When an MCP server responds to a tools/call request, the result object has two fields: `content` (an array of typed blocks) and an optional `isError` boolean. The content array is how the tool delivers its output.
The 2026 spec defines four valid **content types**.
`text` is a plain string payload. Most tools that return descriptions, status messages, or query results use this type. It is the most common by far.
`image` is a base64-encoded image with a mimeType field (e.g., image/png, image/jpeg). A tool that generates a chart, takes a screenshot, or produces a diagram uses this. Multimodal models read the image as visual input on the next turn.
`audio` is a base64-encoded audio clip with a mimeType field (e.g., audio/wav). Added in the 2025 spec to support voice-capable agents. A tool that synthesizes speech or captures a recording uses this.
`resource` is a reference by URI to an MCP resource the host can fetch separately via resources/read. Instead of inlining a large payload, the server returns a pointer, and the host retrieves the data only when the model actually needs it.
A single tool result can mix types. A weather tool might return a text summary alongside an image of the radar map. The host receives the full array, decides which blocks to forward to the model, and includes them in the next inference request.
The type set is closed. Servers cannot invent custom types. Hosts can expect that every content block will be one of these four.
A tools/call result in MCP returns `{content: ContentBlock[], isError?: boolean}`. The content array carries the tool's output as a sequence of typed blocks. The 2026 spec defines four **content types**, and the set is a closed enumeration.
`text` is `{type: "text", text: "..."}`. A UTF-8 string payload used for descriptions, query results, code snippets, status messages, and anything else that fits as text. The most common type in production.
`image` is `{type: "image", data: "<base64>", mimeType: "image/png"}`. Image bytes encoded as base64 with an explicit MIME type. The host forwards this to multimodal models as native image input. Tools that generate charts, take screenshots, or render diagrams use this.
`audio` is `{type: "audio", data: "<base64>", mimeType: "audio/wav"}`. Audio bytes encoded as base64, added in the 2025 spec to support voice-capable agents. Tools that synthesize speech or return recordings use this. On text-only models, the host can transcribe or drop the block.
`resource` is `{type: "resource", resource: {uri: "...", mimeType: "...", text?: "..."}}`. A reference by URI to an MCP resource the host can fetch via resources/read. Instead of inlining large or contingent data, the server returns a pointer; the host fetches only what the model actually reads. A search tool returning twenty matches as twenty resource pointers saves orders of magnitude in tokens compared to inlining all twenty documents.
Mixed arrays are the norm, not an edge case. A browser navigation tool returns a text description of the page alongside a screenshot image. The host receives the array and forwards each block to the model in whatever format the provider expects.
The closed type set means hosts only need to handle four shapes. Servers that need a new content type must wait for a spec revision rather than inventing one unilaterally.
Quick reference
TL;DR
The 2026 spec defines four tools/call content types: text, image (base64 plus mimeType), audio (base64 plus mimeType, added 2025), and resource (URI reference). Mixed arrays are allowed; the set is closed.
Think of a tool result as a small package that can hold different kinds of items. MCP gives tools four kinds of boxes to put things in. A text box holds written messages. An image box holds a picture encoded as data. An audio box holds a sound clip. A resource box holds a card that says 'the actual thing lives at this address, go fetch it when you need it.' A single package can contain several boxes at once, like a written summary next to a chart picture, and the host figures out how to show each piece to the model. Servers cannot invent new box types on their own.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Encoding images or audio as strings inside a text block. The host cannot route the payload to the model's multimodal input path. Use the typed image or audio block so the host knows the data format.
Returning large payloads inline when the model may not read them. Inlining costs tokens and bandwidth on every call. Return a resource reference and let the host fetch only when the model asks.
Inventing a custom content type for a special payload format. Hosts will not handle unknown types reliably. Use the closest existing type or advocate for a spec extension.
Quick scan · 2–4 min
Everything important, quickly.
4 min: enumerate the four content types with their shapes, note that audio was added in 2025, explain why resource exists (lazy fetch), give one example of a mixed-type result, and close with the isError vs JSON-RPC error distinction.
Real-world examples
Real products, models, and research that use this idea.
1A browser automation server returns a text description of the loaded page alongside an image screenshot; multimodal models read the screenshot as visual input to decide what to click next.
2A speech synthesis server returns an audio block with a wav payload and a text block describing the utterance; voice-capable hosts route the audio to the user, text-only hosts display the description.
3A filesystem search server returns text matches plus per-file resource blocks pointing to the full file content; the host fetches a resource only when the model decides to inspect that specific file.
4MCP Inspector surfaces each content block separately with its type label so server authors can verify typed output rather than discovering everything was accidentally stringified into text.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QWhy does MCP define a resource content type instead of always inlining data in text or image blocks?
Sign in or go Pro to see the approach for this follow-up.
QWhat should a host do when a server returns an unrecognized content type?
Sign in or go Pro to see the approach for this follow-up.
QWhen is result-level isError more useful than a JSON-RPC error response?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Assuming every tool result is plain text. MCP tools can return images, audio, and resource references in the same array, and a single call can mix types freely.