Python: Introducing Local MCP Servers

[Eduard van Valkenburg (eavanvalkenburg)]

Motivation and Context

Adds MCP tools to AF

Closes #292

Description

Contribution Checklist

• The code builds clean without any errors or warnings
• The PR follows the Contribution Guidelines
• All unit tests pass, and I have added new tests where possible
• I didn't break anyone 😄

[Copilot]

Pull Request Overview

This PR adds MCP (Model Context Protocol) tools integration to the Agent Framework, enabling agents to connect to and use MCP servers for extended functionality. The implementation provides multiple connection types (stdio, SSE, websocket, and streamable HTTP) and seamlessly integrates MCP tools with existing agent capabilities.

• Adds comprehensive MCP server integration with support for multiple transport protocols
• Extends ChatClientAgent to accept and manage MCP tools alongside traditional tools
• Includes sample implementations demonstrating MCP usage with both OpenAI chat and responses clients

Reviewed Changes

Copilot reviewed 10 out of 11 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
python/packages/main/agent_framework/_mcp.py	Core MCP implementation with server classes and protocol conversion utilities
python/packages/main/agent_framework/_agents.py	Extended ChatClientAgent to support MCP tools with proper context management
python/packages/main/agent_framework/init.py	Exported MCP classes for public API
python/packages/main/pyproject.toml	Added mcp>=1.12 dependency
python/packages/main/tests/main/test_mcp.py	Comprehensive test suite for MCP functionality
python/samples/getting_started/agents/openai_chat_client/openai_chat_client_with_local_mcp.py	Sample showing MCP usage with OpenAI chat client
python/samples/getting_started/agents/openai_responses_client/openai_responses_client_with_local_mcp.py	Sample showing MCP usage with OpenAI responses client
python/pyproject.toml	Updated install command to include all extras
python/.vscode/launch.json	Added justMyCode: false for debugging
.github/workflows/python-merge-tests.yml	Added LOCAL_MCP_URL environment variable for CI

[Eric Zhu (ekzhu)]

Looks great!

Are we planning to also include the sampling and elicitation features of MCP used currently?

See the AG branch that uses them to drive the agent actions: microsoft/autogen#6833

[Eduard van Valkenburg (eavanvalkenburg)]

Eric Zhu (@ekzhu) sampling already works, you can pass a ChatClient to the constructor of the McpServer and that is used to do sampling if requested. I have not looked at elicitation, but I will, thanks for pointing to that

[Eduard van Valkenburg (eavanvalkenburg)]

added a issue to track elicitation #398

[Victor Dibia (victordibia)]

This looks great Eduard van Valkenburg (@eavanvalkenburg)

One future thing to track is our thinking on how to “pipe” in streamed updates from tools. Many tools have agentic behaviors (long running) can send notifications (resource updates, progress notifications, logging ). These may include useful information or even partial results which the rest of the application can benefit from (as well as UI updates). Might be worth exploring a setup where devs can pass in a message handler callback for these scenarios.

related post on long runnning tools here.

[Eduard van Valkenburg (eavanvalkenburg)]

thanks Victor Dibia (@victordibia), peterychang is working on broader support for long-running things, and I think this might fit in there. we do already have a message handler in here, so that can easily be extended to handle other updates, and elicitation depends on broader user request structures we are working on, as noted, already a issue created for tracking that.

[Eduard van Valkenburg (eavanvalkenburg)]

Python Test Coverage Report

File	Stmts	Miss	Cover	Missing
packages/azure/agent_framework_azure
_assistants_client.py	36	8	77%	100, 103, 114–117, 121–122
_chat_client.py	60	5	91%	103, 151, 153, 166–167
_entra_id_authentication.py	23	16	30%	35–36, 40–44, 46, 66–67, 71–75, 77
_responses_client.py	31	2	93%	92, 99
_shared.py	81	8	90%	158–159, 221, 224, 237, 239, 249, 257
packages/foundry/agent_framework_foundry
_chat_client.py	229	120	47%	144–145, 152–153, 156–159, 161–162, 165–167, 169, 178, 200, 218, 230, 233, 237–238, 241, 244, 247–248, 264–269, 289, 292–293, 295, 297, 301–302, 305, 308, 314, 318–319, 321–322, 328–329, 335, 337, 342–343, 345, 347, 349, 359–362, 364–365, 371–374, 382–386, 394, 399–401, 409, 414, 421, 430, 441, 443–449, 453, 458, 483–491, 493–494, 496–500, 505, 511, 528–529, 531, 538–545, 561, 586, 605–606
packages/main/agent_framework
_agents.py	192	22	88%	67, 222, 417, 437–440, 447, 457, 527–528, 530, 533, 651–652, 654, 657, 765, 777, 782–783, 785
_cancellation_token.py	31	31	0%	3–6, 10, 13–16, 18, 20–24, 26, 28–29, 31, 33–35, 37, 39, 41–43, 46–47, 49–50
_clients.py	156	17	89%	56, 66–68, 126–127, 145–146, 189–190, 393, 437, 504, 584, 634, 662, 664
_logging.py	8	0	100%
_mcp.py	308	120	61%	133, 143–144, 165, 188, 198, 202, 204, 206, 253, 257–262, 265–266, 276–278, 281–283, 285–290, 292–294, 297–298, 310–311, 315–320, 326–327, 331–332, 336, 338, 342–343, 347, 361, 376–386, 391, 394–395, 399, 414, 417–418, 422, 443, 448, 450, 457–458, 463, 465, 471–474, 481–485, 558, 563–567, 632, 635–643, 710, 713–723, 779, 782–784
_pydantic.py	27	1	96%	61
_tools.py	131	5	96%	113, 159, 239, 264, 313
_types.py	739	77	89%	162, 180–184, 216, 456–461, 463, 466–469, 471–472, 492, 497, 501, 505–507, 560–565, 567, 570–573, 575–576, 592–602, 605–609, 712, 942, 946, 948, 1437, 1441–1443, 1474, 1638, 1643, 1677, 1734, 1738, 1761, 1793, 1805, 1837–1839, 1906, 1911
exceptions.py	31	0	100%
guard_rails.py	10	10	0%	4, 6–7, 9, 12–13, 16, 21–22, 25
telemetry.py	297	27	90%	315–317, 420, 422, 424, 426, 428, 430, 432, 434, 436, 476, 481, 654, 656, 658, 660, 662, 664, 666, 668, 670, 672, 674, 701, 724
packages/main/agent_framework/openai
_assistants_client.py	188	128	31%	146, 159, 162, 166–167, 170, 173, 176–177, 209, 211, 213, 215, 218, 221, 224, 228, 232–233, 235–238, 242, 244, 249–250, 252, 254, 256, 259, 261–264, 272–276, 278–280, 288–291, 299, 304–305, 312, 321, 332, 334–339, 341, 349, 351–355, 357–358, 360–369, 371–372, 374–376, 381, 386–387, 392–393, 395, 400–403, 405, 407, 409–413, 416–419, 421–424, 431–432, 434–435, 437, 443–444, 446–447, 451, 453, 460, 462–463, 465–467, 469, 479–480
_chat_client.py	179	36	79%	98–99, 103, 116–123, 125–126, 133, 137, 159, 181, 208, 219–221, 246–247, 249, 255, 298, 301, 303, 305, 307–309, 323–324, 330, 343
_exceptions.py	33	0	100%
_responses_client.py	283	218	22%	142, 228, 251, 262–265, 269–270, 272–273, 278–282, 286, 290–291, 303–307, 311–312, 315–316, 318, 322–323, 326–329, 333, 337–338, 348–361, 367–370, 379–380, 382–383, 387, 389–392, 394–395, 397, 399–406, 425–428, 433–434, 436, 446, 448–451, 470–474, 477–483, 492–493, 503–504, 517–518, 534–538, 541–547, 554–560, 568, 570–572, 581–583, 590–593, 601–609, 618–621, 623–647, 656–657, 659, 669, 674–678, 686–687, 690–707, 716–718, 725, 727, 732–736, 741–742, 746–747, 750
_shared.py	73	13	82%	55–60, 62–65, 67, 171, 182
packages/workflow/agent_framework_workflow
_checkpoint.py	122	3	97%	40, 165–166
_const.py	1	0	100%
_edge.py	64	5	92%	71, 74, 101, 139, 153
_events.py	49	9	81%	37, 41, 49, 53, 100, 124, 128, 136, 140
_executor.py	109	20	81%	62, 80, 236–240, 245, 247–250, 254–256, 258, 262, 264–265, 330
_runner.py	170	29	82%	84–86, 88, 134, 167–169, 189–193, 197, 201–202, 223–224, 236–237, 248–250, 254, 261–263, 268–269
_runner_context.py	117	6	94%	205, 208, 227, 248, 270, 285
_shared_state.py	35	5	85%	28–29, 33–34, 68
_typing_utils.py	17	8	52%	26, 33–36, 43–44, 50
_validation.py	206	21	89%	178, 296, 299, 357, 361, 448, 450, 453–454, 456–458, 460–462, 465–469, 474
_workflow.py	176	29	83%	44, 109, 131, 181, 189–191, 203, 217, 319, 339–340, 342–344, 359–361, 369–370, 379–385, 390, 392
_workflow_context.py	37	9	75%	40, 77, 86, 91, 99–100, 104–106
TOTAL	4249	1008	76%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
422	31 💤	0 ❌	0 🔥	7.527s ⏱️