* Adopt streamable_http_client API from MCP SDK
- Update import to use new streamable_http_client function
- Convert httpx_client_factory to httpx.AsyncClient before passing to new API
- Maintain backward compatibility by continuing to accept factories
- Add deprecation warning for sse_read_timeout parameter
The new API accepts httpx.AsyncClient directly instead of factories.
We continue accepting factories for OAuth compatibility, converting
them to clients at the boundary with the MCP SDK.
* Fix timeout type conversion for streamable_http_client
Convert read_timeout_seconds from timedelta to float before passing
to httpx, matching the pattern used in the SSE transport.
* Enable redirect following in httpx client
* Fix httpx client resource leak
* Fix tool_choice to always require tools when result_type is set
* Consolidate sampling examples with rich output
* Replace eval() with explicit add/multiply tools
* Add AnthropicSamplingHandler
Adds a sampling handler for the Anthropic API at
fastmcp.client.sampling.handlers.anthropic, alongside the existing
OpenAI handler. Includes full support for tool calling.
Install with: pip install fastmcp[anthropic]
* Update default model
* Update sampling docs to cover both OpenAI and Anthropic handlers
* Use AsyncAnthropic, fix falsy value handling, handle tool_choice none
* Propagate isError to Anthropic, join multiple text blocks, fix docs
* Unify SamplingHandler and promote OpenAI handler
Consolidates ServerSamplingHandler and ClientSamplingHandler into a single
SamplingHandler type alias. Moves OpenAISamplingHandler from experimental
to fastmcp.client.sampling.handlers.openai as the canonical location.
Backwards compatibility maintained for imports from experimental.
* Remove unreachable code paths in OpenAI handler
* Fix docstring and use elif for mutually exclusive branches
* MCP → SDK (vocab change only)
* WIP: Sampling API with SamplingResult[T] and result_type
* SEP-1577: Sampling with tools
- Add tools and result_type parameters to ctx.sample()
- Update OpenAI handler for tool content types
- Client advertises sampling.tools capability by default
- Collect tool results into single message with list content
* Fix tool result content handling in OpenAI handler
* Remove @sampling_tool decorator - pass functions directly to sample()
Functions passed to ctx.sample(tools=[...]) are now auto-converted
via SamplingTool.from_function(). Users can still use that method
directly for custom name/description overrides.
* Remove auto-conversion of MCP tools to sampling tools
Users want MCP tools passed to ctx.sample() to go through the full MCP
machinery (middleware, native responses) rather than being auto-converted
to direct function calls. Now only SamplingTool and plain callables are
accepted - passing a FastMCP Tool raises a clear TypeError.
Also bumps mcp dependency to >=1.24.0 for required sampling features.
* Refactor sampling API: replace sample_iter() with sample_step()
Replace the mutable SampleRun/sample_iter() pattern with a simpler stateless
sample_step() function. sample_step() makes a single LLM call and returns a
SampleStep with the response and history. sample() now loops sample_step()
internally.
Key changes:
- Add sample_step() for fine-grained control over the sampling loop
- Remove SampleRun class and sample_iter() method
- Structured output uses tool description only (no prompt modification)
- execute_tools parameter controls automatic vs manual tool execution
* Address CodeRabbit nitpicks
* Address CodeRabbit review feedback for sampling tools
- Fix temperature=0.0 being dropped due to falsy evaluation
- Add ToolChoice.name support for forcing specific tools
- Replace assert statements with explicit RuntimeError checks
- Add mask_error_details parameter to sample()/sample_step() with ToolError escape hatch
- Fix hasattr patterns with proper isinstance checks
- Document mask_error_details and add OpenAI prerequisites to docs
* Address additional CodeRabbit review feedback
- Catch ValidationError specifically instead of bare Exception
- Update result_type docs to mention dataclasses and basic types
- Raise ValueError for unknown tool_choice modes
- Validate sampling_handler_behavior to catch typos
- Remove ToolChoice.name handling (not part of MCP spec)
- Validate tool_choice string in sample_step()
* Review fixes for sampling tools PR
- Remove internal functions from sampling __init__.py exports
- Remove fragile is_text property, use not is_tool_use instead
- Inline call_client into context.py, remove from run.py
- Fix SamplingMessage docs to use TextContent
- Handle result.text being None in doc examples
- Simplify client sampling docs to recommend OpenAISamplingHandler
- Add sampling_capabilities override documentation
- Raise iteration limit from 50 to 100
- Remove _parse_model_preferences duplication
- Use AsyncOpenAI in OpenAISamplingHandler
- Fix tool_choice docstring
* Fix OpenAI handler tests to use AsyncOpenAI
* Address remaining CodeRabbit review comments
- Fix message ordering in OpenAI handler: tool results now correctly
follow assistant message with tool_calls
- sample_step() now always includes assistant message in history
- Raise ValueError on JSON parse errors instead of silent {}
- Add has_sampling capability check when behavior is None
- Raise RuntimeError when structured output receives text response
- Wrap primitive result_type schemas in object wrapper
- Fix docs example using invalid SamplingMessage construction
- Add comprehensive client_sampling_test.py example
* Add return type annotation to OpenAISamplingHandler.__init__
* Use explicit 'is not None' check for sampling_capabilities defaulting
---------
Co-authored-by: claude[bot] <41898282+claude[bot]@users.noreply.github.com>
Co-authored-by: Bill Easton <strawgate@users.noreply.github.com>
* fix: skip TextIO log file test on Windows
Avoids PytestUnraisableExceptionWarning caused by ProactorEventLoop
cleanup timing issues with subprocess pipe transports.
* Use WindowsSelectorEventLoopPolicy to fix Windows test warnings
Re-add the SelectorEventLoop fix from bcd2e594 that was inadvertently
removed in cf101c2a. This fixes ProactorEventLoop cleanup warnings
on Windows CI without needing to skip individual tests.
The function set experimental={"tasks": {}} but per the MCP spec:
1. Tasks belong in capabilities.tasks, not capabilities.experimental
2. Clients only need to declare task capabilities if receiving task-augmented
requests from the server (bi-directional support)
For client→server task requests (tools/call with task=True), only the server
needs to declare task capabilities. The SDK's native session.initialize()
handles this correctly.
* feat: add PromptResult as canonical internal type for prompts
Applies the same pattern as ResourceContent to prompts. PromptResult
wraps messages with description and meta. Public render() can return
either list[PromptMessage] or PromptResult (backwards compatible),
while private _render() always returns PromptResult.
* docs: fix incorrect PromptResult return type in example
* feat: add PromptResult canonical type with meta support
* fix: address PR #2600 review comments
Fixes test failures and code quality issues identified in PR review:
- Update 3 tests in test_server_interactions.py to access PromptResult.messages[0] instead of indexing directly
- Fix ProxyPromptManager to preserve meta field when converting GetPromptResult to PromptResult
- Fix ProxyPrompt.render() to return PromptResult instead of deprecated list[PromptMessage], preventing fastmcp tags from leaking into runtime meta
- Fix mask_error_details initialization to respect explicit False values
- Fix exception re-raising to preserve tracebacks (use bare raise instead of raise e)
- Update testing documentation to use pytest -n auto for parallel execution
* feat: make ResourceContent the canonical internal type for resources
Add Resource._read() private method that always returns ResourceContent,
maintaining backwards compatibility for custom resources returning str/bytes
from read(). Includes deprecation warning when str/bytes is returned.
* fix: address review feedback for ResourceContent
- Remove ResourceContent from root exports (import from fastmcp.resources)
- Fix FunctionResource.read() return type to str | bytes | ResourceContent
- Decode base64 blobs in proxy when receiving from remote servers
- Preserve meta in ProxyResource cached content
* fix: add empty result guards in proxy resource reads
When a prompt function returned `mcp.types.PromptMessage` objects directly
and was executed as a task, the result serialization failed with
"'PromptMessage' object has no attribute 'to_mcp'".
The task result converter was calling `.to_mcp()` on what it thought was a
FastMCP wrapper type, but the import actually pulls in `mcp.types.PromptMessage`
directly, which is already the final MCP type. Removed the unnecessary
conversion.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* feat: handle error from the initialize middleware
In some situation, the initialize middleware can check the status of the
server and decide to raise an error.
Example use case: in a FastMCPProxy, an initialization middleware
overrides the on_initialize method and connect to the underlying proxied
client. When client respond with error, I want to pass this error to the
client.
* docs update
* test: use McpError assertions now that exception propagation is fixed
- Update tests to catch McpError specifically instead of generic Exception
- Remove commented-out code in low_level.py
---------
Co-authored-by: Jeremiah Lowin <153965+jlowin@users.noreply.github.com>
anyio task groups suppress exceptions when cancel_scope.cancel() is
called during cleanup. Capture exceptions before cleanup and re-raise
after task group exits cleanly.
Also preserve McpError type in client _connect() so callers can catch
protocol-level errors specifically.
When upstream OAuth providers don't return expires_in (like GitHub OAuth
Apps), use smart defaults: 1 hour if refresh token available, 1 year if
not. Adds fallback_access_token_expiry_seconds parameter to override.
Tools, resources, and prompts from servers mounted more than 2 levels
deep failed to invoke even though they were correctly listed.
The bug was in the routing methods which used manager methods that only
search locally, not through nested mounted servers. Changed to use
server-level methods that search recursively.
Fixes#2583
Proxied tool results now properly forward the meta attribute from upstream servers through ProxyToolManager and ProxyTool.
Co-authored-by: Jeremiah Lowin <153965+jlowin@users.noreply.github.com>
* Add comprehensive documentation for read-only tool patterns
Created new patterns guide explaining readOnlyHint annotation usage,
including practical examples, client-specific behavior, and best
practices for marking tools as read-only.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-authored-by: William Easton <strawgate@users.noreply.github.com>
* Reorganize read-only tools docs into tools section
Moved read-only tools documentation from standalone patterns page into the tools.mdx file as a "Using Annotation Hints" subsection. Condensed from 218 lines to ~50 lines focusing on practical usage while maintaining essential information about readOnlyHint and other annotations.
Changes:
- Added "Using Annotation Hints" subsection in tools.mdx after MCP Annotations
- Removed docs/patterns/read-only-tools.mdx
- Updated docs.json navigation to remove patterns entry
- Content now positioned as core tool feature rather than advanced pattern
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Bill Easton <strawgate@users.noreply.github.com>
---------
Co-authored-by: claude[bot] <41898282+claude[bot]@users.noreply.github.com>
Co-authored-by: William Easton <strawgate@users.noreply.github.com>
* [feat] expose get_session_id callback
* [test] add test for session id callback
* [fix] add test for uninitialized case and default to None
* [fix] add in changes based on reviewers
* SEP-1330 enum schema support for elicitation
* Add version badges for 2.14.0 elicitation features
* Fix Context.elicit() to handle SEP-1330 enum syntaxes
* Guard against empty list in elicit response_type
* Add guards for empty dict/list edge cases in elicit
* Refactor elicit: extract parsing and response handling to elicitation.py
Addresses code review feedback:
- Extract `get_task_capabilities()` to avoid duplicating the SEP-1686
capability structure across transports
- Add `_should_enable_component()` check before task routing for tools,
resources, and prompts to respect enable/tag filtering
- Simplify tasks/__init__.py to avoid circular import issues
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Add EventStore and SSE polling support (SEP-1699)
* Add close_sse_stream() method to Context
* Add SSE polling documentation
* Fix missing Context import in docs example
* Remove EventStore from root __init__.py, update docs imports
- Removed EventStore import and export from src/fastmcp/__init__.py
- Updated docs to import EventStore from fastmcp.server.event_store
- Resolves merge conflict by not exporting EventStore from root package
The task protocol (SEP-1686) is now always enabled - server always
registers task handlers and advertises task capabilities. Users still
opt into background execution at the server level (tasks=True) or
component level (task=True on tools, prompts, resources).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
The base Tool class now has an optional `execution` field for storing
task execution metadata (SEP-1686). This lets gateways/proxies preserve
execution info when forwarding tools from backends - previously this
metadata was lost because Tool had no way to store it.
FunctionTool continues to derive execution from task_config as before.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
When multiple servers with task-enabled tools are mounted into a parent,
their functions were all registered with Docket using `fn.__name__`. This
meant two mounted servers each having a function named `add` would both
register under `"add"`, with the second overwriting the first.
Now mounted functions use prefixed names matching their client-facing tool
names (e.g., `c1_add`, `c2_add`). Root server functions still use their
original names with no prefix.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
When a server is mounted on another, both were creating their own
Docket and Worker instances. With memory:// URLs they share the same
FakeServer queue but have separate result_storage instances. This
caused a race condition where results could be stored in one Docket's
storage but looked up in another's, returning None.
The fix marks mounted servers with `_is_mounted=True` flag so they
skip creating their own Docket/Worker. The parent's Docket handles
all task execution for mounted servers.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
