feat: support MCP tool hints mapped to ToolAnnotations

.agents/skills/naftiko-capability/SKILL.md

@@ -111,7 +111,9 @@ Specification directly.
 9. `ForwardConfig` requires `targetNamespace` (single string, not array)
    and `trustedHeaders` (at least one entry).
 10. MCP tools must have `name` and `description`. MCP tool input parameters
-    must have `name`, `type`, and `description`.
+    must have `name`, `type`, and `description`. Tools may declare optional
+    `hints` (readOnly, destructive, idempotent, openWorld) — these map to
+    MCP `ToolAnnotations` on the wire.
 11. ExposedOperation supports exactly two modes (oneOf): simple (`call` +
     optional `with`) or orchestrated (`steps` + optional `mappings`). Never
     mix fields from both modes.

.agents/skills/naftiko-capability/references/design-guidelines.md

@@ -98,6 +98,11 @@ Avoid:
 - Use tools for actions and resources for read-only data access.
 - Prefer small tools with crisp, typed `inputParameters`.
 - If an MCP tool becomes complex, switch to orchestration and document it clearly.
+- Use `hints` to signal tool behavior to clients:
+    - Set `readOnly: true` for tools that only read data (GET-like).
+    - Set `destructive: true` for tools that delete or overwrite (DELETE, PUT).
+    - Set `idempotent: true` for tools safe to retry.
+    - Set `openWorld: true` for tools calling external APIs; `false` for closed-domain tools (local data, caches).
 
 ## Orchestration guidelines (steps + mappings)
 

.agents/skills/naftiko-capability/references/wrap-api-as-mcp.md

@@ -96,15 +96,20 @@ For each MCP tool:
 
 1. `name` (kebab-case / IdentifierKebab) is required and must be stable (used as the MCP tool name).
 2. `description` is required (agent discovery depends on it).
-3. If tool is simple:
+3. `hints` is optional — declares behavioral hints mapped to MCP `ToolAnnotations`:
+    - `readOnly` (bool) — tool does not modify its environment (default: false)
+    - `destructive` (bool) — tool may perform destructive updates (default: true, meaningful only when readOnly is false)
+    - `idempotent` (bool) — repeating the call has no additional effect (default: false, meaningful only when readOnly is false)
+    - `openWorld` (bool) — tool interacts with external entities (default: true)
+4. If tool is simple:
     - must define `call: {namespace}.{operationName}`
     - may define `with`
     - should define `outputParameters` (typed) when you want structured results.
-4. If tool is orchestrated:
+5. If tool is orchestrated:
     - must define `steps` (min 1), each step has `name`
     - may define `mappings`
     - `outputParameters` must use orchestrated output parameter objects (named + typed)
-5. Tool `inputParameters`:
+6. Tool `inputParameters`:
     - each parameter must have `name`, `type`, `description`
     - set `required: false` explicitly for optional params (default is true)
 

src/main/java/io/naftiko/engine/exposes/mcp/McpServerAdapter.java

@@ -28,6 +28,7 @@
 import io.naftiko.spec.InputParameterSpec;
 import io.naftiko.spec.exposes.McpServerSpec;
 import io.naftiko.spec.exposes.McpServerToolSpec;
+import io.naftiko.spec.exposes.McpToolHintsSpec;
 
 /**
  * MCP Server Adapter implementation.
@@ -149,8 +150,31 @@ private McpSchema.Tool buildMcpTool(McpServerToolSpec toolSpec) {
                 schemaProperties.isEmpty() ? null : schemaProperties,
                 required.isEmpty() ? null : required, null, null, null);
 
+        // Build ToolAnnotations from spec hints and label
+        McpSchema.ToolAnnotations annotations = buildToolAnnotations(toolSpec);
+
         return McpSchema.Tool.builder().name(toolSpec.getName())
-                .description(toolSpec.getDescription()).inputSchema(inputSchema).build();
+                .description(toolSpec.getDescription()).inputSchema(inputSchema)
+                .annotations(annotations).build();
+    }
+
+    /**
+     * Build MCP ToolAnnotations from the tool spec's hints and label. Returns null if neither hints
+     * nor label are present.
+     */
+    McpSchema.ToolAnnotations buildToolAnnotations(McpServerToolSpec toolSpec) {
+        McpToolHintsSpec hints = toolSpec.getHints();
+        String label = toolSpec.getLabel();
+
+        if (hints == null && label == null) {
+            return null;
+        }
+
+        return new McpSchema.ToolAnnotations(label,
+                hints != null ? hints.getReadOnly() : null,
+                hints != null ? hints.getDestructive() : null,
+                hints != null ? hints.getIdempotent() : null,
+                hints != null ? hints.getOpenWorld() : null, null);
     }
 
     public McpServerSpec getMcpServerSpec() {

src/main/java/io/naftiko/engine/exposes/mcp/ProtocolDispatcher.java

@@ -165,6 +165,29 @@ private ObjectNode handleToolsList(JsonNode id) {
                 toolNode.set("inputSchema", mapper.valueToTree(tool.inputSchema()));
             }
 
+            if (tool.annotations() != null) {
+                ObjectNode annotationsNode = mapper.createObjectNode();
+                McpSchema.ToolAnnotations ann = tool.annotations();
+                if (ann.title() != null) {
+                    annotationsNode.put("title", ann.title());
+                }
+                if (ann.readOnlyHint() != null) {
+                    annotationsNode.put("readOnlyHint", ann.readOnlyHint());
+                }
+                if (ann.destructiveHint() != null) {
+                    annotationsNode.put("destructiveHint", ann.destructiveHint());
+                }
+                if (ann.idempotentHint() != null) {
+                    annotationsNode.put("idempotentHint", ann.idempotentHint());
+                }
+                if (ann.openWorldHint() != null) {
+                    annotationsNode.put("openWorldHint", ann.openWorldHint());
+                }
+                if (!annotationsNode.isEmpty()) {
+                    toolNode.set("annotations", annotationsNode);
+                }
+            }
+
             toolsArray.add(toolNode);
         }
 

src/main/java/io/naftiko/spec/exposes/McpServerToolSpec.java

@@ -48,6 +48,9 @@ public class McpServerToolSpec {
     @JsonInclude(JsonInclude.Include.NON_EMPTY)
     private final List<OperationStepSpec> steps;
 
+    @JsonInclude(JsonInclude.Include.NON_NULL)
+    private volatile McpToolHintsSpec hints;
+
     @JsonInclude(JsonInclude.Include.NON_EMPTY)
     private final List<StepOutputMappingSpec> mappings;
 
@@ -124,4 +127,12 @@ public List<OutputParameterSpec> getOutputParameters() {
         return outputParameters;
     }
 
+    public McpToolHintsSpec getHints() {
+        return hints;
+    }
+
+    public void setHints(McpToolHintsSpec hints) {
+        this.hints = hints;
+    }
+
 }

src/main/java/io/naftiko/spec/exposes/McpToolHintsSpec.java

@@ -0,0 +1,71 @@
+/**
+ * Copyright 2025-2026 Naftiko
+ * 
+ * Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except
+ * in compliance with the License. You may obtain a copy of the License at
+ * 
+ * http://www.apache.org/licenses/LICENSE-2.0
+ * 
+ * Unless required by applicable law or agreed to in writing, software distributed under the License
+ * is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express
+ * or implied. See the License for the specific language governing permissions and limitations under
+ * the License.
+ */
+package io.naftiko.spec.exposes;
+
+/**
+ * MCP Tool Hints Specification Element.
+ * 
+ * Optional hints describing tool behavior to MCP clients. All properties are advisory. Mapped to
+ * ToolAnnotations in the MCP protocol.
+ */
+public class McpToolHintsSpec {
+
+    private Boolean readOnly;
+    private Boolean destructive;
+    private Boolean idempotent;
+    private Boolean openWorld;
+
+    public McpToolHintsSpec() {}
+
+    public McpToolHintsSpec(Boolean readOnly, Boolean destructive, Boolean idempotent,
+            Boolean openWorld) {
+        this.readOnly = readOnly;
+        this.destructive = destructive;
+        this.idempotent = idempotent;
+        this.openWorld = openWorld;
+    }
+
+    public Boolean getReadOnly() {
+        return readOnly;
+    }
+
+    public void setReadOnly(Boolean readOnly) {
+        this.readOnly = readOnly;
+    }
+
+    public Boolean getDestructive() {
+        return destructive;
+    }
+
+    public void setDestructive(Boolean destructive) {
+        this.destructive = destructive;
+    }
+
+    public Boolean getIdempotent() {
+        return idempotent;
+    }
+
+    public void setIdempotent(Boolean idempotent) {
+        this.idempotent = idempotent;
+    }
+
+    public Boolean getOpenWorld() {
+        return openWorld;
+    }
+
+    public void setOpenWorld(Boolean openWorld) {
+        this.openWorld = openWorld;
+    }
+
+}

src/main/resources/schemas/examples/skill-adapter.yml

@@ -47,6 +47,9 @@ capability:
       tools:
         - name: get-current-weather
           description: "Retrieve current weather conditions for a location"
+          hints:
+            readOnly: true
+            openWorld: true
           inputParameters:
             - name: location
               type: string

src/main/resources/schemas/naftiko-schema.json

@@ -895,6 +895,9 @@
             "$ref": "#/$defs/StepOutputMapping"
           }
         },
+        "hints": {
+          "$ref": "#/$defs/McpToolHints"
+        },
         "outputParameters": {
           "type": "array"
         }
@@ -936,6 +939,29 @@
       ],
       "additionalProperties": false
     },
+    "McpToolHints": {
+      "type": "object",
+      "description": "Optional hints describing tool behavior to MCP clients. All properties are advisory. Mapped to ToolAnnotations in the MCP protocol.",
+      "properties": {
+        "readOnly": {
+          "type": "boolean",
+          "description": "If true, the tool does not modify its environment. Default: false."
+        },
+        "destructive": {
+          "type": "boolean",
+          "description": "If true, the tool may perform destructive updates. Meaningful only when readOnly is false. Default: true."
+        },
+        "idempotent": {
+          "type": "boolean",
+          "description": "If true, calling the tool repeatedly with the same arguments has no additional effect. Meaningful only when readOnly is false. Default: false."
+        },
+        "openWorld": {
+          "type": "boolean",
+          "description": "If true, the tool may interact with external entities. Default: true."
+        }
+      },
+      "additionalProperties": false
+    },
     "McpToolInputParameter": {
       "type": "object",
       "description": "Declares an input parameter for an MCP tool. These become properties in the tool's JSON Schema input definition.",

src/main/resources/wiki/FAQ.md

@@ -808,9 +808,10 @@ Check the naftiko field in your YAML to specify the version.
 
 1. **Use `type: mcp`** in `exposes`
 2. **Define `tools`** - each tool is an MCP tool your capability provides
-3. **Use stdio transport** - for native Claude Desktop integration
-4. **Test with Claude** - configure Claude Desktop with your MCP server
-5. **Publish** - share your capability spec with the community
+3. **Add `hints`** (optional) - declare behavioral hints like `readOnly`, `destructive`, `idempotent`, `openWorld` to help clients understand tool safety
+4. **Use stdio transport** - for native Claude Desktop integration
+5. **Test with Claude** - configure Claude Desktop with your MCP server
+6. **Publish** - share your capability spec with the community
 
 See [Tutorial - Part 1](https://github.com/naftiko/framework/wiki/Tutorial-MCP-Part-1) for a full MCP example, then continue with [Tutorial - Part 2](https://github.com/naftiko/framework/wiki/Tutorial-MCP-Part-2) for Skill and REST exposure.
 

src/main/resources/wiki/Specification-‐-Schema.md

@@ -343,6 +343,7 @@ An MCP tool definition. Each tool maps to one or more consumed HTTP operations,
 | **name** | `string` | **REQUIRED**. Technical name for the tool. Used as the MCP tool name. MUST match pattern `^[a-zA-Z0-9-]+$`. |
 | **label** | `string` | Human-readable display name for the tool. Mapped to MCP `title` in protocol responses. |
 | **description** | `string` | **REQUIRED**. A meaningful description of the tool. Essential for agent discovery. |
+| **hints** | `McpToolHints` | Optional behavioral hints for MCP clients. Mapped to `ToolAnnotations` in the MCP protocol. See [3.5.5.1 McpToolHints Object](#3551-mctoolhints-object). |
 | **inputParameters** | `McpToolInputParameter[]` | Tool input parameters. These become the MCP tool's input schema (JSON Schema). |
 | **call** | `string` | **Simple mode only**. Reference to a consumed operation. Format: `{namespace}.{operationId}`. MUST match pattern `^[a-zA-Z0-9-]+\.[a-zA-Z0-9-]+$`. |
 | **with** | `WithInjector` | **Simple mode only**. Parameter injection for the called operation. |
@@ -376,6 +377,37 @@ An MCP tool definition. Each tool maps to one or more consumed HTTP operations,
 - Input parameters are accessed via namespace-qualified references of the form `{mcpNamespace}.{paramName}`.
 - No additional properties are allowed.
 
+#### 3.5.5.1 McpToolHints Object
+
+Optional behavioral hints describing a tool to MCP clients. All properties are advisory — clients SHOULD NOT make trust decisions based on these values from untrusted servers. Mapped to `ToolAnnotations` in the MCP protocol wire format.
+
+**Fixed Fields:**
+
+| Field Name | Type | Description |
+| --- | --- | --- |
+| **readOnly** | `boolean` | If `true`, the tool does not modify its environment. Default: `false`. |
+| **destructive** | `boolean` | If `true`, the tool may perform destructive updates. Meaningful only when `readOnly` is `false`. Default: `true`. |
+| **idempotent** | `boolean` | If `true`, calling the tool repeatedly with the same arguments has no additional effect. Meaningful only when `readOnly` is `false`. Default: `false`. |
+| **openWorld** | `boolean` | If `true`, the tool may interact with external entities (e.g. web APIs). If `false`, the tool's domain is closed (e.g. local data). Default: `true`. |
+
+**Rules:**
+
+- All fields are optional. Omitted fields fall back to their defaults.
+- `destructive` and `idempotent` are only meaningful when `readOnly` is `false`.
+- No additional properties are allowed.
+
+**McpToolHints Example:**
+
+```yaml
+tools:
+  - name: get-current-weather
+    description: "Retrieve current weather conditions"
+    hints:
+      readOnly: true
+      openWorld: true
+    call: weather-api.get-current
+```
+
 #### 3.5.6 McpToolInputParameter Object
 
 Declares an input parameter for an MCP tool. These become properties in the tool's JSON Schema input definition.

src/main/resources/wiki/Tutorial-‐-Part-1.md

@@ -317,6 +317,22 @@ The response gets shaped too — flat fields like `departurePort`/`arrivalPort`
 
 The agent went from observer to operator.
 
+> **💡 Tip: Tool hints.** Now that you have both read-only tools (`list-ships`, `get-ship`) and write tools (`create-voyage`), you can declare behavioral hints that help clients distinguish them:
+> ```yaml
+> - name: list-ships
+>   hints:
+>     readOnly: true
+>   # ...
+> - name: create-voyage
+>   hints:
+>     readOnly: false
+>     destructive: false
+>     idempotent: false
+>     openWorld: true
+>   # ...
+> ```
+> Hints map to MCP `ToolAnnotations` and are advisory — clients use them to decide which tools need confirmation, can be retried safely, etc. See [McpToolHints](https://github.com/naftiko/framework/wiki/Specification-Schema#3551-mctoolhints-object) in the spec.
+
 **What you learned:** `POST` operations, `body` template, array-type inputs, write tools.
 
 ---