What causes "Anthropic does not support custom tool calls" in zed-industries/zed?

Use tools whose inputs are JSON (structured input schema) when Anthropic is the backend.

Anthropic does not support custom tool calls

Error message

Anthropic does not support custom tool calls

What it means

When converting a conversation into Anthropic Messages API requests, tool_use content must carry structured JSON input (Anthropic requires an 'input' object). LanguageModelToolUseInput::Text — free-form text produced by custom/untyped tools — has no faithful Anthropic encoding, so the conversion refuses it rather than sending a malformed request.

Source

Thrown at crates/anthropic/src/completion.rs:171

                Ok(None)
            }
        }
        MessageContent::Image(image) => Ok(Some(RequestContent::Image {
            source: ImageSource {
                source_type: "base64".to_string(),
                media_type: "image/png".to_string(),
                data: image.source.to_string(),
            },
            cache_control: None,
        })),
        MessageContent::ToolUse(tool_use) => match tool_use.input {
            LanguageModelToolUseInput::Json(input) => Ok(Some(RequestContent::ToolUse {
                id: tool_use.id.to_string(),
                name: tool_use.name.to_string(),
                input,
                cache_control: None,
            })),
            LanguageModelToolUseInput::Text(_) => Err(anyhow::anyhow!(
                "Anthropic does not support custom tool calls"
            )),
        },
        MessageContent::ToolResult(tool_result) => {
            let content = match tool_result.content.as_slice() {
                [LanguageModelToolResultContent::Text(text)] => {
                    ToolResultContent::Plain(text.to_string())
                }
                _ => {
                    let parts = tool_result
                        .content
                        .into_iter()
                        .map(|part| match part {
                            LanguageModelToolResultContent::Text(text) => ToolResultPart::Text {
                                text: text.to_string(),
                            },
                            LanguageModelToolResultContent::Image(image) => ToolResultPart::Image {
                                source: ImageSource {

View on GitHub (pinned to bc538def45)

Solutions

Use tools whose inputs are JSON (structured input schema) when Anthropic is the backend.
Start a new thread after switching to Anthropic instead of replaying old tool calls.
As a developer: convert Text input before serialization (e.g. wrap it as a JSON object) or skip incompatible turns with a warning.

Example fix

// before
LanguageModelToolUseInput::Text(_) => Err(anyhow::anyhow!(
    "Anthropic does not support custom tool calls"
)),
// after: wrap free-form text as JSON input
LanguageModelToolUseInput::Text(text) => Ok(Some(RequestContent::ToolUse {
    id: tool_use.id.to_string(),
    name: tool_use.name.to_string(),
    input: serde_json::json!({ "text": text.to_string() }),
    cache_control: None,
})),

Defensive patterns

Strategy: type-guard

Validate before calling

for message in &request.messages {
    for content in &message.content {
        if let MessageContent::ToolUse(tool_use) = content {
            if matches!(tool_use.input, LanguageModelToolUseInput::Text(_)) {
                // convert to JSON or reject before handing the request to Anthropic
            }
        }
    }
}

Type guard

fn is_anthropic_compatible(request: &LanguageModelRequest) -> bool {
    request
        .messages
        .iter()
        .flat_map(|message| message.content.iter())
        .all(|content| match content {
            MessageContent::ToolUse(tool_use) => {
                matches!(tool_use.input, LanguageModelToolUseInput::Json(_))
            }
            _ => true,
        })
}

Prevention

Define tool input schemas as JSON objects rather than free-form text.
Gate provider switching on compatibility, or start a fresh thread when replaying is impossible.
Convert Text tool inputs at the boundary where custom tools are invoked.

When it happens

Trigger: Replaying a thread that contains a tool call with text input into the Anthropic backend; switching models/providers mid-thread; custom tools that attach unstructured text as tool input.

Common situations: Switching the agent from another provider to Anthropic on an existing thread; custom tools without JSON schemas; imported transcripts or fixtures containing Text tool uses.

Related errors

AI-assisted analysis of zed-industries/zed@bc538def45 (2026-08-16). Data as JSON: /api/errors/d5c584e59bb35c9a. Report an issue: GitHub.