feat: CLI --research flag + MCP cloud fallback + structured research output

- --research "query": deep research via cloud API, saves JSON file with
  report + sources + findings, prints report to stdout
- --deep: longer, more thorough research mode
- MCP extract/summarize: cloud fallback when no local LLM available
- MCP research: returns structured JSON instead of raw text
- Bump to v0.3.7

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

CHANGELOG.md

@@ -3,6 +3,15 @@
 All notable changes to webclaw are documented here.
 Format follows [Keep a Changelog](https://keepachangelog.com/).
 
+## [0.3.7] — 2026-04-03
+
+### Added
+- **`--research` CLI flag**: run deep research via the cloud API. Prints report to stdout and saves full result (report + sources + findings) to a JSON file. Supports `--deep` for longer reports.
+- **MCP extract/summarize cloud fallback**: when no local LLM is available, these tools now fall back to the cloud API instead of erroring. Set `WEBCLAW_API_KEY` for automatic fallback.
+- **MCP research structured output**: the research tool now returns structured JSON (report + sources + findings + metadata) instead of raw text, so agents can reference individual findings and source URLs.
+
+---
+
 ## [0.3.6] — 2026-04-02
 
 ### Added

Cargo.lock

@@ -3055,7 +3055,7 @@ dependencies = [
 
 [[package]]
 name = "webclaw-cli"
-version = "0.3.6"
+version = "0.3.7"
 dependencies = [
  "clap",
  "dotenvy",
@@ -3075,7 +3075,7 @@ dependencies = [
 
 [[package]]
 name = "webclaw-core"
-version = "0.3.6"
+version = "0.3.7"
 dependencies = [
  "ego-tree",
  "once_cell",
@@ -3093,7 +3093,7 @@ dependencies = [
 
 [[package]]
 name = "webclaw-fetch"
-version = "0.3.6"
+version = "0.3.7"
 dependencies = [
  "bytes",
  "calamine",
@@ -3115,7 +3115,7 @@ dependencies = [
 
 [[package]]
 name = "webclaw-llm"
-version = "0.3.6"
+version = "0.3.7"
 dependencies = [
  "async-trait",
  "reqwest",
@@ -3128,7 +3128,7 @@ dependencies = [
 
 [[package]]
 name = "webclaw-mcp"
-version = "0.3.6"
+version = "0.3.7"
 dependencies = [
  "dotenvy",
  "reqwest",
@@ -3148,7 +3148,7 @@ dependencies = [
 
 [[package]]
 name = "webclaw-pdf"
-version = "0.3.6"
+version = "0.3.7"
 dependencies = [
  "pdf-extract",
  "thiserror",

Cargo.toml

@@ -3,7 +3,7 @@ resolver = "2"
 members = ["crates/*"]
 
 [workspace.package]
-version = "0.3.6"
+version = "0.3.7"
 edition = "2024"
 license = "AGPL-3.0"
 repository = "https://github.com/0xMassi/webclaw"

crates/webclaw-cli/src/main.rs

@@ -268,6 +268,15 @@ struct Cli {
     #[arg(long)]
     cloud: bool,
 
+    /// Run deep research on a topic via the cloud API. Requires --api-key.
+    /// Saves full result (report + sources + findings) to a JSON file.
+    #[arg(long)]
+    research: Option<String>,
+
+    /// Enable deep research mode (longer, more thorough report). Used with --research.
+    #[arg(long)]
+    deep: bool,
+
     /// Output directory: save each page to a separate file instead of stdout.
     /// Works with --crawl, batch (multiple URLs), and single URL mode.
     /// Filenames are derived from URL paths (e.g. /docs/api -> docs/api.md).
@@ -2067,6 +2076,141 @@ fn has_llm_flags(cli: &Cli) -> bool {
     cli.extract_json.is_some() || cli.extract_prompt.is_some() || cli.summarize.is_some()
 }
 
+async fn run_research(cli: &Cli, query: &str) -> Result<(), String> {
+    let api_key = cli
+        .api_key
+        .as_deref()
+        .ok_or("--research requires WEBCLAW_API_KEY (set via env or --api-key)")?;
+
+    let client = reqwest::Client::builder()
+        .timeout(std::time::Duration::from_secs(600))
+        .build()
+        .map_err(|e| format!("http client error: {e}"))?;
+
+    let mut body = serde_json::json!({ "query": query });
+    if cli.deep {
+        body["deep"] = serde_json::json!(true);
+    }
+
+    eprintln!("Starting research: {query}");
+    if cli.deep {
+        eprintln!("Deep mode enabled (longer, more thorough)");
+    }
+
+    // Start job
+    let resp = client
+        .post("https://api.webclaw.io/v1/research")
+        .header("Authorization", format!("Bearer {api_key}"))
+        .json(&body)
+        .send()
+        .await
+        .map_err(|e| format!("API error: {e}"))?
+        .json::<serde_json::Value>()
+        .await
+        .map_err(|e| format!("parse error: {e}"))?;
+
+    let job_id = resp
+        .get("id")
+        .and_then(|v| v.as_str())
+        .ok_or("API did not return a job ID")?
+        .to_string();
+
+    eprintln!("Job started: {job_id}");
+
+    // Poll
+    for poll in 0..200 {
+        tokio::time::sleep(std::time::Duration::from_secs(3)).await;
+
+        let status_resp = client
+            .get(format!("https://api.webclaw.io/v1/research/{job_id}"))
+            .header("Authorization", format!("Bearer {api_key}"))
+            .send()
+            .await
+            .map_err(|e| format!("poll error: {e}"))?
+            .json::<serde_json::Value>()
+            .await
+            .map_err(|e| format!("parse error: {e}"))?;
+
+        let status = status_resp
+            .get("status")
+            .and_then(|v| v.as_str())
+            .unwrap_or("unknown");
+
+        match status {
+            "completed" => {
+                let report = status_resp
+                    .get("report")
+                    .and_then(|v| v.as_str())
+                    .unwrap_or("");
+
+                // Save full result to JSON file
+                let slug: String = query
+                    .chars()
+                    .map(|c| {
+                        if c.is_alphanumeric() || c == ' ' {
+                            c
+                        } else {
+                            ' '
+                        }
+                    })
+                    .collect::<String>()
+                    .split_whitespace()
+                    .collect::<Vec<_>>()
+                    .join("-")
+                    .to_lowercase();
+                let slug = if slug.len() > 50 { &slug[..50] } else { &slug };
+                let filename = format!("research-{slug}.json");
+
+                let json = serde_json::to_string_pretty(&status_resp).unwrap_or_default();
+                std::fs::write(&filename, &json)
+                    .map_err(|e| format!("failed to write {filename}: {e}"))?;
+
+                let elapsed = status_resp
+                    .get("elapsed_ms")
+                    .and_then(|v| v.as_i64())
+                    .unwrap_or(0);
+                let sources = status_resp
+                    .get("sources_count")
+                    .and_then(|v| v.as_i64())
+                    .unwrap_or(0);
+                let findings = status_resp
+                    .get("findings_count")
+                    .and_then(|v| v.as_i64())
+                    .unwrap_or(0);
+
+                eprintln!(
+                    "Research complete: {sources} sources, {findings} findings, {:.1}s",
+                    elapsed as f64 / 1000.0
+                );
+                eprintln!("Saved to: {filename}");
+
+                // Print report to stdout
+                if !report.is_empty() {
+                    println!("{report}");
+                }
+
+                return Ok(());
+            }
+            "failed" => {
+                let error = status_resp
+                    .get("error")
+                    .and_then(|v| v.as_str())
+                    .unwrap_or("unknown error");
+                return Err(format!("Research failed: {error}"));
+            }
+            _ => {
+                if poll % 10 == 9 {
+                    eprintln!("Still researching... ({:.0}s)", (poll + 1) as f64 * 3.0);
+                }
+            }
+        }
+    }
+
+    Err(format!(
+        "Research timed out after ~10 minutes. Check status: GET /v1/research/{job_id}"
+    ))
+}
+
 #[tokio::main]
 async fn main() {
     dotenvy::dotenv().ok();
@@ -2126,6 +2270,15 @@ async fn main() {
         return;
     }
 
+    // --research: deep research via cloud API
+    if let Some(ref query) = cli.research {
+        if let Err(e) = run_research(&cli, query).await {
+            eprintln!("error: {e}");
+            process::exit(1);
+        }
+        return;
+    }
+
     // Collect all URLs from args + --urls-file
     let entries = match collect_urls(&cli) {
         Ok(u) => u,

crates/webclaw-mcp/src/server.rs

@@ -319,34 +319,46 @@ impl WebclawMcp {
     }
 
     /// Extract structured data from a web page using an LLM. Provide either a JSON schema or a natural language prompt.
-    /// Automatically falls back to the webclaw cloud API when bot protection is detected.
+    /// Falls back to the webclaw cloud API when no local LLM is available or bot protection is detected.
     #[tool]
     async fn extract(
         &self,
         Parameters(params): Parameters<ExtractParams>,
     ) -> Result<String, String> {
         validate_url(&params.url)?;
-        let chain = self.llm_chain.as_ref().ok_or(
-            "No LLM providers available. Set OPENAI_API_KEY or ANTHROPIC_API_KEY, or run Ollama locally.",
-        )?;
 
         if params.schema.is_none() && params.prompt.is_none() {
             return Err("Either 'schema' or 'prompt' is required for extraction.".into());
         }
 
-        // For extract, if we get a cloud fallback we call the cloud extract endpoint directly
+        // No local LLM — fall back to cloud API directly
+        if self.llm_chain.is_none() {
+            let cloud = self.cloud.as_ref().ok_or(
+                "No LLM providers available. Set OPENAI_API_KEY, ANTHROPIC_API_KEY, or WEBCLAW_API_KEY for cloud fallback.",
+            )?;
+            let mut body = json!({"url": params.url});
+            if let Some(ref schema) = params.schema {
+                body["schema"] = json!(schema);
+            }
+            if let Some(ref prompt) = params.prompt {
+                body["prompt"] = json!(prompt);
+            }
+            let resp = cloud.post("extract", body).await?;
+            return Ok(serde_json::to_string_pretty(&resp).unwrap_or_default());
+        }
+
+        let chain = self.llm_chain.as_ref().unwrap();
+
         let llm_content = match self.smart_fetch_llm(&params.url).await? {
             SmartFetchResult::Local(extraction) => {
                 webclaw_core::to_llm_text(&extraction, Some(&params.url))
             }
-            SmartFetchResult::Cloud(resp) => {
-                // Use the LLM format from cloud, fall back to markdown
-                resp.get("llm")
-                    .or_else(|| resp.get("markdown"))
-                    .and_then(|v| v.as_str())
-                    .unwrap_or("")
-                    .to_string()
-            }
+            SmartFetchResult::Cloud(resp) => resp
+                .get("llm")
+                .or_else(|| resp.get("markdown"))
+                .and_then(|v| v.as_str())
+                .unwrap_or("")
+                .to_string(),
         };
 
         let data = if let Some(ref schema) = params.schema {
@@ -364,16 +376,32 @@ impl WebclawMcp {
     }
 
     /// Summarize the content of a web page using an LLM.
-    /// Automatically falls back to the webclaw cloud API when bot protection is detected.
+    /// Falls back to the webclaw cloud API when no local LLM is available or bot protection is detected.
     #[tool]
     async fn summarize(
         &self,
         Parameters(params): Parameters<SummarizeParams>,
     ) -> Result<String, String> {
         validate_url(&params.url)?;
-        let chain = self.llm_chain.as_ref().ok_or(
-            "No LLM providers available. Set OPENAI_API_KEY or ANTHROPIC_API_KEY, or run Ollama locally.",
-        )?;
+
+        // No local LLM — fall back to cloud API directly
+        if self.llm_chain.is_none() {
+            let cloud = self.cloud.as_ref().ok_or(
+                "No LLM providers available. Set OPENAI_API_KEY, ANTHROPIC_API_KEY, or WEBCLAW_API_KEY for cloud fallback.",
+            )?;
+            let mut body = json!({"url": params.url});
+            if let Some(sentences) = params.max_sentences {
+                body["max_sentences"] = json!(sentences);
+            }
+            let resp = cloud.post("summarize", body).await?;
+            let summary = resp.get("summary").and_then(|v| v.as_str()).unwrap_or("");
+            if summary.is_empty() {
+                return Ok(serde_json::to_string_pretty(&resp).unwrap_or_default());
+            }
+            return Ok(summary.to_string());
+        }
+
+        let chain = self.llm_chain.as_ref().unwrap();
 
         let llm_content = match self.smart_fetch_llm(&params.url).await? {
             SmartFetchResult::Local(extraction) => {
@@ -535,15 +563,32 @@ impl WebclawMcp {
 
             match status {
                 "completed" => {
-                    let report = status_resp
-                        .get("report")
-                        .and_then(|v| v.as_str())
-                        .unwrap_or("");
+                    // Return structured result: report + sources + findings
+                    let mut result = json!({
+                        "id": job_id,
+                        "status": "completed",
+                    });
 
-                    if report.is_empty() {
-                        return Ok(serde_json::to_string_pretty(&status_resp).unwrap_or_default());
+                    if let Some(report) = status_resp.get("report") {
+                        result["report"] = report.clone();
                     }
-                    return Ok(report.to_string());
+                    if let Some(sources) = status_resp.get("sources") {
+                        result["sources"] = sources.clone();
+                    }
+                    if let Some(findings) = status_resp.get("findings") {
+                        result["findings"] = findings.clone();
+                    }
+                    if let Some(elapsed) = status_resp.get("elapsed_ms") {
+                        result["elapsed_ms"] = elapsed.clone();
+                    }
+                    if let Some(sc) = status_resp.get("sources_count") {
+                        result["sources_count"] = sc.clone();
+                    }
+                    if let Some(fc) = status_resp.get("findings_count") {
+                        result["findings_count"] = fc.clone();
+                    }
+
+                    return Ok(serde_json::to_string_pretty(&result).unwrap_or_default());
                 }
                 "failed" => {
                     let error = status_resp