94 lines
5.1 KiB
Markdown
94 lines
5.1 KiB
Markdown
# **2026-05-13 MCP: Transport Working Group (Weekly)**
|
|
|
|
# **TL;DR**
|
|
|
|
The meeting finalized task notification semantics and introduced the Triggers
|
|
and Events Working Group proposal. The group reached consensus on explicitly
|
|
disallowing progress and logging notifications on tasks under the subscription
|
|
model, choosing to table these in favor of defining a robust intermediate
|
|
results pattern later. Peter Alexander presented the three delivery modes of the
|
|
triggers and events primitive: polling, push stream, and web hooks. Tensions
|
|
arose concerning web hook security, specifically DDoS amplification attacks, and
|
|
whether cursor compatibility should be mandated between polling and web hook
|
|
delivery to ensure reliable pub/sub delivery.
|
|
|
|
# **Detailed Summary**
|
|
|
|
## **Task Notification Strategy**
|
|
|
|
* **Reworking Notifications:** The team discussed how progress and logging
|
|
notifications fit into tasks. Under the stateless SEP, normal requests use an
|
|
SSE stream for notifications, which is incompatible with task polling.
|
|
* **Tension Point \- Fragmented Notifications:** The group debated three
|
|
options: subscribing to all task notifications in `subscriptionslist`
|
|
(disjointed from base spec), opening an SSE stream on `tasks.get` (fragmented
|
|
and complex for polling), or disallowing progress/logging on tasks
|
|
temporarily.
|
|
* **Consensus \- Temporary Disallowance:** The group agreed to Solution 3, which
|
|
explicitly disallows SSE streams on tasks, supporting only the
|
|
`subscriptionslist` for task status updates.
|
|
* **Tension Point \- Progress vs Intermediary Results:** Shaun Smith advocated
|
|
for progress notifications as a major UX improvement. Caitie McCaffrey and
|
|
Peter Alexander argued that progress is too simple and what users actually
|
|
want is intermediate results with steering input (e.g., telling a task to
|
|
change a generated color).
|
|
|
|
## **Triggers and Events Primitive**
|
|
|
|
* **Events Primitive Introduction:** Peter Alexander introduced a new events
|
|
primitive (`events.list`) for low-rate server-side events (e.g., Discord
|
|
messages, build alerts), excluding high-throughput or multimedia streams.
|
|
* **Three Delivery Modes:** The proposal supports polling (clients pull at
|
|
intervals), push streams (single SSE connection per subscription), and web
|
|
hooks (soft TTL-based subscription where the server posts to the client).
|
|
* **Tension Point \- Cursors and Reliability:** Caitie McCaffrey noted that
|
|
implementing cursors for replay makes this more like pub/sub than best-effort
|
|
notifications. Cursors are generated by the server and passed by the client to
|
|
resume streams.
|
|
|
|
## **Web Hook Architecture & Security**
|
|
|
|
* **Two-Hop Architecture:** Shaun Smith noted that web hooks require an open
|
|
HTTP point on both the MCP server and the host client, creating an extra hop.
|
|
Peter Alexander defended this, stating it provides a constrained, filtered API
|
|
for hosts.
|
|
* **Tension Point \- DDoS and Handshake Security:** Kurtis Van Gent raised
|
|
concerns that malicious actors could register fake web hook endpoints to
|
|
orchestrate amplification DDoS attacks. The team agreed on adding an endpoint
|
|
verification challenge-response step and client-generated HMAC signing
|
|
secrets.
|
|
* **Tension Point \- Web Hook Cursors:** Kurtis Van Gent questioned the
|
|
necessity of cursors for web hooks, arguing that standard web hooks start from
|
|
the subscription moment. Caitie McCaffrey and Clare Liguori noted that
|
|
combining web hooks as a "wake up" signal with background polling is a common,
|
|
reliable pattern.
|
|
* **Tension Point \- Mandating Cursor Compatibility:** Kurtis Van Gent expressed
|
|
concerns about not mandating cursor compatibility if both polling and web
|
|
hooks are supported, as arbitrary clients and servers must align.
|
|
|
|
# **Next Steps & Timeline**
|
|
|
|
Core maintainers will move forward with task notification specifications without
|
|
progress/logging. Peter Alexander will update the triggers and events proposal
|
|
with details on transport rationale and handshake verification, while Caitie
|
|
McCaffrey will follow up on official task acceptance votes.
|
|
|
|
# **Action Items**
|
|
|
|
- [ ] Caitie McCaffrey Follow up on the official acceptance status of tasks and
|
|
check the necessity of a voting process.
|
|
- [ ] Caitie McCaffrey Investigate real-world progress notification use cases
|
|
and compare them against Shaun Smith's examples.
|
|
- [ ] Peter Alexander Update the events proposal document to clarify the
|
|
coexistence rationale for pull, push, and web hook modes.
|
|
- [ ] Peter Alexander Add endpoint verification steps and security measures to
|
|
the web hook specification to prevent DDoS attacks.
|
|
- [ ] Shaun Smith Analyze the complexities of many-to-many relationships between
|
|
servers, hosts, and multiple web hooks.
|
|
- [ ] Clare Liguori Develop and test client-side architectures showing protocol
|
|
implementation based on these definitions.
|
|
- [ ] All Participants Conduct larger system demos with external event sources
|
|
(Discord, Telegram) to evaluate production readiness.
|
|
|
|
Source: [Meeting Notes — May 13,
|
|
2026](https://docs.google.com/document/d/1KFwTal-yFzfszl6VOsCjNLu7L6joNLbipBversMrM6Q/edit?usp=meet_tnfm_email)
|