ref: Remove agent mode (#1192)

Remove the legacy agent mode across the MCP core, CLI, Cloudflare app,
telemetry, tests, and documentation. The server now exposes the standard
tool surface without the `--agent` / `?agent=1` mode switch or the
`use_sentry` meta-tool.

**Embedded agent routing remains supported**

The unrelated `--agent-provider` option is preserved for selecting the
LLM provider used by embedded-agent functionality.

**Obsolete test infrastructure removed**

This removes the agent CLI test package and benchmark script alongside
agent-mode-specific tests and docs.

<!-- junior-request-attribution:start -->
Requested by **David Cramer** via Junior.
<!-- junior-request-attribution:end -->

<!-- junior-session-footer:start -->
<!-- junior-conversation-id:slack%3AC08J1NSPU6S%3A1785177165.113189 -->

--

[View Junior
Session](https://junior-prod.sentry.dev/conversations/slack%3AC08J1NSPU6S%3A1785177165.113189)

<!-- junior-session-footer:end -->

---------

Co-authored-by: sentry-junior[bot] <264270552+sentry-junior[bot]@users.noreply.github.com>

Author: sentry-junior[bot]

.agents/skills/mcp-qa/SKILL.md

@@ -77,12 +77,6 @@ that is not part of the tool response. Review raw tool output against
 and free of raw API JSON, internal implementation IDs, empty placeholders, and
 unrelated instructions.
 
-If your changes involve agent mode or experimental tools:
-
-```bash
-pnpm -w run cli --transport stdio --agent "show me my recent errors"
-pnpm -w run cli --transport stdio --experimental "your query"
-```
 
 ## 3. Real Agent Clients
 
@@ -91,9 +85,6 @@ in a real agent client:
 
 ```bash
 pnpm -w run build
-pnpm -w run agent-cli-test auth login
-pnpm -w run agent-cli-test --provider claude --setup stdio
-pnpm -w run agent-cli-test --provider codex --setup stdio
 ```
 
 What this verifies:
@@ -105,7 +96,6 @@ For agent-client-specific behavior, replace the default harness prompt with the
 same realistic prod prompt used for stdio QA. Passing QA requires the same
 changed-tool transcript evidence, not only `whoami`.
 
-The `stdio` setup uses `packages/agent-cli-test/projects/stdio/.sentry/mcp.json`
 as an isolated auth cache. Real clients do not give stdio subprocesses a TTY,
 so warm the cache before running the harness. It also runs the built
 `packages/mcp-server/dist/index.js`, so build first to avoid stale code.
@@ -139,8 +129,6 @@ pnpm -w run cli --mcp-host=http://localhost:5173/mcp/<org> \
   "<same realistic prod prompt used for stdio QA>"
 pnpm -w run cli --mcp-host=http://localhost:5173/mcp/<org>/<project> \
   "<same realistic prod prompt used for stdio QA>"
-pnpm -w run agent-cli-test --provider claude --setup repo
-pnpm -w run agent-cli-test --provider codex --setup repo
 ```
 
 Look for `Connected to MCP server (<resolved MCP URL>)` to confirm HTTP

.github/workflows/mcp-server-package.yml

@@ -102,7 +102,6 @@ jobs:
           declare const server: Parameters<typeof startStdio>[0];
 
           const telemetryContext: StdioServerContext = {
-            agentMode: false,
             experimentalMode: false,
             mcpUrl: "https://example.com/mcp",
             sentryHost: "sentry.io",

.github/workflows/token-cost.yml

@@ -89,7 +89,7 @@ jobs:
 
           ---
 
-          **Note:** This measures the static overhead of tool definitions sent to LLM clients with every request. Lower is better. The \`use_sentry\` tool is excluded as it's only available in agent mode.
+          **Note:** This measures the static overhead of tool definitions sent to LLM clients with every request. Lower is better.
           REPORT_EOF
 
           # Display report

AGENTS.md

@@ -95,7 +95,6 @@ pnpm run build                            # Build all packages
 # Testing
 pnpm -w run cli --transport stdio "q"      # Test MCP tools
 pnpm -w run cli --transport stdio --access-token=TOKEN "q"
-pnpm -w run cli --transport stdio --agent "query"
 
 # Quality (run before committing)
 pnpm run tsc && pnpm run lint && pnpm run test

README.md

@@ -245,9 +245,6 @@ pnpm eval
 # Test with local dev server (default: http://localhost:5173)
 pnpm -w run cli "who am I?"
 
-# Test agent mode (use_sentry tool only)
-pnpm -w run cli --agent "who am I?"
-
 # Test against production
 pnpm -w run cli --mcp-host=https://mcp.sentry.dev "query"
 

TELEMETRY.md

@@ -4,7 +4,6 @@
 
 Use this when investigating Sentry MCP production incidents across the
 Cloudflare HTTP server, stdio package, MCP tools, OAuth flows, and test-client
-agent mode.
 
 Primary backend: Sentry Logs, Issues, Spans/Traces, and Metrics in the MCP
 server projects. Start with a Sentry event, trace ID, route, user, client
@@ -38,7 +37,6 @@ the pivots and recipes below.
 | `app.rate_limit.scope` | local rate-limit scope | metrics | IP vs user rate limits |
 | `app.route.group` | coarse route family | metrics | `mcp`, `oauth`, `chat`, `search` |
 | `app.transport` | MCP transport | tags, spans | `http`, `sse`, or `stdio` |
-| `app.server.mode.agent` | agent-mode flag | metrics, spans, tags | `?agent=1` or stdio `--agent` adoption |
 | `app.server.mode.experimental` | experimental-mode flag | metrics, spans, tags | `?experimental=1` or stdio `--experimental` adoption |
 | `mcp.session.id` | MCP session identity | spans | session timeline |
 | `gen_ai.tool.name` | MCP tool being called | spans, issues | tool timeline |
@@ -86,16 +84,13 @@ HTTP response rates by route and status.
 
 ```text
 dataset=tracemetrics query='metric:app.server.response http.route:"<route>"'
-fields=timestamp,metric,http.request.method,http.route,http.response.status_code,app.response.status_class,app.route.group,app.client.family,app.server.mode.agent,app.server.mode.experimental,value
 aggregate=sum(value) by http.route,http.response.status_code
 ```
 
 MCP mode adoption by client family.
 
 ```text
 dataset=tracemetrics query='metric:app.server.response http.route:"/mcp/:organizationSlug?/:projectSlug?"'
-fields=timestamp,metric,app.client.family,app.server.mode.agent,app.server.mode.experimental,value
-aggregate=sum(value) by app.client.family,app.server.mode.agent,app.server.mode.experimental
 ```
 
 Local rate-limit volume and scope.
@@ -150,7 +145,6 @@ Tool execution timeline for a slow or failing tool.
 
 ```text
 dataset=spans query='gen_ai.tool.name:"<tool_name>" app.consent.skill.<skill>.granted:true'
-fields=timestamp,trace,span_id,span.op,span.duration,gen_ai.tool.name,app.transport,app.client.family,app.server.mode.agent,app.server.mode.experimental,app.constraint.organization_slug,app.constraint.project_slug,gen_ai.tool.call.arguments.organizationSlug,gen_ai.tool.call.arguments.projectSlugOrId,error.type
 sort=-timestamp
 ```
 
@@ -177,7 +171,6 @@ Stdio sessions by configured host or mode.
 
 ```text
 dataset=spans query='app.transport:stdio app.upstream.host:"<host>"'
-fields=timestamp,trace,span_id,app.server.version,app.server.mode.agent,app.server.mode.experimental,app.url.full,error.type
 sort=-timestamp
 ```
 
@@ -194,7 +187,6 @@ Attributes: `http.request.method`, `http.route`,
 `http.response.status_code`, `app.response.status_class`,
 `app.route.group`, `app.response.reason`, `app.rate_limit.scope`,
 `app.request.duration_ms`. MCP responses also include
-`app.client.family`, `app.server.mode.agent`, and
 `app.server.mode.experimental`.
 
 ### OAuth And Client Registration
@@ -238,10 +230,8 @@ Attributes: `app.resource.type`, `gen_ai.tool.name`, `trace_id`, `span_id`
 
 ### Stdio Transport
 
-Local package startup, host selection, agent mode, experimental mode, or token
 resolution behaves differently than the hosted server.
 
-Attributes: `app.server.version`, `app.transport`, `app.server.mode.agent`,
 `app.server.mode.experimental`, `app.upstream.host`, `app.url.full`
 
 ### Agent And GenAI

benchmark-agent.sh

@@ -221,7 +221,7 @@ Ordinary operation modules live as flat files under
 need more implementation structure.
 `packages/mcp-core/src/tools/catalog-runtime/` contains the shared filtering,
 search, schema, and execution helpers. Wrapper/gateway tools such as
-`search_sentry_tools`, `execute_sentry_tool`, and `use_sentry` live in
+`search_sentry_tools`, `execute_sentry_tool` live in
 `packages/mcp-core/src/tools/special/`.
 
 `packages/mcp-core/src/tools/surfaces.ts` only centralizes the subsets of
@@ -332,66 +332,3 @@ This pattern is used when:
 - Embedded agent errors are returned as UserInputError
 - Calling agent sees the error and can retry
 - No internal retry loops - single responsibility
-
-### use_sentry Tool Architecture
-
-The `use_sentry` tool provides a natural language interface to all Sentry MCP tools using an in-memory MCP client-server architecture:
-
-**Architecture**:
-1. Creates linked pair of `InMemoryTransport` from MCP SDK
-2. Builds internal MCP server with all 18 tools (excludes use_sentry to prevent recursion)
-3. Connects server to serverTransport within ServerContext
-4. Creates MCP client with clientTransport
-5. Embedded GPT-5 agent accesses tools through MCP protocol
-6. Zero network overhead - all communication is in-memory
-
-**Data Flow**:
-```
-User request → use_sentry handler
-  ↓
-Creates InMemoryTransport pair
-  ↓
-Builds internal MCP server (18 tools)
-  ↓
-Creates MCP client
-  ↓
-Embedded agent calls tools via MCP protocol
-  ↓
-MCP server executes tool handlers
-  ↓
-Results returned through MCP protocol
-  ↓
-Agent processes and returns final result
-```
-
-**Benefits**:
-- Full MCP protocol compliance throughout
-- Architectural consistency - all tool access via MCP
-- Zero performance overhead (no network, no serialization)
-- Proper tool isolation at protocol level
-- No recursion risk (use_sentry excluded from internal server)
-
-**Implementation**: Uses built-in `InMemoryTransport.createLinkedPair()` from `@modelcontextprotocol/sdk/inMemory.js` for reliable in-process communication.
-
-## Security Model
-
-- Access tokens never logged
-- OAuth tokens encrypted in KV
-- Per-organization isolation
-- CORS configured for security
-
-## Testing Architecture
-
-Three levels of testing:
-
-1. **Unit tests**: Fast, isolated, snapshot-based
-2. **Integration tests**: With mocked API
-3. **Evaluation tests**: Real-world scenarios with LLM
-
-## References
-
-- MCP SDK: `@modelcontextprotocol/sdk`
-- Build config: `turbo.json`
-- TypeScript config: `packages/mcp-server-tsconfig/`
-- API client: `packages/mcp-core/src/api-client/`
-- stdio package: `packages/mcp-server/`

docs/architecture/overview.md

@@ -8,7 +8,6 @@ Not every tool is exposed to every consumer. We rely on several mechanisms to ke
 
 - **Catalog by default** — Most tools are searchable/executable through `search_sentry_tools` + `execute_sentry_tool` automatically. Search uses the tool's existing name and description.
 - **Catalog registry** — `packages/mcp-core/src/tools/catalog/index.ts` lists ordinary Sentry operation tools. The catalog directory is intentionally flat: one tool entry per file.
-- **Special tools** — Wrapper/gateway tools (`search_sentry_tools`, `execute_sentry_tool`, `use_sentry`) live in `packages/mcp-core/src/tools/special/`. They still use the same tool types, but they are kept out of the ordinary catalog.
 - **Central direct exposure policy** — `packages/mcp-core/src/tools/surfaces.ts` lists the catalog tools that are also exposed directly through MCP `tools/list`. The direct surface intentionally keeps long-tail tools catalog-only because `search_sentry_tools` + `execute_sentry_tool` are available as primary primitives.
 - **`requiredCapabilities`** — Tools declare which project capabilities they need (e.g. `profiles`, `replays`, `traces`). If the upstream project doesn't have a capability enabled, the tool is automatically hidden.
 - **`experimental` / `hideInExperimentalMode`** — Feature flags for tools that are being tested or replaced.

docs/contributing/adding-tools.md

@@ -8,7 +8,6 @@ Sentry MCP uses embedded AI agents for the following tools:
 - `search_events` - Natural language search across events, metrics, and session replays
 - `search_issues` - Natural language search across issues
 - `search_issue_events` - Search events within a specific issue
-- `use_sentry` - Unified natural language interface to all Sentry operations
 
 These tools require an LLM provider (OpenAI, Azure OpenAI, Anthropic, or OpenRouter) to be configured.