-
Notifications
You must be signed in to change notification settings - Fork 4k
Add Skills extension
#3485
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Open
vijaydeepsinha
wants to merge
20
commits into
modelcontextprotocol:main
Choose a base branch
from
vijaydeepsinha:sep-2640-python-sdk-support
base: main
Could not load branches
Branch not found: {{ refName }}
Loading
Could not load tags
Nothing to show
Loading
Are you sure you want to change the base?
Some commits from the old base branch may be removed from the timeline,
and old review comments may become outdated.
+1,869
−0
Open
Add Skills extension
#3485
Changes from all commits
Commits
Show all changes
20 commits
Select commit
Hold shift + click to select a range
a4c6ae3
Add SEP-2640 protocol types and validation
573fc06
Add SEP-2640 server support
0cc8dbc
Add SEP-2640 client support
0729d95
Document SEP-2640 Python SDK support
3a3725b
Strengthen SEP-2640 validation test coverage
d31511c
Rename skill resource-URI validator for clarity
c00a1e7
Extract parallel directory-URI validator in Skills handlers
b4edd53
Pin the skill-name grammar rejection cases
b845490
Fix documentation accuracy in the Skills guide
2f363d5
Add end-to-end dynamic-skill and cursor-resume client tests
1adba1d
Report Skills handler-output faults as INTERNAL_ERROR, not INVALID_PA…
70493de
Reject directory-shaped skill resource URIs and dot-segment directory…
38706c1
Seed the pagination cursor and preserve request _meta across skills p…
f1d2920
Correct Skills docs on read_skill_uri return type and verify on dynam…
63254f8
Construct ListSkillsParams _meta via model_validate in the skills cli…
7e32fc9
Test that request _meta is camelCase on the wire and snake_case to a …
473804e
Empty commit to re-trigger CI after a transient PyPI network flake
a2e49d5
empty coomit to re-trigger ci check
071ad0e
added changes for cache-attributes and its tests
6972162
Merge branch 'main' into sep-2640-python-sdk-support
vijaydeepsinha File filter
Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
There are no files selected for viewing
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
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,109 @@ | ||
| # Skills | ||
|
|
||
| [SEP-2640](https://github.com/modelcontextprotocol/modelcontextprotocol/pull/2640) defines a | ||
| convention for serving [Agent Skills](https://agentskills.io/) over MCP: a skill is a directory | ||
| of files — minimally a `SKILL.md` with YAML frontmatter — exposed as ordinary MCP resources, | ||
| conventionally under a `skill://` URI. A server enumerates its skills with `skills/list`, | ||
| answers for any one of them by URI with `skills/get`, and — optionally — lists a directory's | ||
| direct children with `resources/directory/read`. | ||
|
|
||
| The SDK ships this as the built-in `Skills` extension (`io.modelcontextprotocol/skills`). If | ||
| [Extensions](extensions.md) are new to you, skim that page first. | ||
|
|
||
| `Skills` provides the **protocol** primitives: request/response handling, capability | ||
| advertisement, and SEP-2640 conformance validation. It does not discover, read, or hash skills | ||
| from a filesystem — you supply handlers that answer from wherever your catalog actually lives | ||
| (a database, a generated index, an in-memory list, or a directory you walk yourself), and serve | ||
| each skill's files as ordinary resources through `MCPServer.add_resource` or an | ||
| `@mcp.resource(...)` template handler. | ||
|
|
||
| ## Serving a skill | ||
|
|
||
| ```python title="server.py" hl_lines="31-41 44-51 55" | ||
| --8<-- "docs_src/skills/tutorial001.py" | ||
| ``` | ||
|
|
||
| Three moves: | ||
|
|
||
| * `Skill(uri=..., frontmatter=..., resources=[...])`: one entry, identical in shape whether it | ||
| comes back from `skills/list` or `skills/get`. `resources` is the skill's complete file | ||
| manifest — every file, `SKILL.md` included, each with a `sha256:...` digest and byte size — or | ||
| the string `"dynamic"` for content generated on demand. | ||
| * `list_skills`/`get_skill`: plain async callables, invoked per request. `get_skill` **must** | ||
| answer for a skill even if a real `list_skills` implementation omitted it — SEP-2640 requires | ||
| a server to answer by URI for every skill it serves, listed or not. | ||
| * `mcp.add_resource(TextResource(uri=SKILL_URI, ...))`: the skill's actual file content, served | ||
| through the SDK's ordinary resource machinery. `Skills` never reads or writes resource content | ||
| itself. | ||
|
|
||
| `Skills(list_skills=..., get_skill=...)` is all a server needs; `resources/directory/read` is | ||
| optional (below). | ||
|
|
||
| ## Fetching a skill | ||
|
|
||
| ```python title="client.py" hl_lines="5" | ||
| --8<-- "docs_src/skills/tutorial001_client.py" | ||
| ``` | ||
|
|
||
| `list_skills` and `read_directory` follow `nextCursor` to completion, so you get every page's | ||
| skills or resources in one call; `get_skill` costs exactly one request. These three validate the | ||
| server's response against the SEP-2640 conformance rules before returning it — a name that doesn't | ||
| match its URI, a digest in the wrong shape, or an incomplete manifest raises `ValueError` rather | ||
| than reaching your code. `read_skill_uri` is the exception: a thin, discoverable alias for | ||
| `resources/read` that returns a `ReadResourceResult` (text or blob contents) and validates nothing | ||
| itself (see the next paragraph). | ||
|
|
||
| `verify_skill_resource(skill, uri, content)` checks a file's bytes — size, then SHA-256 digest — | ||
| against the entry you hold for it. Call it after `read_skill_uri` and before treating the content | ||
| as trustworthy: `resources/read` returns whatever bytes the server sends *right now*, verification | ||
| is what ties those bytes back to the manifest you already validated. It applies to a static | ||
| manifest only — a `"dynamic"` skill carries no digests, so calling it on one raises `ValueError`. | ||
|
|
||
| !!! warning | ||
| Skill content is untrusted model input, exactly like any other server-provided text. SEP-2640 | ||
| requires a host to tag it with its originating server before it reaches the model, and to | ||
| never grant the frontmatter's `allowed-tools` field (or any other permission-widening field) | ||
| without explicit per-skill user approval. Both are host responsibilities the SDK cannot | ||
| discharge for you — see the SEP's [Security Implications](https://github.com/modelcontextprotocol/modelcontextprotocol/pull/2640) | ||
| section before building a host on top of this extension. | ||
|
|
||
| ## Directory reads | ||
|
|
||
| A skill's instructions often point at a directory rather than a file ("pick the matching | ||
| template from `templates/`"). `resources/list` cannot answer that — it enumerates a server's | ||
| entire resource space, not one subtree — so SEP-2640 adds `resources/directory/read`, gated | ||
| behind the `directoryRead` capability setting: | ||
|
|
||
| ```python | ||
| mcp = MCPServer( | ||
| "catalog", | ||
| extensions=[ | ||
| Skills( | ||
| list_skills=list_skills, | ||
| get_skill=get_skill, | ||
| read_directory=read_directory, # lists uri's direct children | ||
| ) | ||
| ], | ||
| ) | ||
| ``` | ||
|
|
||
| Supplying `read_directory` advertises `{"directoryRead": true}` under the extension's | ||
| capabilities; omitting it advertises neither the setting nor the method — a client calling | ||
| `resources/directory/read` against such a server gets `METHOD_NOT_FOUND`. | ||
| `mcp.client.skills.read_directory` raises before sending if the connected server hasn't | ||
| advertised the setting. | ||
|
|
||
| ## Protocol version and caching | ||
|
|
||
| In protocol version `2026-07-28` and later, `skills/list` and `skills/get` results carry the base | ||
| protocol's caching fields, [`ttlMs` and `cacheScope`](https://github.com/modelcontextprotocol/modelcontextprotocol/pull/2549) — the | ||
| same freshness hint `tools/list`, `resources/list`, and `resources/read` carry. `Skills` fills | ||
| `cacheScope` with `"public"` when your handler leaves it unset, and omits both fields entirely on an | ||
| older connection, so you don't have to branch on protocol version yourself. | ||
|
|
||
| ## What this SDK doesn't do | ||
|
|
||
| `Skills` is a protocol adapter, not a skills provider. It has no opinion on where a skill's | ||
| bytes live, how they're indexed, or when a catalog is refreshed — that's for a higher-level | ||
| library, or your own handler, to decide. If you're looking for "scan this directory and serve | ||
| whatever's in it," you're looking for a provider built on top of `Skills`, not `Skills` itself. |
Empty file.
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
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,54 @@ | ||
| import hashlib | ||
| from typing import Any | ||
|
|
||
| from mcp.server.context import ServerRequestContext | ||
| from mcp.server.mcpserver import MCPServer | ||
| from mcp.server.mcpserver.resources import TextResource | ||
| from mcp.server.skills import Skills | ||
| from mcp.shared.exceptions import MCPError | ||
| from mcp.shared.skills import ( | ||
| GetSkillParams, | ||
| GetSkillResult, | ||
| ListSkillsParams, | ||
| ListSkillsResult, | ||
| Skill, | ||
| SkillResource, | ||
| ) | ||
| from mcp.types import INVALID_PARAMS | ||
|
|
||
| SKILL_URI = "skill://git-workflow/SKILL.md" | ||
| SKILL_MD = """\ | ||
| --- | ||
| name: git-workflow | ||
| description: Follow this team's Git conventions for branching and commits | ||
| --- | ||
| Branch from `main` using `type/short-description`. Write commit subjects in the | ||
| imperative mood, under 72 characters. | ||
| """ | ||
|
|
||
| GIT_WORKFLOW = Skill( | ||
| uri=SKILL_URI, | ||
| frontmatter={"name": "git-workflow", "description": "Follow this team's Git conventions for branching and commits"}, | ||
| resources=[ | ||
| SkillResource( | ||
| uri=SKILL_URI, | ||
| digest=f"sha256:{hashlib.sha256(SKILL_MD.encode()).hexdigest()}", | ||
| size=len(SKILL_MD.encode()), | ||
| ) | ||
| ], | ||
| ) | ||
|
|
||
|
|
||
| async def list_skills(ctx: ServerRequestContext[Any, Any], params: ListSkillsParams) -> ListSkillsResult: | ||
| return ListSkillsResult(skills=[GIT_WORKFLOW]) | ||
|
|
||
|
|
||
| async def get_skill(ctx: ServerRequestContext[Any, Any], params: GetSkillParams) -> GetSkillResult: | ||
| if params.uri != SKILL_URI: | ||
| raise MCPError(code=INVALID_PARAMS, message=f"unknown skill: {params.uri}") | ||
| return GetSkillResult(skill=GIT_WORKFLOW) | ||
|
|
||
|
|
||
| mcp = MCPServer("catalog", extensions=[Skills(list_skills=list_skills, get_skill=get_skill)]) | ||
| mcp.add_resource(TextResource(uri=SKILL_URI, name="SKILL.md", mime_type="text/markdown", text=SKILL_MD)) |
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
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,22 @@ | ||
| import anyio | ||
|
|
||
| from mcp import Client | ||
| from mcp.client.skills import get_skill, list_skills, read_skill_uri, verify_skill_resource | ||
| from mcp.types import TextResourceContents | ||
|
|
||
|
|
||
| async def main() -> None: | ||
| async with Client("http://localhost:8000/mcp") as client: | ||
| for skill in await list_skills(client.session): | ||
| print(skill.uri, skill.frontmatter["description"]) | ||
|
|
||
| skill = await get_skill(client.session, "skill://git-workflow/SKILL.md") | ||
| result = await read_skill_uri(client.session, skill.uri) | ||
| content = result.contents[0] | ||
| if isinstance(content, TextResourceContents): | ||
| verify_skill_resource(skill, skill.uri, content.text.encode()) | ||
| print(content.text) | ||
|
|
||
|
|
||
| if __name__ == "__main__": | ||
| anyio.run(main) |
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
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
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,130 @@ | ||
| """Client-side convenience wrappers for the Skills extension (SEP-2640). | ||
|
|
||
| SEP-2640 needs no client-side method registration: `skills/list`, `skills/get`, | ||
| and `resources/directory/read` are ordinary vendor requests sent through | ||
| `ClientSession.send_request`, exactly like [Extension verbs](../advanced/extensions.md#extension-verbs). | ||
| The functions below are the thin, named wrappers SEP-2640's "SDKs: Convenience | ||
| Wrappers" section recommends — each validates the server's advertised support | ||
| before sending, and `list_skills`/`read_directory` follow `nextCursor` to | ||
| completion so a caller sees one page's worth of ergonomics regardless of how | ||
| many requests it took. | ||
|
|
||
| async with Client("http://localhost:8000/mcp") as client: | ||
| for skill in await list_skills(client.session): | ||
| print(skill.uri, skill.frontmatter["description"]) | ||
| """ | ||
|
|
||
| from __future__ import annotations | ||
|
|
||
| from mcp_types import ReadResourceResult, Resource | ||
|
|
||
| from mcp.client.session import ClientSession | ||
| from mcp.shared.skills import ( | ||
| EXTENSION_ID, | ||
| GetSkillParams, | ||
| GetSkillRequest, | ||
| GetSkillResult, | ||
| ListSkillsParams, | ||
| ListSkillsRequest, | ||
| ListSkillsResult, | ||
| ReadDirectoryParams, | ||
| ReadDirectoryRequest, | ||
| ReadDirectoryResult, | ||
| Skill, | ||
| validate_directory_result, | ||
| validate_list_result, | ||
| validate_skill, | ||
| ) | ||
| from mcp.shared.skills import verify_skill_resource as verify_skill_resource | ||
|
|
||
| __all__ = ["get_skill", "list_skills", "read_directory", "read_skill_uri", "verify_skill_resource"] | ||
|
Member
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I don't think this is the API that we want here. The There's probably something smart we can do type wise to offer methods on the client's side... Something to explore: async with Client(extensions={"skills": Skill()}) as client:
client.extensions["skills"].<autocomplete would work here>Can you check this, please? |
||
|
|
||
|
|
||
| def _require_extension(session: ClientSession, *, directory_read: bool = False) -> None: | ||
| capabilities = session.server_capabilities | ||
| settings = (capabilities.extensions or {}).get(EXTENSION_ID) if capabilities else None | ||
| if settings is None: | ||
| raise ValueError(f"server does not advertise the {EXTENSION_ID!r} extension") | ||
| if directory_read and not settings.get("directoryRead"): | ||
| raise ValueError(f"server does not advertise {EXTENSION_ID!r}'s directoryRead setting") | ||
|
|
||
|
|
||
| async def list_skills(session: ClientSession, params: ListSkillsParams | None = None) -> list[Skill]: | ||
| """Call `skills/list`, following `nextCursor` to completion, and validate the result. | ||
|
|
||
| Raises: | ||
| ValueError: If the server doesn't advertise the Skills extension, or | ||
| its response is not SEP-2640 conformant. | ||
| """ | ||
| _require_extension(session) | ||
| base = params if params is not None else ListSkillsParams() | ||
| cursor = base.cursor | ||
| skills: list[Skill] = [] | ||
| seen_cursors: set[str] = {cursor} if cursor is not None else set() | ||
| while True: | ||
| page = await session.send_request( | ||
| ListSkillsRequest(params=base.model_copy(update={"cursor": cursor})), ListSkillsResult | ||
| ) | ||
| validate_list_result(page) | ||
| skills.extend(page.skills) | ||
| if page.next_cursor is None: | ||
| return skills | ||
| if page.next_cursor in seen_cursors: | ||
| raise ValueError(f"server repeated skills/list pagination cursor {page.next_cursor!r}") | ||
| seen_cursors.add(page.next_cursor) | ||
| cursor = page.next_cursor | ||
|
|
||
|
|
||
| async def get_skill(session: ClientSession, uri: str) -> Skill: | ||
| """Call `skills/get` for `uri` and validate the result. | ||
|
|
||
| Unlike `list_skills`, this succeeds for a skill absent from any listing — | ||
| per SEP-2640, a server MUST answer `skills/get` for every skill it serves. | ||
|
|
||
| Raises: | ||
| ValueError: If the server doesn't advertise the Skills extension, its | ||
| response names a different skill, or the skill is not conformant. | ||
| """ | ||
| _require_extension(session) | ||
| result = await session.send_request(GetSkillRequest(params=GetSkillParams(uri=uri)), GetSkillResult) | ||
| if result.skill.uri != uri: | ||
| raise ValueError(f"server returned skill {result.skill.uri!r} for requested {uri!r}") | ||
| validate_skill(result.skill) | ||
| return result.skill | ||
|
|
||
|
|
||
| async def read_skill_uri(session: ClientSession, uri: str) -> ReadResourceResult: | ||
| """Read a skill file's content via `resources/read`. | ||
|
|
||
| A thin, discoverable alias: works for any `skill://` (or other-scheme) | ||
| file regardless of whether the skill was ever enumerated. Verify the | ||
| result against a held `Skill` entry with `verify_skill_resource` before | ||
| treating it as trusted content — this call does not verify anything itself. | ||
| """ | ||
| return await session.read_resource(uri) | ||
|
|
||
|
|
||
| async def read_directory(session: ClientSession, uri: str, params: ReadDirectoryParams | None = None) -> list[Resource]: | ||
| """Call `resources/directory/read` for `uri`, following `nextCursor` to completion. | ||
|
|
||
| Raises: | ||
| ValueError: If the server doesn't advertise the `directoryRead` | ||
| setting, or its response is not a valid child listing of `uri`. | ||
| """ | ||
| _require_extension(session, directory_read=True) | ||
| base = params if params is not None else ReadDirectoryParams(uri=uri) | ||
| cursor = base.cursor | ||
| resources: list[Resource] = [] | ||
| seen_cursors: set[str] = {cursor} if cursor is not None else set() | ||
| while True: | ||
| page = await session.send_request( | ||
| ReadDirectoryRequest(params=base.model_copy(update={"uri": uri, "cursor": cursor})), ReadDirectoryResult | ||
| ) | ||
| validate_directory_result(uri, page) | ||
| resources.extend(page.resources) | ||
| if page.next_cursor is None: | ||
| return resources | ||
| if page.next_cursor in seen_cursors: | ||
| raise ValueError(f"server repeated resources/directory/read pagination cursor {page.next_cursor!r}") | ||
| seen_cursors.add(page.next_cursor) | ||
| cursor = page.next_cursor | ||
Oops, something went wrong.
Oops, something went wrong.
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.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Probably not your fault, but we are doing a lot of
validate_*calls everywhere. That a smell.We should either leverage Pydantic a bit more (using
Annotated), or not have them.This point is only actionable if you can find ways to avoid this many
validate_... Which you probably can 👀