Repository navigation
MCP analytics for Rust MCP servers - #308
Merged
Merged
Conversation
Contributor
Author
some screenshots:1) opencode and deepseek calling 2 rmcp servers
2) failed call sent with
|
wthrajat
force-pushed
the
feat/mcp-analytics
branch
from
October 26, 2026 13:14
63f2ae5 to
e8c8929
Compare
Commit signatures resolvedGitHub now verifies every commit signature in this PR. No further action is needed. |
lucasheriques
approved these changes
Oct 26, 2026
Contributor
There was a problem hiding this comment.
Reviewed against the cross-SDK MCP analytics spec and ran the example server end to end over stdio; the events match the spec. I pushed a few fixes on top: panics in tool handlers are now recorded as failed calls (then re-raised), $mcp_interface is never dropped by the size bound, payload redaction cost is bounded per event, a panic isn't double-reported when capture_panics is on, and get_more_tools is listed on the first tools/list page like TS and Python. Each commit says how it was tested. Thanks @wthrajat, this is a really solid port.
Contributor
Author
|
thanks @lucasheriques!! :D (looks like your commits are missing signatures) |
rmcp does not catch handler panics, so a panicking tool sent no $mcp_tool_call at all: the worst failures were invisible. The spec counts a panic as a final outcome. The wrapper now catches the unwind, sends the failed call ($mcp_error_type "Panic", the panic message) and its $exception, then resumes the panic so the server behaves exactly as it would uninstrumented. Tested: new a_panicking_handler_is_a_failed_call fails before, passes after; cargo test --features mcp --test test_mcp_connection (21 pass); cargo clippy --features mcp -D warnings; cargo +1.88.0 check --locked --lib --features mcp.
$mcp_interface is required on every MCP event, but it was not in the always-kept list, so an event over the byte bound (a listing of a few thousand tools) lost it along with the other optional properties. Tested: an_oversized_event_keeps_its_interface fails without the fix, passes with it.
Each payload string was redacted up to its own 32,768-character limit, so a result of 100 large text blocks scanned ~3.3M characters on the response path, only for the event bound to cut them to 256 afterwards. Strings in a payload now share a budget of one event's size; once it is spent, the rest are cut to 64 characters before redaction scans them. Tested: strings_past_the_event_budget_are_cut_short fails without the budget, passes with it. Release build, 99 x 40 KB blocks: ~120 ms -> ~10 ms per test run. cargo test --features mcp,tower: 484 pass.
From the pre-PR review: - Catch a panic raised while the handler builds its future, not only while it runs, by awaiting it inside an async block. - Report a panic's $exception as unhandled: rmcp never turns it into a response, unlike a returned error. - Debit the payload budget by what redaction scans, not what it keeps, so strings redacted to a short token cannot keep every scan at full cost. - Tests pin the exact cut (first string at its limit, strings past the budget at 64) and that the client gets no answer from a panic. Tested: cargo test --features mcp (481 pass); cargo clippy --features mcp -D warnings.
With the client's opt-in capture_panics hook installed, a panicking tool was reported twice: once by the hook, once by MCP analytics. The hook already sent the $exception, so the call now sends only its $mcp_tool_call. The panic test also asserts the client gets no answer within 250 ms, which pins that the panic is resumed. Tested: cargo test --features mcp,tower (484 pass); cargo check --no-default-features --features mcp. The hook branch has no test of its own: the hook is process-global and installable only through the global client.
The virtual tool was appended to the last page, so a client that reads only the first page of a paginated server never saw it. TS (#4967) and Python (#962) list it on the first page, the one every client reads; Rust now matches. Tested: the_virtual_tool_is_listed_once_on_the_first_page fails before ((0, 1)), passes after; cargo test --features mcp,tower (485 pass); cargo clippy --features mcp -D warnings.
lucasheriques
force-pushed
the
feat/mcp-analytics
branch
from
October 26, 2026 15:41
451004d to
2db1c93
Compare
lucasheriques
added a commit
to PostHog/posthog.com
that referenced
this pull request
Oct 28, 2026
* docs(mcp-analytics): add the Rust SDK posthog-rs 0.29.0 (on crates.io) ships experimental MCP analytics for rmcp servers behind the `mcp` feature (PostHog/posthog-rs#308). Add a Rust installation page (install, instrument, McpOptionsBuilder options, identify, sessions, get_more_tools, the manual API, events with $lib posthog-rs-mcp, limits), a Rust column in the SDK feature table, the sidebar entry, and Rust tabs/mentions on start-here, events, and intent. Tested: every Rust snippet compiled with `cargo check` against posthog-rs =0.29.0 + rmcp 3.7; changed MDX parses with @mdx-js/mdx v1; prettier --check passes on src/navs/index.js. * docs(mcp-analytics): correct Rust SDK claims from the pre-PR review Two-model review (Opus, Sonnet) of the Rust page against posthog-rs 0.29.0 found overclaims and gaps. Fix: manual API sends tool-call events only; error type naming and the rmcp `validation` case; client_metadata intent source; get_more_tools collision check and placement; required llm_model, appended conversation handle, composed-schema skip; argument stripping needs Registered or list_tools; rmcp 3.7+ and Clone; panic $exception via hook; Limits (tasks, stateless client info, no properties callback, no stack trace). Mark Rust in the Go-only notes on events, identifying-users, privacy, intent, start-here, and sdk-v2. Tested: all changed MDX parses with @mdx-js/mdx v1; each claim checked against sdks/rust at tag 0.29.0. * docs(mcp-analytics): answer failed Rust manual calls, narrow client identity limit The custom dispatcher example recorded a failed call but never answered it, so a copied dispatcher left clients waiting for a timeout. The limits listed client name and version as missing on every stateless server, but posthog-rs 0.29.0 reads them from a 2026-07-28 request's _meta (RequestScope::client_identity), so only older clients without a replayed token lack them. * docs(mcp-analytics): correct the Rust panic, initialize session, and identify notes Addresses wthrajat's review: the panic hook needs capture_panics on the global client, not just the error-tracking feature; $mcp_initialize gets its own ses_ id because initialize precedes Mcp-Session-Id; Rust resolves identify once per request. Also drops a duplicated Go sentence on start-here. * docs(mcp-analytics): scope the BeforeSend note to Go * docs(mcp-analytics): make the Rust snippets compile, check events by query, add troubleshooting - Every Rust block on the Rust and start-here pages now compiles against posthog-rs 0.29.0 as written: imports added, the identify user type and the custom-dispatcher types spelled out, a compiled Streamable HTTP snippet. - The example run command reuses POSTHOG_PROJECT_TOKEN, the one variable the page asks for. - "Check your first events" is a SQL query, so an agent can run it through the PostHog MCP server's execute-sql tool. - New troubleshooting table keyed by symptom. Tested: a scratch crate compiles each ```rust block extracted from the .mdx files (9/9 compile); removing one import from a copy makes that block fail. * Correct Rust $exception properties after the master merge 0.29.0 sends no $mcp_server_build, and the Rust note contradicted the paragraph above it. State the Rust differences once, in that paragraph. Found by an independent verification pass against the 0.29.0 crate source.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
💡 Motivation and Context
closes #258
This adds mcp analytics like the ruby, python, go and node sdks have.
For more context, I had a bit of discussion with @pauldambra on twitter here about this.
Also screenshots are attached here MCP analytics for Rust MCP servers #308 (comment)
mcpfeatureposthog-rs-mcpas the lib so they're easy to tell apart1.88, the rest of the crate still works on1.78💚 How did you test it?
stdioandhttp)📝 Checklist
If releasing new changes
sampo addto generate a changeset file