We recently added outputSchema support for our MCP tools (not yet for
worker routes yet). Today, we always return structuredContent. On tool
execution errors we return structuredContent: {"error": "..."} with
isError: True, even when that shape does not match the tool’s declared
outputSchema. Since the MCP spec says clients SHOULD validate
structuredContent against outputSchema, some clients reject these
responses.
Since structuredContent is optional, we’re going to omit it when
isError: true.
<!-- CURSOR_SUMMARY -->
---
> [!NOTE]
> **Medium Risk**
> Changes the shape of tool error responses across the MCP server, which
may break clients or tools that previously relied on
`structuredContent["error"]` for failures. Behavior is more
spec-compliant but touches core request/response paths and test
expectations.
>
> **Overview**
> Prevents MCP tool error responses from violating a tool’s declared
`outputSchema` by **always setting `structuredContent=None` when
`isError=True`** (server execution errors, unknown tools, middleware
exceptions, and `Context.tools.call_raw` JSON-RPC errors).
>
> Updates requirement-failure error formatting to put the human-friendly
message in `content[0]` and (when present) serialize extra
machine-readable fields (e.g. `authorization_url`, `llm_instructions`)
into an additional `content` item. Examples and integration/unit tests
are updated to read errors from `content[0].text`, and
`arcade-mcp-server` is bumped to `1.19.2`.
>
> <sup>Reviewed by [Cursor Bugbot](https://cursor.com/bugbot) for commit
4213bdd4aa44362de85c30f5f31c576243c132d5. Bugbot is set up for automated
code reviews on this repo. Configure
[here](https://www.cursor.com/dashboard/bugbot).</sup>
<!-- /CURSOR_SUMMARY -->
**Implements**: [SEP-2448: server execution telemetry]
(https://github.com/modelcontextprotocol/modelcontextprotocol/pull/2448)
**Description:**
**The Observability Gap (The Problem)**
MCP clients propagate trace context to servers, but server-side
execution remains a black box. The client sees a single tools/call or
resources/read span; everything the server does (auth checks, policy
evaluation, API calls, sub-tool invocations) is invisible. In
cross-organization deployments, clients and servers use separate
observability backends with no shared collector access, making
traditional span export useless.
<img width="1015" height="450" alt="Screenshot 2026-03-23 at 3 43 21 PM"
src="https://github.com/user-attachments/assets/58c817b5-fee6-46a3-9877-d523a25368ad"
/>
**Server Execution Telemetry (The Solution)**
Servers advertise serverExecutionTelemetry and return a curated slice of
their execution spans directly in _meta.otel of the response. Clients
ingest these verbatim OTLP spans into their own collector, stitching
server-side execution into their distributed trace; no shared
infrastructure required. The black box becomes transparent.
<img width="945" height="574" alt="Screenshot 2026-03-23 at 3 43 44 PM"
src="https://github.com/user-attachments/assets/38d97c94-aa73-4e62-9b4e-3264600e5ed0"
/>
.
**Summary:**
Implement MCP serverExecutionTelemetry capability that enables
cross-organization distributed tracing by returning server-side
OpenTelemetry spans to clients inline via _meta.otel.traces.
Server-side (middleware):
- TelemetryPassbackMiddleware intercepts tools/call and resources/read
- ContextVarSpanCollector isolates span collection per-request via
ContextVar
- Propagates traceparent from client request for distributed trace
stitching
- Serializes collected spans to verbatim OTLP JSON (resourceSpans
format), directly POSTable to /v1/traces
- Top-level span filtering by default; full span tree via detailed
opt-in
- Middleware advertises capabilities via get_capabilities() on the
Middleware base class
- Provisional API: FutureWarning emitted until SEP-2448 is ratified
Client-side (reference agent):
- LangChain ReAct agent connects to MCP server via
streamable_http_client with OAuth 2.1
- Detects serverExecutionTelemetry capability at initialization
- Dynamically wraps discovered MCP tools with traceparent propagation
and _meta.otel span request
- Ingests returned server spans into Jaeger (OTLP JSON) and Galileo
(OTLP protobuf)
- Two-act demo: --no-passback (black box) vs default (full server-side
visibility)
Dependencies:
- opentelemetry-api and opentelemetry-sdk added to arcade-mcp-server
Bump arcade-mcp-server version to 1.18.0.
Versions:
* arcade-mcp\==1.0.0rc1
* arcade-mcp-server\==1.0.0rc1
* arcade-core\==2.5.0rc1
* arcade-tdk\==2.6.0rc1
* arcade-serve\==2.2.0rc1
### Summary
Adds first-class MCP support across Arcade, introduces a new MCP server
and CLI, unifies the project under the arcade-mcp name, overhauls
templates/scaffolding, and improves developer tooling, secrets
management, and examples.
### Highlights
- **MCP Server & Core**
- New MCP server with stdio and HTTP/SSE transports, session management,
resumability, and lifecycle handling.
- FastAPI-like `MCPApp` for building servers with lazy init; integrated
worker+MCP HTTP app option.
- Middleware system (logging and error handling), robust exception
hierarchy, and Pydantic-based settings.
- Async-safe managers for tools, resources, and prompts backed by
registries and locks.
- Developer-facing, transport-agnostic runtime context interfaces (logs,
tools, prompts, resources, sampling, UI, notifications).
- Conversion from Arcade ToolDefinition to MCP tool schema; OpenAI JSON
tool schema converter.
- Parser supports `@app.tool`/`@app.tool(...)` decorators.
- **CLI**
- New `mcp` command to run MCP servers with stdio or HTTP/SSE.
- New `secret` command to set/list/unset tool secrets (supports .env
input, preserves original casing for lookups).
- `new` command refactored; option to create a full toolkit package with
scaffolding.
- `chat` command removed.
- `serve.py` imports updated to `arcade_serve.fastapi.telemetry`;
version retrieval now uses `arcade-mcp`.
- `show.py` refactor to use new local catalog utilities.
- `display_tool_details` improved: adds “Default” column and handles
nested properties.
- **Configuration & Discovery**
- New `configure.py` to set up Claude Desktop, Cursor, and VS Code to
connect to local or Arcade Cloud MCP servers.
- Discovery utilities to find/install toolkits, build `ToolCatalog`s,
analyze files for tools, load kits from directories (pyproject parsing),
and build minimal toolkits.
- Better handling of provider API key resolution and evaluation suite
loading.
- **Templates & Scaffolding**
- Reorganized template structure (minimal vs full); moved
`.pre-commit-config.yaml`, `.ruff.toml`, license, Makefile, README,
tests, and tools layout to correct paths.
- Minimal template adds `.env.example` for runtime secret injection.
- Template pyproject updated for MCP servers; includes sample server
with greeting and secret-reveal tools.
- Authorization flow in templates simplified.
- **Repo-wide Renaming & Examples**
- Migrates references from `arcade-ai` to `arcade-mcp` across READMEs,
scripts, and package metadata.
- Examples updated (LangChain/LangGraph/AI SDK/TypeScript) and package
name changed to `arcade-mcp-sdk`.
- **Evals & Core Utilities**
- Evals now use OpenAI tooling format (`OpenAIToolList`, `to_openai`);
`tool_eval` takes `provider_api_key`.
- Core utilities: fixed `does_function_return_value` by dedenting before
parse; version bump to `2.5.0rc1` and dependency cleanup.
- **Tooling & CI**
- `setup-uv-env` action splits toolkit vs contrib dependency
installation.
- Pre-commit: excludes `libs/arcade-mcp-server/mkdocs.yml` and
`libs/tests/` from YAML and Ruff hooks; Ruff per-file ignores (e.g.,
C901 in `libs/**/*.py`, TRY400 in server docs paths).
- Makefile updates for uv env setup, quality checks, tests, builds, and
new `shell` target.
- Added Makefile to MCP server library to streamline dev workflow.
- **Cleanup**
- Removed `claude.json` config.
- Simplified stdio entrypoint; removed unused imports (`arcade_gmail`,
`arcade_search`).
### Breaking Changes
- **CLI**: `chat` command removed; use `mcp`, `secret`, and updated
`new`.
- **Naming**: All users should update references from `arcade-ai` to
`arcade-mcp`.
- **Templates**: File paths moved; downstream scripts referencing old
template locations may need updates.
### Getting Started
- Run an MCP server:
- `arcade mcp --stdio --toolkits your_toolkit`
- `arcade mcp --http --toolkits your_toolkit`
- Manage secrets:
- `arcade secret set your_toolkit KEY=value`
- `arcade secret list your_toolkit`
- `arcade secret unset your_toolkit KEY`
- Configure clients:
- `arcade configure` to set up Claude Desktop, Cursor, and VS Code for
local/Arcade Cloud MCP.
---------
Co-authored-by: Sam Partee <sam@arcade-ai.com>
Co-authored-by: Shub <125150494+shubcodes@users.noreply.github.com>
Eric Gustin