Agent</th>	Method</th>	MCP entity</th></tr></thead>
Singleton</td>	No parameters</td>	Resource</td></tr>
Non-singleton</td>	No parameters</td>	Resource template</td></tr>
Any</td>	Has parameters</td>	Tool</td></tr> </tbody></table> Metadata</h3> For every agent and agent method, we can attach a description</strong> and a prompt</strong>:</p> @</span>description</span>(</span>"Increments the counter by the number provided in the `n` parameter"</span>)</span></span> @</span>prompt</span>(</span>"Increment by a given number"</span>)</span></span> async</span> increment_by</span>(n: number): </span>Promise</span><</span>number</span>></span> {</span></span> // ...</span></span> }</span></span></code></pre>#[description(</span>"Increments the counter by the number provided in the `n` parameter"</span>)]</span></span> #[prompt(</span>"Increment by a given number"</span>)]</span></span> fn</span> increment_by</span>(&</span>mut</span> self</span>,</span> n</span>:</span> u32</span>) -></span> u32</span>;</span></span></code></pre>@description(</span>"Increments the counter by the number provided in the `n` parameter"</span>)</span></span> @prompt(</span>"Increment by a given number"</span>)</span></span> def</span> incrementBy</span>(n:</span> Int</span>):</span> Future</span>[</span>Int</span>]</span></span></code></pre>///\| Increments the counter by the number provided in the `n` parameter</span></span> #derive.prompt_hint</span>(</span>"Increment by a given number"</span>)</span></span> pub fn</span> Counter</span>::</span>increment</span>(</span>self</span> : </span>Self</span>, </span>n</span>: </span>UInt32</span>) </span>-></span> UInt32</span> {</span></span></code></pre></div> Both are optional, and both are added to the MCP metadata.</p> Special data types</h3> It's not strictly related to the MCP feature, and not even new in Golem 1.5</strong>, but the three special data types supported by all the Golem SDKs are a good match for exposing some special tools and resources through the MCP protocol:</p> Unstructured text</h4> Any method parameter or return type can be defined as unstructured text</strong>. Optionally a set of allowed language codes</strong> can be attached to the type:</p> myMethod</span>(</span></span> anyText: UnstructuredText,</span></span> constrainedText: UnstructuredText</span><</span>[</span>'en'</span>,</span> 'de'</span>]</span>></span></span> ) {</span></span> // ...</span></span> }</span></span></code></pre>#[derive(</span>AllowedLanguages</span>)]</span></span> enum</span> MyLangs</span> {</span> En</span>,</span> #[code(</span>"de"</span>)]</span> German</span> }</span></span> </span> fn</span> my_method</span>(</span></span> &</span>self</span>, </span></span> any_text</span>:</span> UnstructuredText</span>, </span></span> constrained_text</span>:</span> UnstructuredText</span><</span>MyLangs</span>></span></span> );</span></span></code></pre>def</span> myMethod</span>(</span></span> anyText:</span> TextSegment</span>[</span>AllowedLanguages</span>.</span>Any</span>],</span></span> constrainedText:</span> TextSegment</span>[</span>MyLangs</span>]</span></span> )</span></span> </span> sealed trait</span> MyLangs</span></span> object</span> MyLangs</span> {</span></span> case object</span> En</span> extends</span> MyLangs</span></span> </span> @golem.runtime.annotations.languageCode(</span>"de"</span>)</span></span> case object</span> German</span> extends</span> MyLangs</span></span> </span> implicit val</span> allowed</span>:</span> AllowedLanguages</span>[</span>MyLangs</span>] =</span></span> golem.runtime.macros.</span>AllowedLanguagesDerivation</span>.derived </span></span> }</span></span> </span></code></pre>#derive.text_languages</span>(</span>"constrained_text"</span>,</span> "en"</span>,</span> "de"</span>)</span></span> pub fn</span> MyAgent</span>::</span>my_method</span>(</span></span> self</span> : </span>Self</span>,</span></span> any_text</span> : </span>UnstructuredText</span>,</span></span> constrained_text</span> : </span>UnstructuredText</span>,</span></span> ) </span>-></span> Unit</span> {</span></span> ...</span></span> }</span></span></code></pre></div> Unstructured binary</h4> Similarly to unstructured text</strong>, we can also use unstructured binary</strong> parameters and return types, and optionally define the allowed MIME types</em> for them:</p> myMethod</span>(</span></span> anyBinary: UnstructuredBinary,</span></span> image: UnstructuredBinary</span><</span>[</span>'image/png'</span>,</span> 'image/jpeg'</span>]</span>></span></span> ) {</span></span> // ...</span></span> }</span></span></code></pre>#[derive(</span>Debug</span>,</span> Clone</span>,</span> AllowedMimeTypes</span>)]</span></span> enum</span> Image</span> {</span></span> #[mime_type(</span>"image/png"</span>)]</span></span> Png</span>,</span></span> #[mime_type(</span>"image/jpeg"</span>)]</span></span> Jpeg</span>,</span></span> }</span></span> </span> fn</span> my_method</span>(</span></span> &</span>self</span>, </span></span> any_binary</span>:</span> UnstructuredBinary</span>, </span></span> image</span>:</span> UnstructuredBinary</span><</span>Image</span>></span></span> );</span></span></code></pre>def</span> myMethod</span>(</span></span> anyBinary:</span> BinarySegment</span>[</span>AllowedMimeTypes</span>.</span>Any</span>],</span></span> image:</span> BinarySegment</span>[</span>Image</span>]</span></span> )</span></span> </span> sealed trait</span> Image</span></span> object</span> Image</span> {</span></span> @golem.runtime.annotations.mimeType(</span>"image/png"</span>)</span></span> case object</span> Png</span> extends</span> Image</span></span> @golem.runtime.annotations.mimeType(</span>"image/jpeg"</span>)</span></span> case object</span> Jpeg</span> extends</span> Image</span></span> </span> implicit val</span> allowed</span>:</span> AllowedMimeTypes</span>[</span>Image</span>] =</span></span> golem.runtime.macros.</span>AllowedMimeTypesDerivation</span>.derived</span></span> }</span></span></code></pre>#derive.mime_types</span>(</span>"image"</span>,</span> "image/png"</span>,</span> "image/jpeg"</span>)</span></span> pub fn</span> MyAgent</span>::</span>my_method</span>(</span></span> self</span> : </span>Self</span>,</span></span> any_binary</span> : </span>UnstructuredBinary</span>,</span></span> image</span> : </span>UnstructuredBinary</span>,</span></span> ) </span>-></span> Unit</span> {</span></span> ...</span></span> }</span></span></code></pre></div> Multimodal</h4> Finally there is a special parameter type called multimodal</strong>, which is a special way to define methods (tools) that can work on multiple types of input. The default multimodal type just allows pasting either text or binary, but it is fully customizable with the above defined language and MIME type constraints, and can also include structured data. By using multimodal types, and not just modelling the same input using custom data types, Golem can map these definitions better to MCP concepts.</p> The simplest version just accepts either an arbitrary text, or an arbitrary binary:</p> textOrBinary</span>(input: Multimodal) {</span> ...</span> }</span></span></code></pre>fn</span> text_or_binary</span>(&</span>self</span>,</span> input</span>:</span> Multimodal</span>) -></span> Multimodal</span> {</span> input</span> }</span></span></code></pre>def</span> textOrBinary</span>(input:</span> MultimodalItems</span>.</span>Basic</span>):</span> Future</span>[</span>MultimodalItems</span>.</span>Basic</span>]</span></span></code></pre>pub fn</span> MyAgent</span>::</span>text_or_binary</span>(</span></span> self</span> : </span>Self</span>,</span></span> input</span> : @types.</span>Multimodal</span>[</span>TextOrBinary</span>],</span></span> ) </span>-></span> String</span> { ... }</span></span></code></pre></div> We can add a third option in the form of a structured data type to this:</p> type</span> MyStructuredType</span> =</span> {</span> ...</span>}</span></span> textOrBinaryOrStructured</span>(input</span>:</span> MultimodalCustom</span><</span>MyStructuredType</span>>) {</span> ...</span> }</span></span></code></pre>#[derive(</span>Schema</span>)]</span></span> struct</span> MyStructuredType</span> {</span> /* ... /</span> }</span></span> </span> fn</span> text_or_binary_or_structured</span>(</span></span> &</span>self</span>,</span></span> input</span>:</span> MultimodalCustom</span><</span>MyStructuredType</span>>,</span></span> ) -></span> MultimodalCustom</span><</span>MyStructuredType</span>> {</span> input</span> }</span></span></code></pre>final case class</span> MyStructuredType</span>(</span>/ ... /</span>)</span></span> object</span> MyStructuredType</span> {</span> implicit val</span> schema</span>:</span> Schema</span>[</span>MyStructuredType</span>] =</span> Schema</span>.derived }</span></span> </span> def</span> textOrBinaryOrStructured</span>(</span></span> input:</span> MultimodalItems</span>.</span>WithCustom</span>[</span>MyStructuredType</span>]</span></span> ):</span> Future</span>[</span>MultimodalItems</span>.</span>WithCustom</span>[</span>MyStructuredType</span>]]</span></span></code></pre>#derive.golem_schema</span></span> pub</span>(</span>all</span>) </span>struct</span> MyStructuredType</span> { / ... / }</span></span> </span> pub fn</span> MyAgent</span>::</span>text_or_binary_or_structured</span>(</span></span> self</span> : </span>Self</span>,</span></span> input</span> : @types.</span>Multimodal</span>[</span>CustomModality</span>[</span>MyStructuredType</span>]],</span></span> ) </span>-></span> String</span> { ... }</span></span></code></pre></div> Or we can fully customize the multimodal behavior by defining our own variant type it maps to:</p> export type</span> TextOrImage</span> =</span></span> \|</span> { tag</span>:</span> 'text'</span>; val</span>:</span> UnstructuredText</span><[</span>'en'</span>,</span> 'de'</span>>] }</span></span> \|</span> { tag</span>:</span> 'image'</span>; val</span>:</span> UnstructuredBinary</span><[</span>'image/jpeg'</span>,</span> 'image/png'</span>]> };</span></span> fullyCustom</span>(input</span>:</span> MultimodalAdvanced</span><</span>TextOrImage</span>>) {</span> ...</span> }</span></span></code></pre>#[derive(</span>AllowedLanguages</span>)]</span></span> enum</span> TextLang</span> {</span> En</span>,</span> #[code(</span>"de"</span>)]</span> German</span> }</span></span> </span> #[derive(</span>AllowedMimeTypes</span>)]</span></span> enum</span> ImageType</span> {</span></span> #[mime_type(</span>"image/jpeg"</span>)]</span> Jpeg</span>,</span></span> #[mime_type(</span>"image/png"</span>)]</span> Png</span>,</span></span> }</span></span> </span> #[derive(</span>Schema</span>,</span> MultimodalSchema</span>)]</span></span> enum</span> TextOrImage</span> {</span></span> Text</span>(</span>UnstructuredText</span><</span>TextLang</span>>),</span></span> Image</span>(</span>UnstructuredBinary</span><</span>ImageType</span>>),</span></span> }</span></span> </span> fn</span> fully_custom</span>(</span></span> &</span>self</span>,</span></span> input</span>:</span> MultimodalAdvanced</span><</span>TextOrImage</span>>,</span></span> ) -></span> MultimodalAdvanced</span><</span>TextOrImage</span>> {</span> input</span> }</span></span></code></pre>sealed trait</span> TextLang</span></span> object</span> TextLang</span> {</span></span> @golem.runtime.annotations.languageCode(</span>"en"</span>)</span></span> case object</span> En</span> extends</span> TextLang</span></span> @golem.runtime.annotations.languageCode(</span>"de"</span>)</span></span> case object</span> De</span> extends</span> TextLang</span></span> implicit val</span> allowed</span>:</span> AllowedLanguages</span>[</span>TextLang</span>] =</span></span> golem.runtime.macros.</span>AllowedLanguagesDerivation</span>.derived</span></span> }</span></span> </span> sealed trait</span> ImageType</span></span> object</span> ImageType</span> {</span></span> @golem.runtime.annotations.mimeType(</span>"image/jpeg"</span>)</span></span> case object</span> Jpeg</span> extends</span> ImageType</span></span> @golem.runtime.annotations.mimeType(</span>"image/png"</span>)</span></span> case object</span> Png</span> extends</span> ImageType</span></span> implicit val</span> allowed</span>:</span> AllowedMimeTypes</span>[</span>ImageType</span>] =</span></span> golem.runtime.macros.</span>AllowedMimeTypesDerivation</span>.derived</span></span> }</span></span> </span> final case class</span> TextOrImage</span>(</span></span> text:</span> TextSegment</span>[</span>TextLang</span>],</span></span> image:</span> BinarySegment</span>[</span>ImageType</span>],</span></span> )</span></span> object</span> TextOrImage</span> {</span> implicit val</span> schema</span>:</span> GolemSchema</span>[</span>TextOrImage</span>] =</span> / derived */</span> }</span></span> </span> def</span> fullyCustom</span>(input:</span> Multimodal</span>[</span>TextOrImage</span>]):</span> Future</span>[</span>Multimodal</span>[</span>TextOrImage</span>]]</span></span></code></pre>#derive.multimodal</span></span> pub</span>(</span>all</span>) </span>enum</span> TextOrImage</span> {</span></span> Text</span>(</span>UnstructuredText</span>)</span></span> Image</span>(</span>UnstructuredBinary</span>)</span></span> }</span></span> </span> #derive.text_languages</span>(</span>"input.Text"</span>,</span> "en"</span>,</span> "de"</span>)</span></span> #derive.mime_types</span>(</span>"input.Image"</span>,</span> "image/jpeg"</span>,</span> "image/png"</span>)</span></span> pub fn</span> MyAgent</span>::</span>fully_custom</span>(</span></span> self</span> : </span>Self</span>,</span></span> input</span> : @types.</span>Multimodal</span>[</span>TextOrImage</span>],</span></span> ) </span>-></span> String</span> { ... }</span></span></code></pre></div> Remarks</h4> None of these special data types are MCP specific - using them in our agent code is not constraining them to be only called through MCP, they can still be invoked through agent-to-agent communication, mapped to HTTP APIs and so on.</p>

Golem 1.5 features - Part 3: MCP

2026-04-11T00:00:00+00:00

Introduction</h2>

I am writing a series of short posts showcasing the new features of Golem 1.5, to be released at the end of April, 2026. The episodes of this series will be short and assume the reader knows what Golem is. Check my other Golem-related posts</a> for more information!

Parts released so far:

Part 1: Code-first routes</a></li>
Part 2: Webhooks</a></li>
Part 3: MCP</a></li>
Part 4: Node.js compatibility</a></li>
Part 5: Scala support</a></li> </ul>
MCP</h2>
MCP (Model Context Protocol)</a> became a standard way to connect AI applications. With the new Golem release any Golem application can be automatically exposed through MCP. It does not require any code written, MCP is available for any agent automatically, but it needs to be enabled in the application manifest:
Enabling</h3>
The same way how we can deploy HTTP APIs with the application manifest, we can add an mcp</code> section and choose which agents to deploy to which subdomains per environment:
mcp: deployments: local: - domain: mcp-demo.localhost:9007 agents: CounterAgent: {}</code></pre>
Security</h3>
The above manifest section already exposes the listed agent through MCP, but it is not protected by any form of authentication. Similar to how we can protect HTTP endpoints with OAuth, we can attach a security scheme to our MCP deployment, let's call it mcp-oauth</code>:
mcp: deployments: local: - domain: mcp-demo.localhost:9007 agents: CounterAgent: securityScheme: mcp-oauth</code></pre> To set that up we need an OAuth provider and have to create a security scheme using the golem</code> CLI. The provider can be one of the common ones like Google, etc., or any custom one. In this post we are going to use the mock-oauth2-server</code> docker container:
docker run -d \ --name "golem-mock-oauth2" \ -p "9099:8080" \ ghcr.io/navikt/mock-oauth2-server:2.1.10 CLIENT_ID="golem-mcp-client" CLIENT_SECRET="golem-mcp-secret" REDIRECT_URL="http://mcp-demo.localhost:9007/mcp/oauth/callback" golem -L api security-scheme create \ --provider-type custom \ --custom-provider-name "mock-oauth2" \ --custom-issuer-url "http://localhost:9099/golem" \ --client-id "${CLIENT_ID}" \ --client-secret "${CLIENT_SECRET}" \ --scope openid \ --scope email \ --scope profile \ --redirect-url "${REDIRECT_URL}" \ mcp-oauth</code></pre> With that set, and running golem deploy</code>, our MCP server is ready to be used at http://mcp-demo.localhost:9007/mcp</code> using streamable HTTP protocol, authenticated by the mock OAuth2 server.
Demo</h3> We can prove this by simply creating the default template (which simply implements a stateful counter), let's use the Rust one for this example and modify it slightly: #[agent_definition(mount = "/counters/{name}")] pub trait CounterAgent { // The agent constructor, its parameters identify the agent fn new(name: String) -> Self; #[description("Increment by a given number")] fn increment_by(&mut self, n: u32) -> u32; } struct CounterImpl { _name: String, count: u32, } #[agent_implementation] impl CounterAgent for CounterImpl { fn new(name: String) -> Self { Self { _name: name, count: 0, } } fn increment_by(&mut self, n: u32) -> u32 { self.count += n; self.count } }</code></pre> We replaced the default increment</code> method with a parametrized increment_by</code>, which is going to be mapped into an MCP tool in our MCP server automatically. If we start the MCP Inspector</a>: npx @modelcontextprotocol/inspector node build/index.js</code></pre> then click on the quick auth flow and connect, we can go to the Tools page and see our counter incrementation tool: We can pass the Counter</code> agent's constructor parameter, name</code> and the increment_by</code> method's n</code> parameter, and invoke it through an MCP tool. Mapping</h3> As demonstrated above, increment_by</code> has been automatically exported as a tool. But MCP not only defines tools, but also resources and resource templates. We support these with an automatic mapping in the following way: Agent</th> Method</th> MCP entity</th></tr></thead> Singleton</td> No parameters</td> Resource</td></tr> Non-singleton</td> No parameters</td> Resource template</td></tr> Any</td> Has parameters</td> Tool</td></tr> </tbody></table> Metadata</h3> For every agent and agent method, we can attach a description and a prompt: @description("Increments the counter by the number provided in the `n` parameter") @prompt("Increment by a given number") async increment_by(n: number): Promise<number> { // ... }</code></pre>#[description("Increments the counter by the number provided in the `n` parameter")] #[prompt("Increment by a given number")] fn increment_by(&mut self, n: u32) -> u32;</code></pre>@description("Increments the counter by the number provided in the `n` parameter") @prompt("Increment by a given number") def incrementBy(n: Int): Future[Int]</code></pre>///| Increments the counter by the number provided in the `n` parameter #derive.prompt_hint("Increment by a given number") pub fn Counter::increment(self : Self, n: UInt32) -> UInt32 {</code></pre></div> Both are optional, and both are added to the MCP metadata. Special data types</h3> It's not strictly related to the MCP feature, and not even new in Golem 1.5, but the three special data types supported by all the Golem SDKs are a good match for exposing some special tools and resources through the MCP protocol: Unstructured text</h4> Any method parameter or return type can be defined as unstructured text. Optionally a set of allowed language codes can be attached to the type: myMethod( anyText: UnstructuredText, constrainedText: UnstructuredText<['en', 'de']> ) { // ... }</code></pre>#[derive(AllowedLanguages)] enum MyLangs { En, #[code("de")] German } fn my_method( &self, any_text: UnstructuredText, constrained_text: UnstructuredText<MyLangs> );</code></pre>def myMethod( anyText: TextSegment[AllowedLanguages.Any], constrainedText: TextSegment[MyLangs] ) sealed trait MyLangs object MyLangs { case object En extends MyLangs @golem.runtime.annotations.languageCode("de") case object German extends MyLangs implicit val allowed: AllowedLanguages[MyLangs] = golem.runtime.macros.AllowedLanguagesDerivation.derived } </code></pre>#derive.text_languages("constrained_text", "en", "de") pub fn MyAgent::my_method( self : Self, any_text : UnstructuredText, constrained_text : UnstructuredText, ) -> Unit { ... }</code></pre></div> Unstructured binary</h4> Similarly to unstructured text, we can also use unstructured binary parameters and return types, and optionally define the allowed MIME types for them: myMethod( anyBinary: UnstructuredBinary, image: UnstructuredBinary<['image/png', 'image/jpeg']> ) { // ... }</code></pre>#[derive(Debug, Clone, AllowedMimeTypes)] enum Image { #[mime_type("image/png")] Png, #[mime_type("image/jpeg")] Jpeg, } fn my_method( &self, any_binary: UnstructuredBinary, image: UnstructuredBinary<Image> );</code></pre>def myMethod( anyBinary: BinarySegment[AllowedMimeTypes.Any], image: BinarySegment[Image] ) sealed trait Image object Image { @golem.runtime.annotations.mimeType("image/png") case object Png extends Image @golem.runtime.annotations.mimeType("image/jpeg") case object Jpeg extends Image implicit val allowed: AllowedMimeTypes[Image] = golem.runtime.macros.AllowedMimeTypesDerivation.derived }</code></pre>#derive.mime_types("image", "image/png", "image/jpeg") pub fn MyAgent::my_method( self : Self, any_binary : UnstructuredBinary, image : UnstructuredBinary, ) -> Unit { ... }</code></pre></div> Multimodal</h4> Finally there is a special parameter type called multimodal, which is a special way to define methods (tools) that can work on multiple types of input. The default multimodal type just allows pasting either text or binary, but it is fully customizable with the above defined language and MIME type constraints, and can also include structured data. By using multimodal types, and not just modelling the same input using custom data types, Golem can map these definitions better to MCP concepts. The simplest version just accepts either an arbitrary text, or an arbitrary binary: textOrBinary(input: Multimodal) { ... }</code></pre>fn text_or_binary(&self, input: Multimodal) -> Multimodal { input }</code></pre>def textOrBinary(input: MultimodalItems.Basic): Future[MultimodalItems.Basic]</code></pre>pub fn MyAgent::text_or_binary( self : Self, input : @types.Multimodal[TextOrBinary], ) -> String { ... }</code></pre></div> We can add a third option in the form of a structured data type to this: type MyStructuredType = { ...} textOrBinaryOrStructured(input: MultimodalCustom<MyStructuredType>) { ... }</code></pre>#[derive(Schema)] struct MyStructuredType { /* ... */ } fn text_or_binary_or_structured( &self, input: MultimodalCustom<MyStructuredType>, ) -> MultimodalCustom<MyStructuredType> { input }</code></pre>final case class MyStructuredType(/* ... */) object MyStructuredType { implicit val schema: Schema[MyStructuredType] = Schema.derived } def textOrBinaryOrStructured( input: MultimodalItems.WithCustom[MyStructuredType] ): Future[MultimodalItems.WithCustom[MyStructuredType]]</code></pre>#derive.golem_schema pub(all) struct MyStructuredType { /* ... */ } pub fn MyAgent::text_or_binary_or_structured( self : Self, input : @types.Multimodal[CustomModality[MyStructuredType]], ) -> String { ... }</code></pre></div> Or we can fully customize the multimodal behavior by defining our own variant type it maps to: export type TextOrImage = | { tag: 'text'; val: UnstructuredText<['en', 'de'>] } | { tag: 'image'; val: UnstructuredBinary<['image/jpeg', 'image/png']> }; fullyCustom(input: MultimodalAdvanced<TextOrImage>) { ... }</code></pre>#[derive(AllowedLanguages)] enum TextLang { En, #[code("de")] German } #[derive(AllowedMimeTypes)] enum ImageType { #[mime_type("image/jpeg")] Jpeg, #[mime_type("image/png")] Png, } #[derive(Schema, MultimodalSchema)] enum TextOrImage { Text(UnstructuredText<TextLang>), Image(UnstructuredBinary<ImageType>), } fn fully_custom( &self, input: MultimodalAdvanced<TextOrImage>, ) -> MultimodalAdvanced<TextOrImage> { input }</code></pre>sealed trait TextLang object TextLang { @golem.runtime.annotations.languageCode("en") case object En extends TextLang @golem.runtime.annotations.languageCode("de") case object De extends TextLang implicit val allowed: AllowedLanguages[TextLang] = golem.runtime.macros.AllowedLanguagesDerivation.derived } sealed trait ImageType object ImageType { @golem.runtime.annotations.mimeType("image/jpeg") case object Jpeg extends ImageType @golem.runtime.annotations.mimeType("image/png") case object Png extends ImageType implicit val allowed: AllowedMimeTypes[ImageType] = golem.runtime.macros.AllowedMimeTypesDerivation.derived } final case class TextOrImage( text: TextSegment[TextLang], image: BinarySegment[ImageType], ) object TextOrImage { implicit val schema: GolemSchema[TextOrImage] = /* derived */ } def fullyCustom(input: Multimodal[TextOrImage]): Future[Multimodal[TextOrImage]]</code></pre>#derive.multimodal pub(all) enum TextOrImage { Text(UnstructuredText) Image(UnstructuredBinary) } #derive.text_languages("input.Text", "en", "de") #derive.mime_types("input.Image", "image/jpeg", "image/png") pub fn MyAgent::fully_custom( self : Self, input : @types.Multimodal[TextOrImage], ) -> String { ... }</code></pre></div> Remarks</h4> None of these special data types are MCP specific - using them in our agent code is not constraining them to be only called through MCP, they can still be invoked through agent-to-agent communication, mapped to HTTP APIs and so on.

vigoo's software development blog - mcp

Golem 1.5 features - Part 3: MCP

Security</h3> The above manifest section already exposes the listed agent through MCP, but it is not protected by any form of authentication. Similar to how we can protect HTTP endpoints with OAuth, we can attach a security scheme</strong> to our MCP deployment, let's call it mcp-oauth</code>:</p>

Special data types</h3> It's not strictly related to the MCP feature, and not even new in Golem 1.5</strong>, but the three special data types supported by all the Golem SDKs are a good match for exposing some special tools and resources through the MCP protocol:</p>

Remarks</h4> None of these special data types are MCP specific - using them in our agent code is not constraining them to be only called through MCP, they can still be invoked through agent-to-agent communication, mapped to HTTP APIs and so on.</p>

Special data types</h3>
It's not strictly related to the MCP feature, and not even new in Golem 1.5</strong>, but the three special data types supported by all the Golem SDKs are a good match for exposing some special tools and resources through the MCP protocol:</p>

Remarks</h4>
None of these special data types are MCP specific - using them in our agent code is not constraining them to be only called through MCP, they can still be invoked through agent-to-agent communication, mapped to HTTP APIs and so on.</p>