — subsequent launches use the npx cache. Why not npmjs anymore? See the announcement at the top: their publishing system got too annoying.

Scopes:

# Per-user (available in all projects)
claude mcp add --scope user proxy-mcp -- npx -y "github:yfe404/proxy-mcp#semver:^3"

# Per-project (shared via .mcp.json, commit to repo)
claude mcp add --scope project proxy-mcp -- npx -y "github:yfe404/proxy-mcp#semver:^3"

To pin an exact release instead, use github:yfe404/proxy-mcp#v3.4.0.

Prerequisites

Node.js 20+

From source (development)

git clone https://github.com/yfe404/proxy-mcp.git
cd proxy-mcp
npm install
npm run build

# stdio transport (default) — used by MCP clients like Claude Code
node dist/index.js

# Streamable HTTP transport — exposes /mcp endpoint for scripting
node dist/index.js --transport http --port 3001

--transport and --port also accept env vars TRANSPORT and PORT.

PROXY_MCP_DEBUG=1 adds debug lines to stderr, one per request the client cancels: request aborted by client: <method> <url>.

PROXY_MCP_UPSTREAM_PASSWORD and PROXY_MCP_UPSTREAM_HOST keep an upstream proxy password out of the transcript — see Keeping the upstream password out of the transcript.

Manual MCP configuration

The configured server alias controls Claude's generated tool prefix. The examples below use proxy-mcp, so Claude Code exposes tools as mcp__proxy-mcp__<tool_name>. If you rename the server key to proxy, use mcp__proxy__<tool_name> instead.

Claude Code CLI:

# stdio (default)
claude mcp add proxy-mcp -- npx -y "github:yfe404/proxy-mcp#semver:^3"

# From local clone
claude mcp add proxy-mcp -- node /path/to/proxy-mcp/dist/index.js

# HTTP transport for scripting
claude mcp add --transport http proxy-mcp http://127.0.0.1:3001/mcp

.mcp.json (project-level, commit to repo):

{
  "mcpServers": {
    "proxy-mcp": {
      "command": "npx",
      "args": ["-y", "github:yfe404/proxy-mcp#semver:^3"]
    }
  }
}

Streamable HTTP transport:

{
  "mcpServers": {
    "proxy-mcp": {
      "type": "streamable-http",
      "url": "http://127.0.0.1:3001/mcp"
    }
  }
}

HTTP Proxy Configuration

1) Start proxy and get endpoint

proxy_start

Use the returned port and endpoint http://127.0.0.1:<port>.

2) Browser setup (recommended: interceptor)

Use the browser interceptor so proxy flags and cert trust are configured automatically. Launches cloakbrowser — a stealth-patched Chromium with source-level C++ fingerprint patches and humanize mode on by default:

interceptor_browser_launch --url "https://example.com"

Drive the page with Playwright-backed tools (no CDP, no sidecar — target_id is all you need):

interceptor_browser_navigate --target_id "browser_<id>" --url "https://apify.com"
interceptor_browser_snapshot  --target_id "browser_<id>"
interceptor_browser_screenshot --target_id "browser_<id>" --file_path "/tmp/shot.png"

3) Browser setup (manual fallback)

If launching a browser manually, pass the proxy flag yourself:

google-chrome --proxy-server="http://127.0.0.1:<port>"

4) CLI/process setup

Route any process through proxy-mcp by setting proxy env vars:

export HTTP_PROXY="http://127.0.0.1:<port>"
export HTTPS_PROXY="http://127.0.0.1:<port>"
export NO_PROXY="localhost,127.0.0.1"

If the client verifies TLS, trust the proxy-mcp CA certificate (see proxy_get_ca_cert) or use the Terminal interceptor (interceptor_spawn) which sets proxy env vars plus common CA env vars (curl, Node, Python requests, Git, npm/yarn, etc.):

interceptor_spawn --command curl --args '["-s","https://example.com"]'

Explicit curl examples:

curl --proxy http://127.0.0.1:<port> http://example.com
curl --proxy http://127.0.0.1:<port> https://example.com

5) Upstream proxy chaining

Set optional proxy chaining from proxy-mcp to another upstream proxy (for geolocation, auth, or IP reputation):

Client/app  →  proxy-mcp (local explicit proxy)  →  upstream proxy (optional chaining layer)

proxy_set_upstream --proxy_url "socks5://user:pass@upstream.example:1080"

Supported upstream URL schemes: socks4://, socks5://, http://, https://, pac+http://.

Keeping the upstream password out of the transcript

Tool calls and tool results are both persisted by the MCP client. To avoid writing an upstream password there on every call, set it in the server's environment and pass a URL with a username but no password.

Two variables are required, and both have to be in the environment of the server process, which the MCP client spawns. Exporting them in the shell you launch the client from may reach it — CLI clients pass their own environment through — but that depends on the client and is lost the moment the server is started any other way. Put them in the client's server config:

variable	meaning
`PROXY_MCP_UPSTREAM_PASSWORD`	the password to fill in
`PROXY_MCP_UPSTREAM_HOST`	the only hostname it may be sent to — a bare hostname, no scheme, port or path

claude mcp add proxy-mcp \
  -e PROXY_MCP_UPSTREAM_PASSWORD=s3cret \
  -e PROXY_MCP_UPSTREAM_HOST=upstream.example \
  -- npx -y "github:yfe404/proxy-mcp#semver:^3"

{
  "mcpServers": {
    "proxy-mcp": {
      "command": "npx",
      "args": ["-y", "github:yfe404/proxy-mcp#semver:^3"],
      "env": {
        "PROXY_MCP_UPSTREAM_PASSWORD": "s3cret",
        "PROXY_MCP_UPSTREAM_HOST": "upstream.example"
      }
    }
  }
}

Then omit the password from the call:

proxy_set_upstream --proxy_url "http://user@upstream.example:1080"
# routes as http://user:s3cret@upstream.example:1080

Why the host variable exists. Without it, a caller who cannot read the password could still name any host and have the password delivered there — the proxy sends it on the first request, and the transcript would show only ***. The hostname is matched case-insensitively and exactly, with no wildcards; the port is not part of the match, so one variable covers a provider offering several. A URL naming any other host is left alone. If PROXY_MCP_UPSTREAM_PASSWORD is set and PROXY_MCP_UPSTREAM_HOST is not, nothing is merged at all: a half-configuration fails closed rather than becoming an unbound credential.

This keeps the password out of tool arguments and responses, not out of reach. interceptor_spawn runs an arbitrary command as the server user, so a caller can read the client config file the password is configured in — and on Linux /proc/<pid>/environ. The variable removes the routine exposure of writing a credential into every tool call; it is not a sandbox, and anyone who can call interceptor_spawn should be treated as able to obtain the password.

The response reports which credential was used — passwordSource is env, url or none. none means no password was applied to a URL that names a user: either the credential is genuinely username-only, or the server does not have both variables set for this host. The field is omitted for a URL with no username, where the question does not arise.

Applies to proxy_set_upstream and proxy_set_host_upstream. A URL that already carries a password is used as-is, so existing calls are unaffected. One credential covers all upstreams at the pinned host; a URL without a username is left alone.

Username-only credentials at the pinned host cannot be expressed. A URL with a username and no password is exactly the syntax that requests the merge, and user:@host cannot signal otherwise — the URL parser erases the empty password before the server sees it. If the pinned host authenticates on the username alone, unset PROXY_MCP_UPSTREAM_PASSWORD for that server.

socks*:// upstreams: no : in the password. socks-proxy-agent splits the credential on the first : and keeps only what follows, so pa:ss would authenticate as pa. Rather than deliver half a password silently, a socks upstream is refused with an error when PROXY_MCP_UPSTREAM_PASSWORD contains :. The truncation itself is a toolchain limitation, not something this introduces — a literal socks5://user:pa%3Ass@host:1080 truncates the same way, and nothing can guard that. http://, https:// and pac+http:// upstreams take the whole password.

A : in the username is refused on every scheme. Basic auth splits the decoded pair at the first colon (RFC 7617), and socks-proxy-agent does the same, so a username of gro:ups with password s3cret reaches the proxy as user gro, password ups:s3cret — the merged password silently discarded. No scheme can carry it, so the merge refuses rather than guess. Put the whole credential in proxy_url instead.

Responses redact credentials — the password in userinfo, with path segments masked and the query and fragment dropped, since a pac+http:// token may live in any of those:

Global upstream set to http://user:***@upstream.example:1080/
Global upstream set to pac+http://pac.example.com/***

proxy_status and the proxy://status resource are redacted the same way. A PAC URL's filename is masked along with the rest of the path, so a confirmation message shows the host and nothing else.

The username is not redacted. For several providers it is configuration rather than a secret — Apify Proxy encodes proxy group, country and sticky-session id there — and showing it is what makes the confirmation useful. If your provider puts a secret in the username field, do not rely on these messages being safe to share.

Typical geo-routing examples:

# Route ALL outgoing traffic from proxy-mcp via a geo proxy
proxy_set_upstream --proxy_url "socks5://user:pass@fr-exit.example.net:1080"

# Bypass upstream for local/internal hosts
proxy_set_upstream --proxy_url "http://user:pass@proxy.example.net:8080" --no_proxy '["localhost","127.0.0.1",".corp.local"]'

# Route only one hostname via a dedicated upstream (overrides global)
proxy_set_host_upstream --hostname "api.example.com" --proxy_url "https://user:pass@us-exit.example.net:443"

# Remove overrides when done
proxy_remove_host_upstream --hostname "api.example.com"
proxy_clear_upstream

For HTTPS MITM, the proxy CA must be trusted in the target environment (proxy_get_ca_cert).

6) Validate and troubleshoot quickly

proxy_list_traffic --limit 20
proxy_search_traffic --query "example.com"

Common issues:

Traffic from the wrong browser instance (fix: always pass target_id from interceptor_browser_launch)
HTTPS cert trust missing on target
NO_PROXY bypassing expected hosts
First launch is slow: cloakbrowser downloads a ~200 MB stealth Chromium binary on first use (cached afterwards)

7) HAR import + replay

Import HAR into a persisted session, then analyze with existing session query/findings tools:

proxy_import_har --har_file "/path/to/capture.har" --session_name "imported-run"
proxy_list_sessions
proxy_query_session --session_id SESSION_ID --hostname_contains "api.example.com"
proxy_get_session_handshakes --session_id SESSION_ID

Replay defaults to dry-run (preview only). Execute requires explicit mode:

# Preview what would be replayed
proxy_replay_session --session_id SESSION_ID --mode dry_run --limit 20

# Execute replay against original hosts
proxy_replay_session --session_id SESSION_ID --mode execute --limit 20

# Optional: override target host/base URL while preserving path+query
proxy_replay_session --session_id SESSION_ID --mode execute --target_base_url "http://127.0.0.1:8081"

Note: imported HAR entries (and entries created by proxy_replay_session) do not carry JA3/JA4/JA3S handshake metadata. Use live proxy-captured traffic to analyze handshake fingerprints.

Boundaries

Only sees traffic configured to route through it (not a network tap or packet sniffer)
Spoofs outgoing JA3 + HTTP/2 fingerprint + header order (via impit — native Rust TLS impersonation), not JA4 (JA4 is capture-only)
Can add, overwrite, or delete HTTP headers; outgoing header order can be controlled via fingerprint spoofing
Returns its own CA certificate — does not expose upstream server certificate chains

TLS ClientHello Passthrough (browser via interceptor)

When cloakbrowser is launched via interceptor_browser_launch, proxy-mcp forwards the browser's original TLS ClientHello to the upstream server for document loads and same-origin sub-resource requests. The target server sees an authentic Chrome TLS fingerprint — not the proxy's.

This is a key difference from typical MITM proxies (mitmproxy, Charles, Fiddler) which re-terminate TLS with their own fingerprint, making MITM trivially detectable by anti-bot systems via JA3/JA4 analysis.

How to verify passthrough is working:

proxy_list_tls_fingerprints --hostname_filter "example.com"

JA3 varies across requests to the same host — this is expected; Chrome randomizes cipher suite order per-connection (feature since Chrome 110+)
JA4 stays stable — same cipher/extension set, just different ordering
JA3 variation + JA4 stability = authentic Chrome TLS passthrough confirmed

When passthrough applies vs. when spoofing is needed:

Traffic source	TLS behavior	Action needed
cloakbrowser via `interceptor_browser_launch` (document loads, same-origin)	Browser's native ClientHello forwarded (passthrough)	None — fingerprint is authentic
cloakbrowser via `interceptor_browser_launch` (cross-origin sub-resources, when spoof active)	Re-issued via impit with spoofed TLS	`proxy_set_fingerprint_spoof` with a browser preset
Non-browser clients (curl, Python, `interceptor_spawn`)	Proxy's own TLS	`proxy_set_fingerprint_spoof` or `proxy_set_ja3_spoof` required
HAR replay (`proxy_replay_session`)	Proxy's own TLS	`proxy_set_fingerprint_spoof` required

Built on stealth browsers + Playwright

Browser automation uses cloakbrowser for stealth-patched Chromium, driven through Playwright. There is no CDP sidecar or hand-rolled stealth script in proxy-mcp. Downstream tools take a browser_* target from interceptor_browser_launch.

Capability	proxy-mcp
See/modify DOM, run JS in page	`interceptor_browser_evaluate` (run JS file, return value), `interceptor_browser_inject_init_script` (pre-document hook, every navigation), `interceptor_browser_add_script_tag` (DOM-visible — avoid for stealth); plus `interceptor_browser_snapshot` for ARIA reads
Read cookies, localStorage, sessionStorage	Yes — `interceptor_browser_list_cookies`, `interceptor_browser_list_storage_keys`
Capture HTTP request/response bodies	Via the MITM proxy (4 KB preview cap by default; `full` capture profile on persisted sessions stores complete bodies)
Modify requests in-flight (headers, body, mock, drop)	Yes (declarative rules, hot-reload)
Upstream proxy chaining (geo, auth)	Global + per-host upstreams across all clients (SOCKS4/5, HTTP, HTTPS, PAC)
TLS fingerprint capture (JA3/JA4/JA3S)	Yes
JA3 + HTTP/2 fingerprint spoofing	Proxy-side (impit re-issues matching requests with spoofed TLS 1.3, HTTP/2 frames, and header order)
Intercept non-browser traffic (curl, Python, Docker containers)	Yes (interceptors)
Human-like mouse/keyboard/scroll input	`humanizer_` tools call Playwright mouse/keyboard primitives on `browser_` targets. Cloakbrowser's own humanize patches apply when enabled at launch.
Locator-based interaction	`humanizer_click` accepts CSS/XPath selector, ARIA role + name, visible text, or form label — no pixel guessing

Standard flow:

Call proxy_start
Optionally enable outbound fingerprint spoofing for cross-origin sub-resources: proxy_set_fingerprint_spoof --preset chrome_136
Call interceptor_browser_launch --url "https://example.com"
Drive the page: interceptor_browser_navigate, interceptor_browser_snapshot, humanizer_click --selector "...", humanizer_type --text "..."
Inspect traffic: proxy_search_traffic --query "<hostname>"

Tools Reference

Lifecycle (4)

Tool	Description
`proxy_start`	Start MITM proxy, auto-generate CA cert
`proxy_stop`	Stop proxy (traffic/cert retained). Deactivates this MCP session's interceptor targets; `all: true` deactivates every session's
`proxy_status`	Running state, port, rule/traffic counts
`proxy_get_ca_cert`	CA certificate PEM + SPKI fingerprint

Upstream Proxy (4)

Tool	Description
`proxy_set_upstream`	Set global upstream proxy
`proxy_clear_upstream`	Remove global upstream
`proxy_set_host_upstream`	Per-host upstream override
`proxy_remove_host_upstream`	Remove per-host override

Interception Rules (7)

Tool	Description
`proxy_add_rule`	Add rule with matcher + handler
`proxy_update_rule`	Modify existing rule
`proxy_remove_rule`	Delete rule
`proxy_list_rules`	List all rules by priority
`proxy_test_rule_match`	Test which rules would match a simulated request or captured exchange, with detailed diagnostics
`proxy_enable_rule`	Enable a disabled rule
`proxy_disable_rule`	Disable without removing

Quick debugging examples:

# Simulate a request and see which rule would win
proxy_test_rule_match --mode simulate --request '{"method":"GET","url":"https://example.com/api/v1/items","headers":{"accept":"application/json"}}'

# Evaluate a real captured exchange by ID
proxy_test_rule_match --mode exchange --exchange_id "ex_abc123"

Traffic Capture (4)

Tool	Description
`proxy_list_traffic`	Paginated traffic list with filters
`proxy_get_exchange`	Full exchange details by ID
`proxy_search_traffic`	Full-text search across traffic
`proxy_clear_traffic`	Clear capture buffer

Modification Shortcuts (3)

Tool	Description
`proxy_inject_headers`	Add/overwrite/delete headers on matching traffic (set value to `null` to remove a header)
`proxy_rewrite_url`	Rewrite request URLs
`proxy_mock_response`	Return mock response for matched requests

TLS Fingerprinting (9)

Tool	Description
`proxy_get_tls_fingerprints`	Get JA3/JA4 client fingerprints + JA3S for a single exchange
`proxy_list_tls_fingerprints`	List unique JA3/JA4 fingerprints across all traffic with counts
`proxy_set_ja3_spoof`	Legacy: enable JA3 spoofing (deprecated, use `proxy_set_fingerprint_spoof`)
`proxy_clear_ja3_spoof`	Disable fingerprint spoofing
`proxy_get_tls_config`	Return current TLS config (server capture, JA3 spoof state)
`proxy_enable_server_tls_capture`	Toggle server-side JA3S capture (monkey-patches `tls.connect`)
`proxy_set_fingerprint_spoof`	Enable full TLS + HTTP/2 fingerprint spoofing via impit. Supports browser presets.
`proxy_list_fingerprint_presets`	List available browser fingerprint presets (e.g. `chrome_131`, `chrome_136`, `chrome_136_linux`, `firefox_133`)
`proxy_check_fingerprint_runtime`	Check fingerprint spoofing backend readiness

Fingerprint spoofing works by re-issuing the request from the proxy via impit (native Rust TLS/HTTP2 impersonation via rustls). TLS 1.3 and HTTP/2 fingerprints (SETTINGS, WINDOW_UPDATE, PRIORITY frames) match real browsers by construction. The origin server sees the proxy's spoofed TLS, HTTP/2, and header order — not the original client's. When a user_agent is set (including via presets), proxy-mcp also normalizes Chromium UA Client Hints headers (sec-ch-ua*) to match the spoofed User-Agent (forwarding contradictory hints is a common bot signal). Browser exception: when cloakbrowser is launched via interceptor_browser_launch, document loads and same-origin requests use the browser's native TLS (no impit), preserving fingerprint consistency for bot detection challenges. Only cross-origin sub-resource requests are re-issued with spoofed TLS. Non-browser clients (curl, spawn, HAR replay) get full TLS + UA spoofing on all requests. Use proxy_set_fingerprint_spoof with a browser preset for one-command setup. proxy_set_ja3_spoof is kept for backward compatibility but custom JA3 strings are ignored (the preset's impit browser target is used instead). JA4 fingerprints are captured (read-only) but spoofing is not supported.

Interceptors (10)

Interceptors configure targets (browsers, processes, containers) to route their traffic through the proxy automatically.

Discovery (3)

Tool	Description
`interceptor_list`	List all interceptors with availability and active target counts
`interceptor_status`	Detailed status of a specific interceptor
`interceptor_deactivate_all`	Emergency cleanup: kill all active interceptors across all types

Browser (3)

Tool	Description
`interceptor_browser_launch`	Launch cloakbrowser (stealth Chromium) with proxy flags, SPKI cert trust, built-in humanize mode
`interceptor_browser_navigate`	Navigate the bound page via Playwright `page.goto` and verify proxy capture
`interceptor_browser_close`	Close a browser instance by target ID

Stealth is source-level: cloakbrowser ships 48+ C++ patches so ja3n/ja4/akamai match real Chrome, navigator.webdriver is false, audio/canvas/WebGL fingerprints match real hardware. No JS stealth injection needed. First launch downloads a ~200 MB Chromium binary (cached afterwards).

Terminal / Process (2)

Tool	Description
`interceptor_spawn`	Spawn a command with proxy env vars pre-configured (HTTP_PROXY, SSL certs, etc.)
`interceptor_kill`	Kill a spawned process and retrieve stdout/stderr

Sets 18+ env vars covering curl, Node.js, Python requests, Deno, Git, npm/yarn.

Docker (2)

Tool	Description
`interceptor_docker_attach`	Inject proxy env vars and CA cert into running container
`interceptor_docker_detach`	Remove proxy config from container

Two modes: exec (live injection, existing processes need restart) and restart (stop + restart container). Uses host.docker.internal for proxy URL.

Browser DevTools-equivalents (12)

Playwright-driven tools for the browser target. Each takes a target_id directly — no session binding, no sidecar.

Tool	Description
`interceptor_browser_snapshot`	ARIA/role YAML snapshot of the page (or selector subtree) — optimized for LLM page reasoning
`interceptor_browser_screenshot`	Screenshot. Writes to `file_path` if provided; otherwise reports byte count only
`interceptor_browser_list_console`	Buffered console messages since launch, with type/text filters and pagination
`interceptor_browser_list_cookies`	Cookie listing with filters, pagination, truncated value previews
`interceptor_browser_get_cookie`	Get one cookie by `cookie_id` (value is capped to keep output bounded)
`interceptor_browser_list_storage_keys`	localStorage/sessionStorage key listing with value previews
`interceptor_browser_get_storage_value`	Get one storage value by `item_id`
`interceptor_browser_list_network_fields`	Header field listing from proxy-captured traffic since the browser was launched
`interceptor_browser_get_network_field`	Get one full header field value by `field_id`
`interceptor_browser_evaluate`	Run a JS file in the page (file body wrapped as `(__args) => { ... }`); returns the result. Runs in the isolated utility world
`interceptor_browser_inject_init_script`	Inject a JS file as `page.addInitScript` — runs before every page script on the next navigation. Injected into the isolated utility world
`interceptor_browser_add_script_tag`	Append a `<script>` to the current page. DOM-visible — avoid for stealth. Use for benign payloads where main-world execution + page visibility is intentional

Network data is sourced from the MITM proxy rather than a browser-side protocol — the proxy sees every wire request regardless of what the browser reported.

Stealth tradeoffs for JS injection:

Method	Cloakbrowser
`evaluate`	Safe (isolated utility world) — rate-limit before reCAPTCHA, each call is CDP traffic
`inject_init_script`	Best for stealth — pre-document, no DOM artifact
`add_script_tag`	Detectable (DOM node, MutationObserver, CSP)

References: Playwright evaluate, Playwright addInitScript.

Worlds and isolation — what your JS can and can't see

Playwright's evaluate runs in an isolated "utility" world that shares globals with the page's main world. An addInitScript patch to navigator.webdriver is visible to (a) your subsequent evaluate probes AND (b) anti-bot code the site loads. This is the model most "stealth playbooks" assume. Detection vectors are CDP-side (Runtime.evaluate chatter) — cloakbrowser's C++ patches mitigate those.

Practical rules:

Use case	Tool
Read DOM / extract data	`interceptor_browser_evaluate`
Modify page state, click via JS	`interceptor_browser_evaluate` (globals are shared with the page)
Spoof navigator / window fingerprints	`interceptor_browser_inject_init_script`
Load a 3rd-party JS lib into the page	`interceptor_browser_add_script_tag` (page sees it — usually OK if intentional)

Sessions (14)

Persistent, queryable on-disk capture for long runs and post-crash analysis.

Tool	Description
`proxy_session_start`	Start persistent session capture (preview or full-body mode)
`proxy_session_stop`	Stop and finalize the active persistent session
`proxy_session_status`	Runtime status for persistence (active session, bytes, disk cap errors)
`proxy_import_har`	Import a HAR file from disk into a new persisted session
`proxy_list_sessions`	List recorded sessions from disk
`proxy_get_session`	Get manifest/details for one session
`proxy_query_session`	Indexed query over recorded exchanges
`proxy_search_session_bodies`	Search request/response bodies stored in a persistent session, with context snippets
`proxy_get_session_handshakes`	Report JA3/JA4/JA3S handshake metadata availability for session entries
`proxy_get_session_exchange`	Fetch one exchange from a session (with optional full bodies)
`proxy_replay_session`	Dry-run or execute replay of selected session requests
`proxy_export_har`	Export full session or filtered subset to HAR
`proxy_delete_session`	Delete a stored session
`proxy_session_recover`	Rebuild indexes from records after unclean shutdown

proxy_get_session_exchange and proxy_export_har automatically decompress response bodies (gzip, deflate, brotli) based on the stored content-encoding header. The returned responseBodyText and responseBodyBase64 contain the decompressed content. Raw compressed bytes are preserved on disk for exact replay fidelity.

Note on proxy_start with persistence_enabled: true: this auto-creates a session. A subsequent proxy_session_start() call returns the existing active session instead of failing — no need to stop and re-start.

Humanizer — Playwright Input (5)

Human-like browser input via Playwright page.mouse / page.keyboard. Works with browser_* targets from interceptor_browser_launch. Cloakbrowser's own humanize patches apply when enabled at launch.

Tool	Description
`humanizer_move`	Move the mouse to `x,y` through the backend Playwright page
`humanizer_click`	Click a locator (`selector` / `role` + `name` / `text` / `label`) or raw `x,y`. Auto-waits for visible + enabled + stable + in-view before clicking
`humanizer_type`	Type text into the focused element via `page.keyboard.type`; optional `delay_ms` passes through to Playwright
`humanizer_scroll`	Dispatch one Playwright `page.mouse.wheel` event
`humanizer_idle`	Simulate idle behavior with mouse micro-jitter and occasional micro-scrolls to defeat idle detection

All tools require target_id from a prior interceptor_browser_launch. The engine maintains tracked mouse position across calls for coordinate-based move/click/idle behavior.

Behavioral details:

Mouse: humanizer_move calls page.mouse.move; locator clicks call Playwright locators and raw-coordinate clicks call page.mouse.click
Typing: humanizer_type calls page.keyboard.type(text, { delay }) when delay_ms is provided; no WPM, typo, or bigram model is implemented in proxy-mcp
Scrolling: humanizer_scroll sends one wheel event with the requested delta
Idle: Periodic micro-jitter (±3px subtle / ±8px normal) and random micro-scrolls at configurable intensity

Resources

URI	Description
`proxy://status`	Proxy running state and config
`proxy://ca-cert`	CA certificate PEM
`proxy://traffic/summary`	Traffic stats: method/status breakdown, top hostnames, TLS fingerprint stats
`proxy://interceptors`	All interceptor metadata and activation status
`proxy://sessions`	Persistent session catalog + runtime persistence status
`proxy://browser/primary`	Current page URL/title for the most recently launched browser instance
`proxy://browser/targets`	Current page state for all active browser instances
`proxy://sessions/{session_id}/summary`	Aggregate stats for one recorded session (resource template)
`proxy://sessions/{session_id}/timeline`	Time-bucketed request/error timeline (resource template)
`proxy://sessions/{session_id}/findings`	Top errors/slow exchanges/host error rates (resource template)

Usage Example

# Start the proxy
proxy_start

# Optional: start persistent session recording
proxy_session_start --capture_profile full --session_name "reverse-run-1"

# Use interceptors to auto-configure targets:
interceptor_browser_launch                    # Launch stealth browser with proxy
interceptor_spawn --command curl --args '["https://example.com"]'  # Spawn proxied process

# Set upstream proxy for geolocation
proxy_set_upstream --proxy_url socks5://user:pass@geo-proxy:1080

# Mock an API response
proxy_mock_response --url_pattern "/api/v1/config" --status 200 --body '{"feature": true}'

# Inject auth headers (set value to null to delete a header)
proxy_inject_headers --hostname "api.example.com" --headers '{"Authorization": "Bearer token123"}'

# View captured traffic
proxy_list_traffic --hostname_filter "api.example.com"
proxy_search_traffic --query "error"

# TLS fingerprinting
proxy_list_tls_fingerprints                # See unique JA3/JA4 fingerprints
proxy_set_ja3_spoof --ja3 "771,4865-..."   # Spoof outgoing JA3 (for non-browser clients)
proxy_set_fingerprint_spoof --preset chrome_136 --host_patterns '["example.com"]'  # Full fingerprint spoof
proxy_list_fingerprint_presets                  # Available browser presets

# Human-like browser interaction (browser_* target)
humanizer_move   --target_id "browser_<id>" --x 500 --y 300
humanizer_click  --target_id "browser_<id>" --selector "#login-button"
humanizer_click  --target_id "browser_<id>" --role "button" --name "Sign in"
humanizer_type   --target_id "browser_<id>" --text "user@example.com" --delay_ms 45
humanizer_scroll --target_id "browser_<id>" --delta_y 300
humanizer_idle   --target_id "browser_<id>" --duration_ms 2000 --intensity subtle

# Run / inject JS in the page
interceptor_browser_evaluate           --target_id "browser_<id>" --script_path /tmp/probe.js
interceptor_browser_inject_init_script --target_id "browser_<id>" --script_path /tmp/hook.js   # applies on next navigation
interceptor_browser_add_script_tag     --target_id "browser_<id>" --script_path /tmp/lib.js    # DOM-visible — avoid for stealth

# Query/export recorded session
proxy_list_sessions
proxy_query_session --session_id SESSION_ID --hostname_contains "api.example.com"
proxy_export_har --session_id SESSION_ID

Architecture

State: ProxyManager singleton manages mockttp server, rules, traffic
Rule rebuild: Rules must be set before mockttp start(), so rule changes trigger stop/recreate/restart cycle
Traffic capture: on('request') + on('response') events, correlated by request ID
Ring buffer: 1000 entries max, body previews capped at 4KB
TLS capture: Client JA3/JA4 from mockttp socket metadata; server JA3S via tls.connect monkey-patch
TLS spoofing: impit (native Rust TLS/HTTP2 impersonation via rustls); in-process, no container needed
Interceptors: Managed by InterceptorManager, each type registers independently
Browser: cloakbrowser (stealth Chromium, ~200 MB binary auto-downloaded on first launch) driven via Playwright BrowserContext / Page
Humanizer: Singleton engine using Playwright's page.mouse / page.keyboard, plus local mouse-position tracking for idle jitter

Testing

npm test              # All tests (unit + integration)
npm run test:unit     # Unit tests only
npm run test:integration  # Integration tests
npm run test:e2e      # E2E fingerprint tests (requires cloakbrowser + internet)

Credits

Core Libraries

Project	Role
mockttp	MITM proxy engine, rule system, CA generation
impit	Native TLS/HTTP2 fingerprint impersonation (Rust via NAPI-RS)
cloakbrowser	Stealth-patched Chromium with source-level C++ fingerprint patches
playwright-core	Browser automation API driving cloakbrowser
@modelcontextprotocol/sdk	MCP server framework

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
scripts		scripts
src		src
test		test
.gitignore		.gitignore
AGENTS.md		AGENTS.md
CHANGELOG.md		CHANGELOG.md
DEMO.md		DEMO.md
README.md		README.md
ROADMAP.md		ROADMAP.md
package-lock.json		package-lock.json
package.json		package.json
tsconfig.json		tsconfig.json

Latest commit

History

Folders and files

Repository files navigation

proxy-mcp

Table of Contents

Setup

Quick install (Claude Code)

Prerequisites

From source (development)

Manual MCP configuration

HTTP Proxy Configuration

1) Start proxy and get endpoint

2) Browser setup (recommended: interceptor)

3) Browser setup (manual fallback)

4) CLI/process setup

5) Upstream proxy chaining

Keeping the upstream password out of the transcript

6) Validate and troubleshoot quickly

7) HAR import + replay

Boundaries

TLS ClientHello Passthrough (browser via interceptor)

Built on stealth browsers + Playwright

Tools Reference

Lifecycle (4)

Upstream Proxy (4)

Interception Rules (7)

Traffic Capture (4)

Modification Shortcuts (3)

TLS Fingerprinting (9)

Interceptors (10)

Discovery (3)

Browser (3)

Terminal / Process (2)

Docker (2)

Browser DevTools-equivalents (12)

Worlds and isolation — what your JS can and can't see

Sessions (14)

Humanizer — Playwright Input (5)

Resources

Usage Example

Architecture

Testing

Credits

Core Libraries

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages