diff --git a/.env.example b/.env.example
index 8558e9c4..9ad0e864 100644
--- a/.env.example
+++ b/.env.example
@@ -46,10 +46,18 @@ MAX_WORKERS=30
 # API Keys and External Services
 # =============================================================================
 
+# Search backend selector. Set to "exa" to route web/scholar search through Exa,
+# or leave as "serper" (default) to use Serper. Existing deployments are unaffected.
+SEARCH_PROVIDER=serper
+
 # Serper API for web search and Google Scholar
 # Get your key from: https://serper.dev/
 SERPER_KEY_ID=your_key
 
+# Exa API for web search and research-paper search (used when SEARCH_PROVIDER=exa)
+# Get your key from: https://dashboard.exa.ai/
+EXA_API_KEY=your_exa_api_key
+
 # Jina API for web page reading
 # Get your key from: https://jina.ai/
 JINA_API_KEYS=your_key
diff --git a/inference/tests/__init__.py b/inference/tests/__init__.py
new file mode 100644
index 00000000..e69de29b
diff --git a/inference/tests/test_tool_exa.py b/inference/tests/test_tool_exa.py
new file mode 100644
index 00000000..d54f83de
--- /dev/null
+++ b/inference/tests/test_tool_exa.py
@@ -0,0 +1,159 @@
+"""Tests for the Exa search backend (``inference/tool_exa.py``).
+
+These tests do not hit the live Exa API. They mock the ``exa_py.Exa`` class
+and assert response parsing, snippet fallback ordering, the integration header,
+and behavior when ``EXA_API_KEY`` is unset.
+"""
+
+import os
+import sys
+import unittest
+from unittest import mock
+
+# The inference package uses sibling-relative imports (e.g. ``from tool_exa import ...``)
+# so we add the inference directory to sys.path for these tests.
+HERE = os.path.dirname(os.path.abspath(__file__))
+INFERENCE_DIR = os.path.dirname(HERE)
+if INFERENCE_DIR not in sys.path:
+    sys.path.insert(0, INFERENCE_DIR)
+
+
+class _FakeResult:
+    def __init__(
+        self,
+        title="",
+        url="",
+        published_date=None,
+        author=None,
+        text=None,
+        highlights=None,
+        summary=None,
+    ):
+        self.title = title
+        self.url = url
+        self.published_date = published_date
+        self.author = author
+        self.text = text
+        self.highlights = highlights
+        self.summary = summary
+
+
+class _FakeResponse:
+    def __init__(self, results):
+        self.results = results
+
+
+class _FakeExa:
+    """Stands in for ``exa_py.Exa`` and records the kwargs it receives."""
+
+    last_kwargs: dict = {}
+    last_query: str = ""
+    last_instance: "Optional[_FakeExa]" = None
+
+    def __init__(self, api_key):
+        self.api_key = api_key
+        self.headers: dict = {}
+        _FakeExa.last_instance = self
+
+    def search_and_contents(self, query, **kwargs):
+        _FakeExa.last_query = query
+        _FakeExa.last_kwargs = kwargs
+        return _FakeResponse(_FakeExa.next_results)
+
+
+_FakeExa.next_results = []
+
+
+def _patch_exa(results):
+    """Install a fake ``exa_py`` module so ``from exa_py import Exa`` returns _FakeExa."""
+    _FakeExa.next_results = results
+    fake_module = mock.MagicMock()
+    fake_module.Exa = _FakeExa
+    return mock.patch.dict(sys.modules, {"exa_py": fake_module})
+
+
+class ExaParsingTests(unittest.TestCase):
+    def setUp(self):
+        os.environ["EXA_API_KEY"] = "test-key"
+
+    def test_search_parses_results_and_sets_integration_header(self):
+        results = [
+            _FakeResult(
+                title="Example",
+                url="https://example.com/a",
+                published_date="2026-04-01",
+                highlights=["Hello world", "More context"],
+            ),
+            _FakeResult(
+                title="Second",
+                url="https://example.com/b",
+                summary="A short summary.",
+            ),
+        ]
+        with _patch_exa(results):
+            from tool_exa import search_with_exa
+
+            output = search_with_exa("python testing")
+
+        self.assertIn("python testing", output)
+        self.assertIn("Example", output)
+        self.assertIn("https://example.com/a", output)
+        self.assertIn("Hello world ... More context", output)
+        self.assertIn("A short summary.", output)
+        self.assertIn("Date published: 2026-04-01", output)
+        self.assertEqual(_FakeExa.last_query, "python testing")
+        self.assertEqual(_FakeExa.last_kwargs.get("type"), "auto")
+        self.assertIn("highlights", _FakeExa.last_kwargs)
+        self.assertIn("text", _FakeExa.last_kwargs)
+        self.assertIsNotNone(_FakeExa.last_instance)
+        self.assertEqual(
+            _FakeExa.last_instance.headers.get("x-exa-integration"),
+            "deepresearch",
+        )
+
+    def test_snippet_falls_back_through_highlights_summary_text(self):
+        from tool_exa import ExaResult
+
+        only_highlights = ExaResult(title="t", url="u", highlights=["h1", "h2"])
+        only_summary = ExaResult(title="t", url="u", summary="s")
+        only_text = ExaResult(title="t", url="u", text="t" * 600)
+        empty = ExaResult(title="t", url="u")
+
+        self.assertEqual(only_highlights.snippet(), "h1 ... h2")
+        self.assertEqual(only_summary.snippet(), "s")
+        self.assertTrue(only_text.snippet().endswith("..."))
+        self.assertLessEqual(len(only_text.snippet()), 504)
+        self.assertEqual(empty.snippet(), "")
+
+    def test_scholar_uses_research_paper_category(self):
+        results = [_FakeResult(title="Paper", url="https://arxiv.org/x", summary="S")]
+        with _patch_exa(results):
+            from tool_exa import scholar_with_exa
+
+            output = scholar_with_exa("transformer scaling laws")
+
+        self.assertIn("Scholar Results", output)
+        self.assertIn("Paper", output)
+        self.assertEqual(_FakeExa.last_kwargs.get("category"), "research paper")
+
+    def test_no_results_returns_friendly_message(self):
+        with _patch_exa([]):
+            from tool_exa import search_with_exa
+
+            output = search_with_exa("zzz nothing matches")
+
+        self.assertIn("No results found", output)
+
+
+class ExaDisabledTests(unittest.TestCase):
+    def test_missing_api_key_returns_error_string(self):
+        os.environ.pop("EXA_API_KEY", None)
+        with _patch_exa([]):
+            from tool_exa import search_with_exa
+
+            output = search_with_exa("anything")
+
+        self.assertIn("EXA_API_KEY is not set", output)
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/inference/tool_exa.py b/inference/tool_exa.py
new file mode 100644
index 00000000..a7f9661b
--- /dev/null
+++ b/inference/tool_exa.py
@@ -0,0 +1,158 @@
+"""Exa search backend.
+
+Provides Exa-powered web search and Exa-powered scholar search. The implementations
+return strings with the same numbered-snippet format produced by the Serper backends
+in ``tool_search.py`` and ``tool_scholar.py`` so the agent prompt format is unchanged.
+"""
+
+import os
+from dataclasses import dataclass
+from typing import Any, List, Optional
+
+
+_EXA_INTEGRATION_HEADER = "deepresearch"
+
+
+@dataclass
+class ExaResult:
+    title: str
+    url: str
+    published_date: Optional[str] = None
+    author: Optional[str] = None
+    text: Optional[str] = None
+    highlights: Optional[List[str]] = None
+    summary: Optional[str] = None
+
+    @classmethod
+    def from_sdk(cls, item: Any) -> "ExaResult":
+        get = lambda key: getattr(item, key, None) if not isinstance(item, dict) else item.get(key)
+        highlights = get("highlights")
+        if highlights is not None and not isinstance(highlights, list):
+            highlights = list(highlights)
+        return cls(
+            title=get("title") or "",
+            url=get("url") or "",
+            published_date=get("published_date") or get("publishedDate"),
+            author=get("author"),
+            text=get("text"),
+            highlights=highlights,
+            summary=get("summary"),
+        )
+
+    def snippet(self) -> str:
+        if self.highlights:
+            return " ... ".join(h.strip() for h in self.highlights if h)
+        if self.summary:
+            return self.summary.strip()
+        if self.text:
+            text = self.text.strip()
+            if len(text) > 500:
+                text = text[:500].rstrip() + "..."
+            return text
+        return ""
+
+
+def _build_client():
+    from exa_py import Exa
+
+    api_key = os.environ.get("EXA_API_KEY")
+    if not api_key:
+        raise RuntimeError("EXA_API_KEY is not set")
+    client = Exa(api_key=api_key)
+    try:
+        client.headers["x-exa-integration"] = _EXA_INTEGRATION_HEADER
+    except AttributeError:
+        pass
+    return client
+
+
+def _run_search(
+    query: str,
+    num_results: int,
+    category: Optional[str] = None,
+    include_domains: Optional[List[str]] = None,
+    exclude_domains: Optional[List[str]] = None,
+    start_published_date: Optional[str] = None,
+    end_published_date: Optional[str] = None,
+) -> List[ExaResult]:
+    client = _build_client()
+
+    kwargs: dict = {
+        "num_results": num_results,
+        "type": "auto",
+        "highlights": {"num_sentences": 3},
+        "text": {"max_characters": 500},
+    }
+    if category:
+        kwargs["category"] = category
+    if include_domains:
+        kwargs["include_domains"] = include_domains
+    if exclude_domains:
+        kwargs["exclude_domains"] = exclude_domains
+    if start_published_date:
+        kwargs["start_published_date"] = start_published_date
+    if end_published_date:
+        kwargs["end_published_date"] = end_published_date
+
+    last_err: Optional[Exception] = None
+    for attempt in range(5):
+        try:
+            response = client.search_and_contents(query, **kwargs)
+            raw_results = getattr(response, "results", None) or []
+            return [ExaResult.from_sdk(r) for r in raw_results]
+        except Exception as e:
+            last_err = e
+            continue
+    raise RuntimeError(f"Exa search failed after 5 attempts: {last_err}")
+
+
+def _format_block(query: str, results: List[ExaResult], header: str, label: str) -> str:
+    if not results:
+        return f"No results found for '{query}'. Try with a more general query."
+
+    snippets: List[str] = []
+    for idx, r in enumerate(results, start=1):
+        date_published = f"\nDate published: {r.published_date}" if r.published_date else ""
+        author_line = f"\nSource: {r.author}" if r.author else ""
+        snippet = r.snippet()
+        snippet_block = f"\n{snippet}" if snippet else ""
+        title = r.title or r.url
+        snippets.append(f"{idx}. [{title}]({r.url}){date_published}{author_line}{snippet_block}")
+
+    return f"{header} for '{query}' found {len(snippets)} results:\n\n## {label}\n" + "\n\n".join(snippets)
+
+
+def search_with_exa(
+    query: str,
+    num_results: int = 10,
+    category: Optional[str] = None,
+    include_domains: Optional[List[str]] = None,
+    exclude_domains: Optional[List[str]] = None,
+    start_published_date: Optional[str] = None,
+    end_published_date: Optional[str] = None,
+) -> str:
+    try:
+        results = _run_search(
+            query,
+            num_results=num_results,
+            category=category,
+            include_domains=include_domains,
+            exclude_domains=exclude_domains,
+            start_published_date=start_published_date,
+            end_published_date=end_published_date,
+        )
+    except RuntimeError as e:
+        return f"[Exa search] {e}"
+    except Exception as e:
+        return f"[Exa search] Unexpected error: {e}"
+    return _format_block(query, results, header="A web search", label="Web Results")
+
+
+def scholar_with_exa(query: str, num_results: int = 10) -> str:
+    try:
+        results = _run_search(query, num_results=num_results, category="research paper")
+    except RuntimeError as e:
+        return f"[Exa scholar] {e}"
+    except Exception as e:
+        return f"[Exa scholar] Unexpected error: {e}"
+    return _format_block(query, results, header="A research-paper search", label="Scholar Results")
diff --git a/inference/tool_scholar.py b/inference/tool_scholar.py
index ae021b38..158f0257 100644
--- a/inference/tool_scholar.py
+++ b/inference/tool_scholar.py
@@ -6,8 +6,11 @@
 from concurrent.futures import ThreadPoolExecutor
 import http.client
 
+from tool_exa import scholar_with_exa
+
 
 SERPER_KEY=os.environ.get('SERPER_KEY_ID')
+SEARCH_PROVIDER=os.environ.get('SEARCH_PROVIDER', 'serper').lower()
 
 
 @register_tool("google_scholar", allow_overwrite=True)
@@ -91,6 +94,11 @@ def google_scholar_with_serp(self, query: str):
             return f"No results found for '{query}'. Try with a more general query."
 
 
+    def _do_scholar_search(self, query: str) -> str:
+        if SEARCH_PROVIDER == "exa":
+            return scholar_with_exa(query)
+        return self.google_scholar_with_serp(query)
+
     def call(self, params: Union[str, dict], **kwargs) -> str:
         # assert GOOGLE_SEARCH_KEY is not None, "Please set the IDEALAB_SEARCH_KEY environment variable."
         try:
@@ -98,13 +106,13 @@ def call(self, params: Union[str, dict], **kwargs) -> str:
             query = params["query"]
         except:
             return "[google_scholar] Invalid request format: Input must be a JSON object containing 'query' field"
-        
+
         if isinstance(query, str):
-            response = self.google_scholar_with_serp(query)
+            response = self._do_scholar_search(query)
         else:
             assert isinstance(query, List)
             with ThreadPoolExecutor(max_workers=3) as executor:
 
-                response = list(executor.map(self.google_scholar_with_serp, query))
+                response = list(executor.map(self._do_scholar_search, query))
             response = "\n=======\n".join(response)
         return response
diff --git a/inference/tool_search.py b/inference/tool_search.py
index 1a3f7b53..0e78bb7c 100644
--- a/inference/tool_search.py
+++ b/inference/tool_search.py
@@ -11,8 +11,11 @@
 
 import os
 
+from tool_exa import search_with_exa
+
 
 SERPER_KEY=os.environ.get('SERPER_KEY_ID')
+SEARCH_PROVIDER=os.environ.get('SEARCH_PROVIDER', 'serper').lower()
 
 
 @register_tool("search", allow_overwrite=True)
@@ -110,22 +113,27 @@ def search_with_serp(self, query: str):
         result = self.google_search_with_serp(query)
         return result
 
+    def _do_search(self, query: str) -> str:
+        if SEARCH_PROVIDER == "exa":
+            return search_with_exa(query)
+        return self.search_with_serp(query)
+
     def call(self, params: Union[str, dict], **kwargs) -> str:
         try:
             query = params["query"]
         except:
             return "[Search] Invalid request format: Input must be a JSON object containing 'query' field"
-        
+
         if isinstance(query, str):
             # 单个查询
-            response = self.search_with_serp(query)
+            response = self._do_search(query)
         else:
             # 多个查询
             assert isinstance(query, List)
             responses = []
             for q in query:
-                responses.append(self.search_with_serp(q))
+                responses.append(self._do_search(q))
             response = "\n=======\n".join(responses)
-            
+
         return response
 
diff --git a/requirements.txt b/requirements.txt
index 5fd43bd1..798bd344 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -40,6 +40,7 @@ einops==0.8.1
 email-validator==2.3.0
 et_xmlfile==2.0.0
 eval_type_backport==0.2.2
+exa-py>=2.0.0
 exceptiongroup==1.3.0
 fastapi==0.116.1
 fastapi-cli==0.0.11