mcp-resource-subscriber

CLI probe for MCP resource subscriptions — connects to any MCP Streamable HTTP server, opens a subscriptions/listen stream for a resource, receives live update notifications, and re-reads updated content.

Protocol: this CLI speaks MCP protocol revision 2026-07-28 only. It pins negotiation to that revision and never falls back to the 2025-era resources/subscribe path — a server that cannot offer 2026-07-28 fails with PROTOCOL_UNSUPPORTED. See docs/protocol-migration.md for the interoperability matrix, the rollout order, and why no legacy path is kept.

Install

# recommended (no install):
pnpm dlx mcp-resource-subscriber --url <mcp-server-url> --uri <resource-uri>

# or install globally:
pnpm add --global mcp-resource-subscriber
mcp-resource-subscriber --url <mcp-server-url> --uri <resource-uri>

パッケージマネージャー方針: このリポジトリと README の利用例は pnpm を前提にしています。pnpm がない環境では、先に pnpm を利用できる状態にしてください。

Note: A reference MCP test server used during compatibility verification is also included in this repository (Docker Compose). See the Lab Server section below.

Related MCP server: SkyStage MCP Test Server

CLI Usage

Against copilot-review-mcp

mcp-resource-subscriber \
  --url http://127.0.0.1:8080/mcp/copilot-review \
  --uri copilot-review://watch/<watch_id> \
  --timeout-ms 900000

copilot-review-mcp is the @scottlz0310/copilot-review-mcp server. Replace <watch_id> with the ID returned by start_copilot_review_watch.

Against the bundled test server

# Start the test server first:
docker compose up --build
# or: pnpm run dev

mcp-resource-subscriber --url http://127.0.0.1:8089/mcp

Note: test://review/status is the default resource URI and is only meaningful against the bundled test server. For any other MCP server, always pass --uri explicitly.

Calling a tool once instead of subscribing

For a single tools/call invocation with no subscription/wait, use the call subcommand — see call mode below.

Options

  --url <url>         MCP server Streamable HTTP endpoint (required)
                      Env: MCP_PROBE_URL
  --uri <uri>         Resource URI to subscribe to
                      Default: test://review/status (bundled test server only)
                      Env: MCP_PROBE_URI
  --auth-token <tok>  Bearer token for Authorization header
                      Prefer MCP_PROBE_AUTH_TOKEN env var (flag is visible in
                      process lists and may be stored in shell history)
                      Env: MCP_PROBE_AUTH_TOKEN (recommended)
  --login             Interactive device-flow login (RFC 8628) against the
                      gateway serving --url. Caches the issued tokens so later
                      runs authenticate and refresh automatically.
                      Cache path env: MCP_PROBE_TOKEN_STORE_PATH
  --logout             Remove the cached token set for the gateway serving --url.
                      Use after a gateway rebuild or DCR store reset so the
                      next --login registers a fresh client.
  --skip-resource-list-check
                      Skip resources/list and assume the URI exists.
                      Use for servers with dynamic resources not in list.
                      Env: MCP_PROBE_SKIP_LIST_CHECK=true
  --timeout-ms <ms>   Notification wait timeout in ms (default: 15000)
                      Env: MCP_PROBE_TIMEOUT_MS
  --json              Emit a single JSON object to stdout instead of line-based output.
                      Diagnostic messages are written to stderr only.
  --version, -v       Print version and exit
  --help, -h          Print this help and exit

Gateway authentication (--login)

When subscribing through an mcp-gateway, run a one-time interactive login instead of provisioning MCP_PROBE_AUTH_TOKEN by hand:

mcp-resource-subscriber --login --url http://127.0.0.1:8080/mcp/subscribe-probe

This performs RFC 7591 dynamic client registration and the RFC 8628 device authorization flow against the gateway origin: it prints a user-code and a verification-uri-complete line, waits while you approve the device in a browser, then caches the issued access_token / refresh_token.

Later probe runs against the same origin then work unattended:

1. An explicit --auth-token / MCP_PROBE_AUTH_TOKEN always wins and skips the cache (existing callers such as MCP_PROBE_AUTH_TOKEN=$(gh auth token) keep working unchanged).

2. Otherwise a cached token for the --url origin is used while still fresh.

3. An expired cached token is renewed automatically via the refresh grant. The gateway rotates refresh tokens on every renewal; the rotated token is persisted immediately.

4. If the refresh token itself is rejected (invalid_grant), the run fails with error-code AUTH_LOGIN_REQUIRED — run --login once more. Transient gateway errors during refresh fail with AUTH_REFRESH_FAILED and can simply be retried.

5. If the gateway no longer recognizes the cached client at all (invalid_client / unauthorized_client — e.g. after a gateway rebuild or DCR store reset), the run also fails with AUTH_LOGIN_REQUIRED. Running --login again automatically registers a fresh client when the cached one is rejected; --logout clears the stale entry outright if you want to force that.

6. Auth resolution (cross-process refresh lock wait + endpoint discovery + refresh grant) is bounded by --timeout-ms, the same budget used for the notification wait. A gateway that accepts the connection but never responds fails with error-code AUTH_TIMEOUT (a plain retry is reasonable) instead of hanging past the requested timeout.

Runs that never used --login do not create the cache and behave exactly as before.

The token cache is a SQLite database (one row per gateway origin) stored under the OS state directory, owner-only permissions:

Override with MCP_PROBE_TOKEN_STORE_PATH. Token values are never printed to stdout/stderr.

JSON output mode

Pass --json to emit a single JSON object to stdout for agent workflow integration:

mcp-resource-subscriber \
  --url http://localhost:3000/mcp \
  --uri queue://review/re-review-requests \
  --timeout-ms 900000 \
  --json

Success output:

{
  "route": "subscription",
  "serverUrl": "http://localhost:3000/mcp",
  "resourceUri": "queue://review/re-review-requests",
  "listenAcknowledged": true,
  "honoredUris": ["queue://review/re-review-requests"],
  "notificationReceived": true,
  "notificationCount": 1,
  "closeReason": "local",
  "errorCode": null,
  "initialText": "...",
  "finalText": "...",
  "recommendedNextAction": null
}

Failure output (same shape with non-null errorCode):

{
  "route": "timeout",
  "serverUrl": "http://localhost:3000/mcp",
  "resourceUri": "queue://review/re-review-requests",
  "listenAcknowledged": true,
  "honoredUris": ["queue://review/re-review-requests"],
  "notificationReceived": false,
  "notificationCount": 0,
  "closeReason": "local",
  "errorCode": "NOTIFICATION_TIMEOUT",
  "initialText": null,
  "finalText": null,
  "recommendedNextAction": null
}

• route: "subscription" | "pre-completion" | "timeout" | "failed"

• listenAcknowledged: true once the server answered subscriptions/listen with notifications/subscriptions/acknowledged

• honoredUris: the resource URIs the server actually honored, taken from that acknowledgement. A requested URI missing here fails with SUBSCRIPTION_NOT_HONORED instead of waiting for a notification that can never arrive

• closeReason: how the listen stream ended — "local" (this CLI closed it), "graceful" (the server ended it deliberately), "remote" (dropped without a response — reported as SUBSCRIPTION_DISCONNECTED), or null

• notificationReceived: true when route === "subscription"

• recommendedNextAction: extracted from finalText if present, otherwise null. On network-level failures (TLS_CERT_UNTRUSTED, DNS_LOOKUP_FAILED, CONNECTION_REFUSED) this is instead a client-generated remediation hint — see Network error classification below.

• If finalText is JSON, callers can parse it themselves

• Diagnostic warnings (e.g. --auth-token flag warning) go to stderr and do not corrupt stdout JSON

call mode (single tools/call invocation)

Invoke any MCP tool once and exit — no subscription, no wait. Reuses the same --url / --auth-token / --login token cache / --timeout-ms / --json flags as subscribe mode:

mcp-resource-subscriber call \
  --url https://gateway.example/mcp/thread-owl \
  --tool enqueue_review \
  --args '{"owner":"scottlz0310","repo":"example","prNumber":123,"reason":"opened"}' \
  --json

Options specific to call mode:

  --tool <name>       MCP tool name to invoke (required)
  --args <json>       JSON object of tool arguments (default: {})

Exit codes are distinct per outcome, so callers can branch on $? alone without parsing stdout:

--json output shape:

{
  "serverUrl": "https://gateway.example/mcp/thread-owl",
  "tool": "enqueue_review",
  "isError": false,
  "errorCode": null,
  "content": [{ "type": "text", "text": "..." }],
  "recommendedNextAction": null
}

content is the raw MCP CallToolResult.content array (verbatim from the server); parse it yourself if it contains JSON text. Line-based (non-JSON) output always prints the same six fields — server-url, tool, is-error, error-code, recommended-next-action, and a content block — for both success and error outcomes, so machine parsers can rely on a single shape: content is the JSON-stringified content array on success, or the literal null when the call never reached the tool (e.g. a communication or auth error); recommended-next-action is null except on network-level errors (see Network error classification).

Note: an unknown tool name (and invalid arguments) is rejected by the server before the tool runs, so it surfaces as exit code 3 / TOOL_REQUEST_REJECTED — not 1 / TOOL_ERROR, which is reserved for a tool that ran and returned isError: true.

Structured line-based output (default)

Every run emits machine-parseable lines:

capabilities {"subscribe":true,"listChanged":false}
resource-found true
resource-uri <resource-uri>
server-url <url>
initial
<initial resource text>
route subscription
listen-acknowledged true
honored-uris ["<resource-uri>"]
notification-received true
notification-count 1
close-reason local
recommended_next_action READ_REVIEW_THREADS
error-code null
notification <resource-uri>
final
<updated resource text>
phase-summary route=subscription url=<url> uri=<uri>

Note: recommended_next_action is only emitted when the final resource text contains it (e.g., from copilot-review-mcp). It is omitted for the bundled test server.

recommended_next_action=POLL_AFTER は非終端状態として扱われます。この場合、CLI は exit せず、同じ subscriptions/listen stream を維持したまま次の notifications/resources/updated を待ちます。--timeout-ms は ack 後の全体待機上限です。

On failure:

error-code SERVER_URL_UNKNOWN
phase-summary route=failed url=unknown error-code=SERVER_URL_UNKNOWN

error-code RESOURCE_NOT_FOUND
phase-summary route=timeout url=<url> uri=<uri> error-code=RESOURCE_NOT_FOUND

error-code NOTIFICATION_TIMEOUT
phase-summary route=timeout url=<url> uri=<uri> error-code=NOTIFICATION_TIMEOUT

error-code PROTOCOL_UNSUPPORTED
recommended-next-action The server does not offer MCP protocol revision 2026-07-28. ...
phase-summary route=failed url=<url> uri=<uri> error-code=PROTOCOL_UNSUPPORTED

Network error classification

Low-level network failures that Node's fetch otherwise flattens into a generic INTERNAL_ERROR / CALL_FAILED are classified into a dedicated errorCode with a recommendedNextAction remediation hint, in both subscribe and call mode (--json and line-based recommended-next-action <text>):

Lab Server

Minimal MCP Streamable HTTP server for testing whether MCP clients correctly handle subscriptions/listen and notifications/resources/updated.

This repository is meant to be a reproducible issue / compatibility lab for CLI AI agents such as Codex CLI, Gemini CLI, OpenCode, GitHub Copilot CLI, Claude Code, Goose, and Crush.

Purpose

The server exposes one fixed MCP resource:

test://review/status

Initial content:

status: pending
version: 1
message: Waiting for simulated review result.

After a client opens a subscriptions/listen stream for the resource, the server waits for MCP_TEST_UPDATE_DELAY_SECONDS, changes the resource, and sends:

{
  "method": "notifications/resources/updated",
  "params": {
    "uri": "test://review/status"
  }
}

Updated content:

status: reviewed
version: 2
message: Simulated review result is now available.

Why Resource Subscriptions Instead Of Tools/Call

tools/call is useful for explicit actions, but many agent workflows depend on context that changes after the original request. Polling every source is noisy and client-specific. MCP resource subscriptions give clients a protocol-level way to learn that a known context object changed and should be re-read.

Examples where subscription behavior matters:

• Copilot review result

• PR review thread

• CI status

• Codecov comment

• GitHub issue discussion

• local build/test result

This test server focuses on whether the client notices a resource update, re-runs resources/read, and reflects the new content in the agent loop / model context.

This is a statement about the bundled reference server's design, not a restriction on the CLI: the call mode is a deliberate, separate escape hatch for callers that need a single tools/call invocation against any MCP server (e.g. triggering a one-off action tool) without standing up a subscription.

Start

docker compose up --build

MCP URL:

http://127.0.0.1:8089/mcp

For local development:

pnpm install --frozen-lockfile
pnpm run dev

Configuration

If MCP_TEST_SEND_LIST_CHANGED=true, the server also sends notifications/resources/list_changed after the simulated update.

Expected Client Behavior

An ideal MCP client should follow this flow:

server/discover                       (negotiate 2026-07-28)
  ↓
resources/list
  ↓
resources/read test://review/status
  ↓
subscriptions/listen { notifications: { resourceSubscriptions: ["test://review/status"] } }
  ↓
receive notifications/subscriptions/acknowledged   (MUST arrive first; check the honored filter)
  ↓
receive notifications/resources/updated            (on the same long-lived SSE stream)
  ↓
resources/read test://review/status again
  ↓
reflect updated status: reviewed in agent context
  ↓
close the stream                       (there is no resources/unsubscribe in 2026-07-28)

The server has no protocol-level session: the subscription lives exactly as long as the subscriptions/listen HTTP request. A stream that ends without a response is an abnormal disconnect, not a clean unsubscribe.

Server Capabilities

The initialize response advertises:

{
  "resources": {
    "subscribe": true,
    "listChanged": false
  }
}

listChanged follows MCP_TEST_SEND_LIST_CHANGED. subscribe: true now means "individual resource updates can be requested through resourceSubscriptions", not that resources/subscribe exists.

Implemented MCP Messages

• server/discover (2026-07-28 only — a 2025-era initialize is rejected with -32022)

• resources/list

• resources/read

• subscriptions/listen + notifications/subscriptions/acknowledged

• notifications/resources/updated

• notifications/resources/list_changed when MCP_TEST_SEND_LIST_CHANGED=true

• GET / DELETE on the MCP endpoint answer 405: the standalone GET SSE endpoint and Mcp-Session-Id sessions no longer exist

• tools/list, tools/call:

  • get_review_status — returns the current review status (same data as reading test://review/status)

  • echo_tool — testing utility for call mode; echoes { message } back as text content, or returns isError: true when called with { shouldError: true }

Logs

The server logs each important message so client behavior can be checked objectively:

[resources/list] requested
[resources/read] uri=test://review/status version=1
[resource/update] uri=test://review/status version=2
[notification/send] notifications/resources/updated uri=test://review/status
[resources/read] uri=test://review/status version=2

The key evidence for resource subscription support is:

subscriptions/listen was acknowledged
notification was sent on that stream
resources/read was received again after the notification

Tests

pnpm test

The test suite verifies:

• resources/list returns test://review/status

• initial resources/read returns version 1

• opening a subscriptions/listen stream triggers an internal update to version 2

• notifications/resources/updated is received

• updated resources/read returns version 2

• repeated probes against the same server process each observe the update

Standalone Subscription Probe Client

The repository also includes a reusable MCP SDK client that exercises the full subscription flow against a running server:

pnpm run probe:subscribe -- --url http://127.0.0.1:8089/mcp

After pnpm run build, the same client can be run directly with Node:

node dist/src/client/cli.js --url http://127.0.0.1:8089/mcp

This client is separate from any AI client's native MCP surface. For Codex CLI, it demonstrates a reproducible agent-driven SDK workaround: if the agent has shell, Node.js, local dependency, and localhost network access, it can run this client to open a subscriptions/listen stream, receive notifications/resources/updated, and re-read the updated resource.

Verification Procedure

Use docs/verification-guide.md for a repeatable client verification procedure.

Record results in results/compatibility-matrix.md.

Historical: the verification guides and the compatibility matrices under results/ were produced against the 2025-era protocol, before this repository moved to 2026-07-28. Their resources/subscribe / resources/unsubscribe steps no longer apply to the server described above — the flow is now the one in Expected Client Behavior. They are kept as a record of the compatibility spike, not as instructions to follow.

Skill Templates

Reusable Codex skill templates are tracked under docs/skills. The pr-review-subscribe template documents a PR review cycle that uses an MCP resource subscription as the primary wait route and polling only as fallback. It predates the 2026-07-28 migration and still describes the 2025-era resources/subscribe wire calls; the wait strategy carries over, the RPC names do not.

Client Compatibility

See results/compatibility-matrix-v2.md for the current Round 2 compatibility matrix (tool + resource testing) across Codex CLI, Gemini CLI, OpenCode, GitHub Copilot CLI, Claude Code, Goose, and Crush.