[Bug] MCP structuredContent-only tool results leave ToolResult.content empty — agent loop sees no output

[Sebastiao-Assuncao]

Problem

Some production MCP servers (e.g. Brex at https://api.brex.com/mcp) return successful tool results where the payload lives entirely in structuredContent, with an empty content array:

{
  "content": [],
  "structuredContent": {
    "id": "cuuser_...",
    "first_name": "...",
    "email": "...",
    "role": "EMPLOYEE",
    "status": "ACTIVE"
  },
  "isError": false
}

The MCP call succeeds (isError: false), but agents behave as if the tool returned nothing: empty tool messages to the model, empty observability spans, and user-facing “no data” responses despite a valid structured payload.

---

What seems fixed already (thank you)

Follow-ups to #7686 landed in #7715 and #8580: structuredContent is no longer discarded and is preserved on ToolResult (via metadata["structured_content"] in recent releases).

That addresses data loss on the ToolResult object — e.g. tool hooks or direct inspection can read the structured payload.

---

What still appears broken

Even on releases that preserve metadata["structured_content"], ToolResult.content stays empty when content: [].

The model / run pipeline still appears to read only .content when building tool messages (e.g. in models/base.py: function_call_output = tool_result.content).

So for structuredContent-only responses:

Field	Observed
ToolResult.metadata["structured_content"]	Populated (2.6.20+)
ToolResult.content	""
Tool message fed back to the LLM	Empty
Typical tracing (tool output string)	Empty

On older Agno versions (e.g. 2.6.9), structuredContent is not preserved at all — only the empty .content path runs.

PR #8580 notes that metadata fields are “not propagated into the run/stream/message pipeline” — which matches the gap above for agents that rely on the standard tool-result → message flow.

---

Use case that would benefit from a fix

Vendor MCP integrations for business APIs (expense management, banking, billing, etc.) where the server exposes typed structured objects rather than plain text blocks in content.

Concrete example: calling get_user_myself on Brex MCP returns the authenticated user as structured JSON. An agent should be able to answer “who am I in Brex?” or chain follow-up tool calls using that identity — without the developer re-implementing MCP result adapters in every app.

The same pattern likely applies to other MCP servers adopting structured output (Stripe and similar), and to MCP Apps / AG-UI flows where structured data must reach either the LLM or the client through the normal Agno tool path.

---

Expected behavior (for discussion)

When isError == false, content is empty, and structuredContent is present, Agno should make the payload visible through the primary agent loop, not only via direct ToolResult.metadata access. Possible directions (Agno’s call):

1. Serialize structuredContent into ToolResult.content when content is empty (e.g. JSON string), and/or
2. Fall back to metadata["structured_content"] in the model tool-message path when .content is blank.

When both content and structuredContent are present, existing behavior should remain (text in .content, structured data additionally in metadata).

---

Temporary workaround on our side (not proposed as upstream design)

We added a small integration-layer shim: a forwarding proxy around the MCP session’s call_tool that injects a synthetic TextContent JSON block when content is empty and structuredContent is present, then feed that session into Agno’s existing get_entrypoint_for_tool.

It unblocks our production path but feels like something that belongs in Agno’s MCP adapter and/or message pipeline rather than in every consumer. We’re reporting this so upstream can choose the right long-term behavior.

---

Environment

• Agno: 2.6.9 (also verified gap on 2.6.20+ for LLM-visible output despite metadata preservation)
• MCP Python SDK: 1.28.x, streamable HTTP transport
• MCP server: Brex (https://api.brex.com/mcp), OAuth Bearer auth

Happy to provide a minimal repro script or additional traces if useful.

[Success6666]

Hi @Sebastiao-Assuncao, thanks for the detailed report. I opened #8654 to address this. It serializes structuredContent into ToolResult.content when the content array is empty, so structured-only MCP results are visible to the standard agent loop.

[Success6666]

Quick update: I opened #8695 with the refreshed branch for this fix. It keeps the structured-only MCP result behavior and updates the regression test to compare the parsed structured payload rather than the exact JSON string representation.