diff --git a/CHANGELOG.md b/CHANGELOG.md
index e8cba41..3ff6f8e 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -52,9 +52,16 @@ et [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
   **discret** au-dessus de chaque réponse rapporte le fournisseur et le modèle
   réellement utilisés (SSE `provider`/`model`) ; une **barre d'actions** au survol
   sous chaque bloc : « Copier » (texte brut + toast) et, pour les réponses,
-  « Ajouter » (insertion dans le document ouvert dans l'éditeur Forge). Fiche :
+  « Ajouter » (insertion dans le document ouvert dans l'éditeur Forge).
+  Complément : la section « N étapes » affiche désormais des libellés humains
+  produits par le backend (`backend/tools/labels.py`, événements SSE en direct
+  pendant l'exécution + ligne « Réflexion »), et deux nouveaux outils web
+  principaux sont disponibles en mode agent : `web_search` (SearXNG
+  auto-hébergé) et `fetch_url` (page publique → texte, garde SSRF). Le reste
+  des catégories d'outils Notion est documenté pour le futur dans la fiche.
+  Fiche :
   [docs/features/ai-assistant-conversation-ux.md](./docs/features/ai-assistant-conversation-ux.md)
-  Tests : `tests/frontend/ai.test.mjs` (74).
+  Tests : `tests/frontend/ai.test.mjs` (76), `tests/test_tool_labels.py` (8), `tests/test_web_tools.py` (8).
 
 ### Corrigé
 
diff --git a/backend/agent/loop.py b/backend/agent/loop.py
index 0473fde..1dfd607 100644
--- a/backend/agent/loop.py
+++ b/backend/agent/loop.py
@@ -30,6 +30,7 @@ from backend.tools.api import (
     call_tool,
     get_tool_schemas,
 )
+from backend.tools.labels import thought_step_label, tool_step_label
 
 logger = logging.getLogger("obsigate.agent.loop")
 
@@ -54,6 +55,9 @@ class ToolCallRecord:
     arguments: dict[str, Any]
     ok: bool
     result: Any
+    # Human-readable « step » label for the Notion-style UI
+    # ({key, params} — see backend.tools.labels).
+    step: dict[str, Any] = field(default_factory=dict)
 
 
 @dataclass
@@ -63,6 +67,9 @@ class AgentResult:
     content: str = ""
     messages: list[dict[str, Any]] = field(default_factory=list)
     tool_calls: list[ToolCallRecord] = field(default_factory=list)
+    # Ordered Notion-style step descriptors ({key, params}); tool steps and
+    # intermediate reasoning notes interleaved by execution order.
+    steps: list[dict[str, Any]] = field(default_factory=list)
     iterations: int = 0
     stopped: str = STOP_DONE
     pending: dict[str, Any] | None = None
@@ -142,7 +149,10 @@ def _execute_confirmed(
         payload = e.to_dict()
         ok = False
 
-    record = ToolCallRecord(name=name, arguments=arguments, ok=ok, result=payload)
+    record = ToolCallRecord(
+        name=name, arguments=arguments, ok=ok, result=payload,
+        step=tool_step_label(name, arguments),
+    )
     executed.append(record)
     if on_tool_call is not None:
         on_tool_call(record)
@@ -164,6 +174,7 @@ async def run_agent(
     max_iterations: int = DEFAULT_MAX_ITERATIONS,
     max_tool_calls: int | None = None,
     on_tool_call: Callable[[ToolCallRecord], None] | None = None,
+    on_thought: Callable[[dict[str, Any]], None] | None = None,
     resume_messages: list[dict[str, Any]] | None = None,
     confirm_pending: dict[str, Any] | None = None,
 ) -> AgentResult:
@@ -195,6 +206,15 @@ async def run_agent(
     if tools is None:
         tools = get_tool_schemas(scope=ToolScope.IN_APP)
     quota = DEFAULT_MAX_TOOL_CALLS if max_tool_calls is None else max_tool_calls
+    steps: list[dict[str, Any]] = []
+
+    def _emit_note(text: str) -> None:
+        """Record an intermediate reasoning note as a visible step."""
+        note = thought_step_label(text)
+        if note["params"]["value"]:
+            steps.append(note)
+            if on_thought is not None:
+                on_thought(note)
 
     convo = [dict(m) for m in (resume_messages if resume_messages is not None else messages)]
     executed: list[ToolCallRecord] = []
@@ -205,6 +225,7 @@ async def run_agent(
                 content="",
                 messages=convo,
                 tool_calls=executed,
+                steps=steps,
                 iterations=0,
                 stopped=STOP_QUOTA_EXCEEDED,
             )
@@ -218,10 +239,13 @@ async def run_agent(
                 content=response.content or "",
                 messages=convo,
                 tool_calls=executed,
+                steps=steps,
                 iterations=iteration,
                 stopped=STOP_DONE,
             )
 
+        # Intermediate reasoning shown alongside tool calls → a "thought" step.
+        _emit_note(response.content or "")
         convo.append(_assistant_tool_message(response.content, response.tool_calls))
 
         for call in response.tool_calls:
@@ -231,6 +255,7 @@ async def run_agent(
                     content=response.content or "",
                     messages=convo,
                     tool_calls=executed,
+                    steps=steps,
                     iterations=iteration,
                     stopped=STOP_QUOTA_EXCEEDED,
                 )
@@ -247,6 +272,7 @@ async def run_agent(
                     content=response.content or "",
                     messages=convo,
                     tool_calls=executed,
+                    steps=steps,
                     iterations=iteration,
                     stopped=STOP_CONFIRMATION_REQUIRED,
                     pending=pending,
@@ -255,8 +281,12 @@ async def run_agent(
                 payload = e.to_dict()
                 ok = False
 
-            record = ToolCallRecord(name=call.name, arguments=call.arguments, ok=ok, result=payload)
+            record = ToolCallRecord(
+                name=call.name, arguments=call.arguments, ok=ok, result=payload,
+                step=tool_step_label(call.name, call.arguments),
+            )
             executed.append(record)
+            steps.append(record.step)
             if on_tool_call is not None:
                 on_tool_call(record)
 
@@ -272,6 +302,7 @@ async def run_agent(
         content="",
         messages=convo,
         tool_calls=executed,
+        steps=steps,
         iterations=max_iterations,
         stopped=STOP_MAX_ITERATIONS,
     )
diff --git a/backend/bookslm_routes.py b/backend/bookslm_routes.py
index e48f019..6e933c9 100644
--- a/backend/bookslm_routes.py
+++ b/backend/bookslm_routes.py
@@ -317,6 +317,21 @@ def _resolve_provider_name(requested: str | None) -> str | None:
     return None
 
 
+def _tool_event_sse(rec) -> str:
+    """Serialize one executed tool call as an SSE ``tool`` event."""
+    payload = json.dumps(
+        {"name": rec.name, "ok": rec.ok, "arguments": rec.arguments, "step": rec.step},
+        ensure_ascii=False,
+    )
+    return f"event: tool\ndata: {payload}\n\n"
+
+
+def _thought_event_sse(note: dict) -> str:
+    """Serialize one intermediate reasoning note as an SSE ``step`` event."""
+    payload = json.dumps({"step": note}, ensure_ascii=False)
+    return f"event: step\ndata: {payload}\n\n"
+
+
 # ── Endpoints ──
 
 
@@ -456,6 +471,8 @@ async def api_bookslm_agent(
         )
 
     async def generate_sse():
+        import asyncio
+
         try:
             cfg_name = _resolve_provider_name(req.provider)
             if cfg_name is None:
@@ -466,20 +483,43 @@ async def api_bookslm_agent(
                 yield f"event: error\ndata: {error_data}\n\n"
                 return
 
-            result = await run_agent(
+            # Stream tool events live: each executed step is pushed on the
+            # queue by the loop callback and emitted as soon as it happens,
+            # so the UI can grow its « N steps » block while thinking.
+            queue: asyncio.Queue = asyncio.Queue()
+
+            def _on_tool(rec) -> None:
+                queue.put_nowait(("tool", rec))
+
+            run_task = asyncio.create_task(run_agent(
                 messages,
                 ctx=ctx,
                 llm=_llm,
                 resume_messages=req.confirm_messages,
                 confirm_pending=req.confirm,
-            )
+                on_tool_call=_on_tool,
+                on_thought=lambda note: queue.put_nowait(("thought", note)),
+            ))
 
-            for rec in result.tool_calls:
-                payload = json.dumps(
-                    {"name": rec.name, "ok": rec.ok, "arguments": rec.arguments},
-                    ensure_ascii=False,
-                )
-                yield f"event: tool\ndata: {payload}\n\n"
+            # Drain every completed step as soon as it lands, while the agent
+            # keeps running in the background. If the client disconnects, the
+            # generator is cancelled: release the run so it cannot orphan.
+            try:
+                while True:
+                    try:
+                        kind, item = await asyncio.wait_for(queue.get(), timeout=0.25)
+                    except asyncio.TimeoutError:
+                        if run_task.done():
+                            break
+                        continue
+                    yield _tool_event_sse(item) if kind == "tool" else _thought_event_sse(item)
+                while not queue.empty():
+                    kind, item = queue.get_nowait()
+                    yield _tool_event_sse(item) if kind == "tool" else _thought_event_sse(item)
+                result = run_task.result()
+            finally:
+                if not run_task.done():
+                    run_task.cancel()
 
             if result.stopped == "confirmation_required":
                 pending = json.dumps(
diff --git a/backend/tools/api.py b/backend/tools/api.py
index 6061c7f..3944572 100644
--- a/backend/tools/api.py
+++ b/backend/tools/api.py
@@ -10,6 +10,7 @@ which ``.gitignore`` excludes via ``_*.py``), hence this explicit facade.
 """
 
 from backend.tools import service as _service  # noqa: F401  (registers tools)
+from backend.tools import web as _web  # noqa: F401  (registers web tools)
 from backend.tools.context import (
     ToolConfirmationRequired,
     ToolContext,
diff --git a/backend/tools/labels.py b/backend/tools/labels.py
new file mode 100644
index 0000000..4507080
--- /dev/null
+++ b/backend/tools/labels.py
@@ -0,0 +1,89 @@
+"""Human-readable labels for tool calls (Notion-style “steps” section).
+
+The assistant's conversation window shows a collapsible « N steps » block
+above each answer. Raw tool names (``search_fulltext``) are developer-centric;
+this module maps every registered tool to an i18n message key plus the
+argument worth surfacing (path, query, …), so the UI can render sentences like
+« Recherche dans le vault : pizza » or « Fichier lu : notes/x.md ».
+
+The label KEY is emitted with each ``tool`` SSE event; the client resolves it
+against its locale dictionaries (``ai.step.<key>``). Unknown tools fall back to
+a generic key with the raw name prettified, so a new tool never breaks the UI.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+# tool name -> (message-key, primary argument name or None)
+# NOTE: argument names MUST match the pydantic input models in
+# backend/tools/schemas.py (verified by tests/test_tool_labels.py).
+_STEP_LABELS: dict[str, tuple[str, str | None]] = {
+    "list_vaults": ("vaults", None),
+    "list_directory": ("directory", "path"),
+    "list_all_files": ("files", "dir"),
+    "read_file": ("file_read", "path"),
+    "read_file_raw": ("file_read", "path"),
+    "get_backlinks": ("backlinks", "path"),
+    "list_backups": ("backups", "path"),
+    "diff_backup": ("backup_diff", "path"),
+    "get_graph": ("graph", None),
+    "search_fulltext": ("search", "q"),
+    "search_advanced": ("search", "q"),
+    "search_paths": ("search_paths", "q"),
+    "list_tags": ("tags", None),
+    "suggest_tags": ("tags_suggest", "q"),
+    "list_recent": ("recent", None),
+    "create_file": ("file_create", "path"),
+    "create_directory": ("dir_create", "path"),
+    "edit_file": ("file_edit", "path"),
+    "append_to_file": ("file_append", "path"),
+    "rename_file": ("file_rename", "path"),
+    "rename_directory": ("dir_rename", "path"),
+    "move_path": ("move", "source_path"),
+    "replace_in_files": ("replace", "find"),
+    "delete_file": ("file_delete", "path"),
+    "delete_directory": ("dir_delete", "path"),
+    "restore_backup": ("backup_restore", "path"),
+    "web_search": ("web_search", "query"),
+    "fetch_url": ("fetch_url", "url"),
+}
+
+GENERIC_KEY = "generic"
+
+
+def _prettify(name: str) -> str:
+    return name.replace("_", " ").strip()
+
+
+def _primary_value(arguments: dict[str, Any], arg: str | None) -> str | None:
+    if not arg:
+        return None
+    value = arguments.get(arg)
+    if isinstance(value, str) and value.strip():
+        return value.strip()
+    return None
+
+
+def tool_step_label(name: str, arguments: dict[str, Any] | None = None) -> dict[str, Any]:
+    """Return ``{key, params}`` describing one tool call in human terms.
+
+    ``key`` resolves client-side to ``ai.step.<key>``; ``params`` carries the
+    optional ``{value}`` placeholder (path, query, …). For an unknown tool the
+    generic key is used with the prettified name as the value.
+    """
+    arguments = arguments or {}
+    key, arg = _STEP_LABELS.get(name, (GENERIC_KEY, None))
+    if key == GENERIC_KEY:
+        return {"key": GENERIC_KEY, "params": {"tool": _prettify(name)}}
+    value = _primary_value(arguments, arg)
+    params: dict[str, str] = {}
+    if value is not None:
+        params["value"] = value
+    return {"key": key, "params": params}
+
+
+def thought_step_label(text: str) -> dict[str, Any]:
+    """Step descriptor for one intermediate reasoning note of the model."""
+    cleaned = " ".join((text or "").split())
+    return {"key": "thought", "params": {"value": cleaned[:400]}}
diff --git a/backend/tools/schemas.py b/backend/tools/schemas.py
index f233255..601b93f 100644
--- a/backend/tools/schemas.py
+++ b/backend/tools/schemas.py
@@ -237,6 +237,22 @@ class RestoreBackupInput(BaseModel):
     version: int = Field(..., description="Backup timestamp to restore")
 
 
+class WebSearchInput(BaseModel):
+    """Search the public web through the configured meta-search engine."""
+
+    query: str = Field(..., description="Search terms")
+    max_results: int = Field(5, ge=1, le=10, description="Number of results to return")
+    category: str = Field("", description="Optional engine category: general, news, it, science")
+    language: str = Field("", description="Optional language code, e.g. 'fr'")
+    page: int = Field(1, ge=1, le=10, description="Result page number")
+
+
+class FetchUrlInput(BaseModel):
+    """Fetch one public web page and return its readable text."""
+
+    url: str = Field(..., description="Absolute http(s) URL of a public page")
+
+
 class ToolResult(BaseModel):
     """Uniform result returned by :func:`backend.tools.registry.call_tool`."""
 
diff --git a/backend/tools/web.py b/backend/tools/web.py
new file mode 100644
index 0000000..aa48344
--- /dev/null
+++ b/backend/tools/web.py
@@ -0,0 +1,217 @@
+"""Web tools for the assistant (Notion-style "research" capabilities).
+
+Phase 1 of the documented web-toolset roadmap:
+
+* ``web_search``  — query the self-hosted SearXNG instance (no API key).
+* ``fetch_url``   — retrieve a public web page and return readable text.
+
+Both are READ-risk tools (no confirmation), rate-limited through the shared
+registry, SSRF-guarded (scheme + private-address rejection), and size-capped.
+
+Configuration (environment):
+* ``OBSIGATE_SEARXNG_URL``   — defaults to https://search.dracodev.net
+* ``OBSIGATE_WEB_TIMEOUT``   — seconds, default 10
+"""
+
+from __future__ import annotations
+
+import html as html_lib
+import ipaddress
+import logging
+import os
+import re
+import socket
+from typing import Any
+from urllib.parse import urlparse
+
+import httpx
+
+from backend.tools.context import ToolError, ToolRisk, ToolScope
+from backend.tools.registry import tool
+from backend.tools.schemas import FetchUrlInput, WebSearchInput
+
+logger = logging.getLogger("obsigate.tools.web")
+
+SEARXNG_URL = os.environ.get("OBSIGATE_SEARXNG_URL", "https://search.dracodev.net")
+WEB_TIMEOUT = float(os.environ.get("OBSIGATE_WEB_TIMEOUT", "10"))
+USER_AGENT = "ObsiGateAssistant/1.0 (+self-hosted vault AI)"
+MAX_FETCH_BYTES = 1_500_000
+MAX_TEXT_CHARS = 20_000
+
+_BLOCKED_TAGS_RE = re.compile(
+    r"<(script|style|noscript|template|svg)\b.*?</\1>", re.IGNORECASE | re.DOTALL
+)
+_TAG_RE = re.compile(r"<[^>]+>")
+_BLOCK_SPLIT_RE = re.compile(
+    r"</?(?:p|div|br|li|h[1-6]|tr|table|ul|ol|section|article|header|footer)\b[^>]*>",
+    re.IGNORECASE,
+)
+
+
+class SSRFError(ToolError):
+    """Raised for a URL whose host is private/loopback or scheme unsupported."""
+
+
+def _assert_public_http_url(url: str) -> str:
+    """Reject non-http(s) schemes and private/loopback/link-local targets."""
+    try:
+        parsed = urlparse(url)
+    except ValueError as e:
+        raise SSRFError("URL invalide", code="invalid_url") from e
+    if parsed.scheme not in ("http", "https"):
+        raise SSRFError("Seuls les schémas http/https sont autorisés", code="invalid_scheme")
+    host = parsed.hostname
+    if not host:
+        raise SSRFError("URL sans hôte", code="invalid_url")
+    # Resolve the host so DNS-rebinding to internal IPs is also caught.
+    try:
+        infos = socket.getaddrinfo(host, None)
+    except socket.gaierror as e:
+        raise SSRFError(f"Hôte introuvable: {host}", code="dns_error") from e
+    for info in infos:
+        ip = ipaddress.ip_address(info[4][0])
+        if (
+            ip.is_private
+            or ip.is_loopback
+            or ip.is_link_local
+            or ip.is_reserved
+            or ip.is_multicast
+            or ip.is_unspecified
+        ):
+            raise SSRFError("Accès aux adresses internes interdit", code="ssrf_blocked")
+    return url
+
+
+def _html_to_text(raw: str) -> str:
+    """Cheap HTML → readable text: strip scripts/styles, tags, then compress.
+
+    Comments (which may carry script-like payloads) are removed first.
+    """
+    text = re.sub(r"<!--.*?-->", " ", raw, flags=re.DOTALL)
+    text = _BLOCKED_TAGS_RE.sub(" ", text)
+    # Keep block boundaries as newlines before dropping the remaining tags.
+    text = _BLOCK_SPLIT_RE.sub("\n", text)
+    text = _TAG_RE.sub("", text)
+    text = html_lib.unescape(text)
+    text = re.sub(r"[ \t]+", " ", text)
+    text = re.sub(r" ?\n ?", "\n", text)
+    text = re.sub(r"\n{3,}", "\n\n", text)
+    return text.strip()
+
+
+@tool(
+    name="web_search",
+    description=(
+        "Search the public web for current information and return ranked results "
+        "(title, url, snippet). Use for facts outside the vault: weather, news, "
+        "documentation, versions, prices, anything that needs live sources."
+    ),
+    input_model=WebSearchInput,
+    risk=ToolRisk.READ,
+    scopes=(ToolScope.IN_APP,),
+)
+def web_search(ctx, params: WebSearchInput) -> dict[str, Any]:
+    """Query the self-hosted SearXNG instance and return trimmed results."""
+    query = params.query.strip()
+    if not query:
+        raise ToolError("Requête vide", code="invalid_arguments")
+    url = SEARXNG_URL.rstrip("/") + "/search"
+    try:
+        resp = httpx.get(
+            url,
+            params={
+                "q": query,
+                "format": "json",
+                "categories": params.category or "general",
+                "pageno": max(1, params.page),
+                **({"language": params.language} if params.language else {}),
+                "safesearch": "1",
+            },
+            headers={"User-Agent": USER_AGENT},
+            timeout=WEB_TIMEOUT,
+            follow_redirects=False,
+        )
+        resp.raise_for_status()
+        data = resp.json()
+    except httpx.HTTPError as e:
+        logger.warning("web_search failed: %s", e)
+        raise ToolError(
+            "Le moteur de recherche web est momentanément indisponible.",
+            code="web_search_unavailable",
+        ) from e
+    results: list[dict[str, Any]] = []
+    for item in (data.get("results") or [])[: params.max_results]:
+        results.append(
+            {
+                "title": (item.get("title") or "")[:300],
+                "url": item.get("url") or "",
+                "snippet": (item.get("content") or "")[:600],
+                "published": item.get("publishedDate"),
+                "score": item.get("score"),
+            }
+        )
+    return {
+        "query": query,
+        "engine": "searxng",
+        "results": results,
+        "count": len(results),
+    }
+
+
+@tool(
+    name="fetch_url",
+    description=(
+        "Fetch a public web page (http/https) and return its readable text. "
+        "Use after web_search to read a promising result in detail. HTML is "
+        "converted to plain text; binary pages are rejected."
+    ),
+    input_model=FetchUrlInput,
+    risk=ToolRisk.READ,
+    scopes=(ToolScope.IN_APP,),
+)
+def fetch_url(ctx, params: FetchUrlInput) -> dict[str, Any]:
+    """Retrieve one page, guard against SSRF, and extract its text."""
+    url = _assert_public_http_url(params.url.strip())
+    try:
+        # Follow redirects manually so every hop is re-checked against the
+        # private-address SSRF guard (a public page can redirect to 127.0.0.1).
+        resp = None
+        for _hop in range(5):
+            resp = httpx.get(
+                url,
+                headers={"User-Agent": USER_AGENT, "Accept": "text/html,application/xhtml+xml,*/*"},
+                timeout=WEB_TIMEOUT,
+                follow_redirects=False,
+            )
+            if resp.status_code in (301, 302, 303, 307, 308):
+                location = resp.headers.get("location") or ""
+                if not location:
+                    break
+                url = str(httpx.URL(url).join(location))
+                url = _assert_public_http_url(url)
+                continue
+            break
+        assert resp is not None
+        resp.raise_for_status()
+    except SSRFError:
+        raise
+    except httpx.HTTPError as e:
+        logger.warning("fetch_url failed for %s: %s", url, e)
+        raise ToolError("Impossible de récupérer la page.", code="fetch_unavailable") from e
+    ctype = (resp.headers.get("content-type") or "").lower()
+    if not any(t in ctype for t in ("html", "xml", "text", "json", "markdown")):
+        raise ToolError(
+            f"Type de contenu non pris en charge: {ctype.split(';')[0] or 'inconnu'}",
+            code="unsupported_content_type",
+        )
+    raw = (resp.content[:MAX_FETCH_BYTES]).decode(resp.encoding or "utf-8", errors="replace")
+    title_match = re.search(r"<title[^>]*>(.*?)</title>", raw, re.IGNORECASE | re.DOTALL)
+    title = html_lib.unescape(title_match.group(1)).strip()[:300] if title_match else ""
+    text = _html_to_text(raw)[:MAX_TEXT_CHARS]
+    return {
+        "url": str(resp.url),
+        "status": resp.status_code,
+        "title": title,
+        "text": text,
+        "truncated": len(raw) > MAX_TEXT_CHARS,
+    }
diff --git a/docs/features/ai-assistant-conversation-ux.md b/docs/features/ai-assistant-conversation-ux.md
index 4e1cb5d..35514f5 100644
--- a/docs/features/ai-assistant-conversation-ux.md
+++ b/docs/features/ai-assistant-conversation-ux.md
@@ -80,4 +80,40 @@
   - L'ancien épinglage flex `order: -1` (première itération de #91) est **remplacé** :
     l'ordre est redevenu chronologique, l'ancrage est fait par défilement explicite.
   - Les pouces haut/bas Notion ne sont **pas** repris : aucun signal n'existe
-    côté backend ; « Ajouter » (insertion éditeur réelle) les remplace utilement.
\ No newline at end of file
+    côté backend ; « Ajouter » (insertion éditeur réelle) les remplace utilement.
+
+## G. Section « steps » enrichie + outils web (complément #91) — ✅ livré
+  - [x] **G1.** Chaque étape est décrite par le **backend** (`backend/tools/labels.py`) :
+    événement SSE `tool` avec `step {key, params}` ; le frontend résout
+    `ai.step.<key>` (FR/EN) → phrases humaines : « Recherche dans le vault : pizza »,
+    « Fichier lu : notes/a.md », etc. Outil inconnu → libellé générique (jamais cassé).
+  - [x] **G2.** Événements en **direct** : les steps sont poussés sur une file
+    `asyncio.Queue` par la boucle agent et émis dès leur exécution — le bloc
+    « N étapes » grandit pendant que l'assistant travaille (plus de liste fin de run).
+  - [x] **G3.** Ligne « Réflexion » (`ai.step.thought`) : quand le modèle émet un
+    texte intermédiaire avec ses appels d'outils, il est publié comme event SSE
+    `step` et affiché dans la liste (`. thought` Notion).
+  - [x] **G4.** Nouveaux outils principaux côté **web** (`backend/tools/web.py`,
+    risques READ, scope IN_APP, SSRF-guard + limite de taille) :
+    `web_search` (SearXNG auto-hébergé, configurable via `OBSIGATE_SEARXNG_URL`)
+    et `fetch_url` (lecture d'une page publique, HTML → texte).
+    Steps affichés : « Recherche sur le web : … » / « Page web consultée : … ».
+
+### Outils restants — documentés pour le futur (hors #91)
+La catégorie Notion « étapes » peut s'étendre ; chaque futur outil devra être un
+tool du registre (`@tool`) + un libellé dans `labels.py` + deux clés i18n.
+Priorités proposées (à transformer en items `#NN` quand implémentés) :
+  - **Créer/supprimer un fichier depuis une réponse** : déjà couvert par les cartes
+    d'action `create_file` / l'outil `create_file` — exposer une action « Créer la
+    note » dans la barre d'actions (post-traitement du texte copié).
+  - **Insérer dans le document courant** : l'équivalent « Ajouter » côté assistant
+    est livré (G) ; l'export vers une sélection précise de l'éditeur reste possible.
+  - **Convertir un format / tableur / document** : outils `convert_format`,
+    `create_spreadsheet` (csv/xlsx via backend) — risque WRITE (confirmation).
+  - **Calendrier / tâches** : intégrer l'API existante `n8n-automation` ou un
+    serveur MCP externe (la porte MCP est déjà ouverte côté ObsiGate).
+  - **Courriel / messagerie** : via le serveur MCP externe (Gmail/IMPT SMTP) —
+    jamais de clé en dur : passer par Infisical.
+  - **Sources connectées (Slack, GitHub, Drive)** : uniquement par **MCP externe**
+    (`docs/features/ai-tools-mcp.md`), pas d'outils natifs — l'attaque surface
+    reste dans le registre + rate-limit + audit.
\ No newline at end of file
diff --git a/frontend/js/bookslm.js b/frontend/js/bookslm.js
index 2f47dea..a2efed6 100644
--- a/frontend/js/bookslm.js
+++ b/frontend/js/bookslm.js
@@ -2058,22 +2058,35 @@ class BooksLM {
 
   // ── Tool activity & confirmations (agent mode) ──────────────────────
 
-  /** Agent tool calls collapse into a discreet “N steps” line (Notion-style). */
-  _renderToolActivity(toolCalls) {
-    const wrap = document.createElement('details');
-    wrap.className = 'bookslm-tool-trace';
-    const summary = document.createElement('summary');
-    summary.textContent = t('ai.steps_count', { count: toolCalls.length });
-    wrap.appendChild(summary);
-    for (const call of toolCalls) {
-      const line = document.createElement('div');
-      line.className = 'bookslm-tool-line' + (call.ok === false ? ' failed' : '');
-      line.textContent = `${call.ok === false ? '⚠' : '🔧'} ${t('ai.tool_call', { name: call.name })}`;
-      if (call.ok === false) line.title = t('ai.tool_call_failed', { name: call.name });
-      wrap.appendChild(line);
+  /** Human one-liner for a recorded step (backend label key or legacy name). */
+    _stepText(call) {
+      const step = call.step;
+      if (step && step.key) {
+        const key = `ai.step.${step.key}`;
+        const label = t(key, step.params || {});
+        if (label && label !== key) return label;
+      }
+      return t('ai.tool_call', { name: call.name });
+    }
+
+    /** Agent tool calls collapse into a discreet “N steps” line (Notion-style). */
+    _renderToolActivity(toolCalls) {
+      const wrap = document.createElement('details');
+      wrap.className = 'bookslm-tool-trace';
+      const summary = document.createElement('summary');
+      const countKey = toolCalls.length > 1 ? 'ai.steps_count_plural' : 'ai.steps_count';
+      summary.textContent = t(countKey, { count: toolCalls.length });
+      wrap.appendChild(summary);
+      for (const call of toolCalls) {
+        const line = document.createElement('div');
+        line.className = 'bookslm-tool-line' + (call.ok === false ? ' failed' : '');
+        line.textContent = `${call.ok === false ? '⚠' : '·'} ${this._stepText(call)}`;
+        if (call.ok === false) line.title = t('ai.tool_call_failed', { name: call.name });
+        else if (call.step && call.step.params && call.step.params.value) line.title = call.step.params.value;
+        wrap.appendChild(line);
+      }
+      return wrap;
     }
-    return wrap;
-  }
 
   _renderConfirmationCard(msg) {
     const conf = msg.confirmation;
@@ -2393,10 +2406,22 @@ class BooksLM {
       }
 
       if (currentEvent === 'tool') {
-        assistantMsg.toolCalls.push({ name: data.name, ok: data.ok !== false, arguments: data.arguments || {} });
+        assistantMsg.toolCalls.push({
+          name: data.name,
+          ok: data.ok !== false,
+          arguments: data.arguments || {},
+          step: data.step || null,
+        });
         this._setActivity('working', t('ai.activity_tool', { name: data.name }));
         return;
       }
+      // Intermediate reasoning note streamed by the agent (Notion-style
+      // ". thought" line inside the steps block).
+      if (currentEvent === 'step' && data.step) {
+        assistantMsg.toolCalls.push({ name: '', ok: true, arguments: {}, step: data.step });
+        this._setActivity('working', t('ai.activity_thinking'));
+        return;
+      }
       if (currentEvent === 'confirmation') {
         assistantMsg.confirmation = {
           pending: data.pending || data,
diff --git a/frontend/locales/en.json b/frontend/locales/en.json
index 1d8e238..d049353 100644
--- a/frontend/locales/en.json
+++ b/frontend/locales/en.json
@@ -1780,7 +1780,37 @@
   "bookslm.insert_hint": "Append the answer to the document open in the editor",
   "bookslm.inserted": "Answer added to the document",
   "bookslm.insert_no_editor": "No document open in the editor",
-  "ai.steps_count": "{count} steps",
+  "ai.steps_count": "{count} step",
+  "ai.steps_count_plural": "{count} steps",
+  "ai.activity_thinking": "Thinking…",
+  "ai.step.thought": "Thought: {value}",
+  "ai.step.backlinks": "Analyzed backlinks",
+  "ai.step.backup_diff": "Compared backup: {value}",
+  "ai.step.backup_restore": "Restored backup: {value}",
+  "ai.step.backups": "Listed backups",
+  "ai.step.dir_create": "Proposed directory: {value}",
+  "ai.step.dir_delete": "Deleted directory: {value}",
+  "ai.step.dir_rename": "Renamed directory: {value}",
+  "ai.step.directory": "Explored directory: {value}",
+  "ai.step.file_append": "Appended to file: {value}",
+  "ai.step.file_create": "Proposed file: {value}",
+  "ai.step.file_delete": "Deleted file: {value}",
+  "ai.step.file_edit": "Edited file: {value}",
+  "ai.step.file_read": "Read file: {value}",
+  "ai.step.file_rename": "Renamed file: {value}",
+  "ai.step.files": "Listed files",
+  "ai.step.generic": "Used tool: {tool}",
+  "ai.step.graph": "Consulted the vault graph",
+  "ai.step.move": "Moved: {value}",
+  "ai.step.recent": "Listed recent files",
+  "ai.step.replace": "Replacements: {value}",
+  "ai.step.search": "Searched the vault: {value}",
+  "ai.step.search_paths": "Searched paths: {value}",
+  "ai.step.tags": "Listed tags",
+  "ai.step.tags_suggest": "Suggested tags for {value}",
+  "ai.step.vaults": "Listed the vaults",
+  "ai.step.fetch_url": "Opened a web page: {value}",
+  "ai.step.web_search": "Searched the web: {value}",
   "bookslm.copied": "Copied to clipboard",
   "bookslm.error": "AI service error",
   "bookslm.regenerate": "Regenerate",
diff --git a/frontend/locales/fr.json b/frontend/locales/fr.json
index 85feacf..9e84935 100644
--- a/frontend/locales/fr.json
+++ b/frontend/locales/fr.json
@@ -1780,7 +1780,37 @@
   "bookslm.insert_hint": "Ajouter la réponse au document ouvert dans l'éditeur",
   "bookslm.inserted": "Réponse ajoutée au document",
   "bookslm.insert_no_editor": "Aucun document ouvert dans l'éditeur",
-  "ai.steps_count": "{count} étapes",
+  "ai.steps_count": "{count} étape",
+  "ai.steps_count_plural": "{count} étapes",
+  "ai.activity_thinking": "Réflexion…",
+  "ai.step.thought": "Réflexion : {value}",
+  "ai.step.backlinks": "Backlinks analysés",
+  "ai.step.backup_diff": "Comparaison de sauvegarde : {value}",
+  "ai.step.backup_restore": "Sauvegarde restaurée : {value}",
+  "ai.step.backups": "Sauvegardes consultées",
+  "ai.step.dir_create": "Dossier proposé : {value}",
+  "ai.step.dir_delete": "Dossier supprimé : {value}",
+  "ai.step.dir_rename": "Dossier renommé : {value}",
+  "ai.step.directory": "Répertoire exploré : {value}",
+  "ai.step.file_append": "Fichier complété : {value}",
+  "ai.step.file_create": "Fichier proposé : {value}",
+  "ai.step.file_delete": "Fichier supprimé : {value}",
+  "ai.step.file_edit": "Fichier modifié : {value}",
+  "ai.step.file_read": "Fichier lu : {value}",
+  "ai.step.file_rename": "Fichier renommé : {value}",
+  "ai.step.files": "Fichiers listés",
+  "ai.step.generic": "Outil utilisé : {tool}",
+  "ai.step.graph": "Graphe du vault consulté",
+  "ai.step.move": "Élément déplacé : {value}",
+  "ai.step.recent": "Fichiers récents consultés",
+  "ai.step.replace": "Remplacements : {value}",
+  "ai.step.search": "Recherche dans le vault : {value}",
+  "ai.step.search_paths": "Chemins recherchés : {value}",
+  "ai.step.tags": "Tags consultés",
+  "ai.step.tags_suggest": "Tags suggérés pour {value}",
+  "ai.step.vaults": "Liste des vaults consultée",
+  "ai.step.fetch_url": "Page web consultée : {value}",
+  "ai.step.web_search": "Recherche sur le web : {value}",
   "bookslm.copied": "Réponse copiée dans le presse-papiers",
   "bookslm.error": "Erreur du service AI",
   "bookslm.regenerate": "Régénérer",
diff --git a/tests/frontend/ai.test.mjs b/tests/frontend/ai.test.mjs
index 9938ef1..f9bb1c7 100644
--- a/tests/frontend/ai.test.mjs
+++ b/tests/frontend/ai.test.mjs
@@ -335,7 +335,7 @@ async function main() {
     await test("agent tool calls collapse into a discreet steps block", () => {
       const b = new BooksLM();
       const trace = b._renderToolActivity([
-        { name: "read_file", ok: true },
+        { name: "read_file", ok: true, step: { key: "file_read", params: { value: "a.md" } } },
         { name: "search", ok: true },
       ]);
       assert.equal(trace.tagName, "DETAILS", "collapsible details element");
@@ -345,6 +345,18 @@ async function main() {
       assert.equal(trace.querySelectorAll(".bookslm-tool-line").length, 2, "each step listed inside");
     });
 
+        await test("_stepText uses the backend label key, falls back to tool name", () => {
+          const b = new BooksLM();
+          // No locales loaded in the JSDOM harness → t() returns the key, so the
+          // fallback path (legacy name) is what surfaces; assert it never throws
+          // and produces a non-empty string for both shapes.
+          const withStep = b._stepText({ name: "read_file", step: { key: "file_read", params: { value: "a.md" } } });
+          const legacy = b._stepText({ name: "read_file" });
+          assert.equal(typeof withStep, "string");
+          assert.equal(typeof legacy, "string");
+          assert.ok(legacy.length > 0);
+        });
+
   // ── 5. Source guards: no window.prompt() and authenticated BooksLM ──
   await test("ai.js no longer uses window.prompt()", async () => {
     const { readFileSync } = await import("node:fs");
diff --git a/tests/test_tool_labels.py b/tests/test_tool_labels.py
new file mode 100644
index 0000000..975d4ca
--- /dev/null
+++ b/tests/test_tool_labels.py
@@ -0,0 +1,67 @@
+"""Unit tests for the Notion-style step labels (backend.tools.labels)."""
+
+from backend.agent.loop import ToolCallRecord
+from backend.tools.labels import GENERIC_KEY, _STEP_LABELS, tool_step_label
+from backend.tools.registry import list_tools
+from backend.tools.context import ToolScope
+
+
+class TestToolStepLabel:
+    def test_every_in_app_tool_has_a_label(self):
+        # A registered tool without a mapping still degrades gracefully to the
+        # generic key, but the curated coverage must stay complete for the UI.
+        names = {spec.name for spec in list_tools(scope=ToolScope.IN_APP)}
+        assert names.issubset(set(_STEP_LABELS)), names - set(_STEP_LABELS)
+
+    def test_label_argument_names_match_input_models(self):
+        # A mapped primary argument that the tool's input model does not
+        # actually declare would silently render "{value}" placeholders —
+        # verify every (name, arg) pair against the real schemas.
+        from backend.tools.registry import get_tool
+
+        for name, (_key, arg) in _STEP_LABELS.items():
+            if arg is None:
+                continue
+            spec = get_tool(name)
+            assert spec is not None, name
+            fields = spec.input_model.model_fields
+            assert arg in fields, f"{name}: input model has no field '{arg}'"
+
+    def test_named_tool_returns_key_and_value(self):
+        label = tool_step_label("read_file", {"vault": "V", "path": "notes/a.md"})
+        assert label["key"] == "file_read"
+        assert label["params"]["value"] == "notes/a.md"
+
+    def test_search_tool_surfaces_the_query(self):
+        label = tool_step_label("search_fulltext", {"q": "pizza", "vault": "V"})
+        assert label["key"] == "search"
+        assert label["params"]["value"] == "pizza"
+
+    def test_no_argument_tools_have_empty_params(self):
+        label = tool_step_label("list_vaults", {})
+        assert label["key"] == "vaults"
+        assert label["params"] == {}
+
+    def test_missing_argument_degrades_to_key_only(self):
+        label = tool_step_label("read_file", {})
+        assert label["key"] == "file_read"
+        assert "value" not in label["params"]
+
+    def test_unknown_tool_falls_back_to_generic(self):
+        label = tool_step_label("future_tool_name", {"x": 1})
+        assert label["key"] == GENERIC_KEY
+        assert label["params"]["tool"] == "future tool name"
+
+    def test_arguments_none_is_accepted(self):
+        assert tool_step_label("list_tags", None)["key"] == "tags"
+
+    def test_tool_call_record_carries_step(self):
+        rec = ToolCallRecord(
+            name="read_file",
+            arguments={"path": "a.md"},
+            ok=True,
+            result={},
+            step=tool_step_label("read_file", {"path": "a.md"}),
+        )
+        assert rec.step["key"] == "file_read"
+        assert rec.step["params"]["value"] == "a.md"
diff --git a/tests/test_web_tools.py b/tests/test_web_tools.py
new file mode 100644
index 0000000..bef536a
--- /dev/null
+++ b/tests/test_web_tools.py
@@ -0,0 +1,112 @@
+"""Unit tests for the web tools (backend.tools.web): web_search + fetch_url."""
+
+from typing import Any
+
+import pytest
+
+import backend.tools.web as web
+from backend.tools.api import get_tool_schemas
+from backend.tools.context import ToolContext, ToolError, ToolMode, ToolRisk
+from backend.tools.registry import get_tool
+
+
+class FakeResponse:
+    def __init__(self, payload: Any = None, json_data: Any = None, status_code: int = 200,
+                 content: bytes = b"", headers: dict | None = None, url: str = "https://example.com/x"):
+        self._payload = payload
+        self._json = json_data
+        self.status_code = status_code
+        self.content = content
+        self.headers = headers or {}
+        self.url = url
+        self.encoding = "utf-8"
+
+    def json(self):
+        return self._json
+
+    def raise_for_status(self):
+        if self.status_code >= 400:
+            raise web.httpx.HTTPStatusError("boom", request=None, response=self)  # type: ignore[arg-type]
+
+
+def _ctx() -> ToolContext:
+    return ToolContext(user={"username": "tester", "vaults": []}, mode=ToolMode.IN_APP)
+
+
+class TestRegistration:
+    def test_tools_registered_read_only_in_app(self):
+        for name in ("web_search", "fetch_url"):
+            spec = get_tool(name)
+            assert spec is not None, name
+            assert spec.risk == ToolRisk.READ
+            assert name in [s["function"]["name"] for s in get_tool_schemas()]
+
+
+class TestWebSearch:
+    def test_returns_trimmed_results(self, monkeypatch):
+        captured = {}
+
+        def fake_get(url, params=None, **kw):
+            captured["url"] = url
+            captured["params"] = params
+            return FakeResponse(json_data={
+                "results": [
+                    {"title": "A", "url": "https://a.dev", "content": "x" * 900, "score": 1.0,
+                     "publishedDate": None},
+                ] * 12
+            })
+
+        monkeypatch.setattr(web.httpx, "get", fake_get)
+        out = web.web_search(_ctx(), web.WebSearchInput(query="pizza", max_results=3))
+        assert out["engine"] == "searxng"
+        assert out["count"] == 3
+        assert len(out["results"][0]["snippet"]) <= 600
+        assert captured["params"]["q"] == "pizza"
+
+    def test_empty_query_rejected(self):
+        with pytest.raises(ToolError):
+            web.web_search(_ctx(), web.WebSearchInput(query="   "))
+
+    def test_engine_unavailable_maps_to_tool_error(self, monkeypatch):
+        def boom(*a, **kw):
+            raise web.httpx.ConnectError("down")
+
+        monkeypatch.setattr(web.httpx, "get", boom)
+        with pytest.raises(ToolError) as ei:
+            web.web_search(_ctx(), web.WebSearchInput(query="x"))
+        assert ei.value.code == "web_search_unavailable"
+
+
+class TestFetchUrl:
+    def test_html_converted_to_text(self, monkeypatch):
+        html = (b"<html><head><title>T&</title><style>b{}</style>"
+                b"<script>evil()</script></head><body><p>hello</p><ul>"
+                b"<li>one</li><li>two</li></ul></body></html>")
+        monkeypatch.setattr(web.httpx, "get",
+                            lambda *a, **k: FakeResponse(content=html,
+                                                         headers={"content-type": "text/html; charset=utf-8"}))
+        out = web.fetch_url(_ctx(), web.FetchUrlInput(url="https://example.com/x"))
+        assert out["title"] == "T&"
+        assert "hello" in out["text"]
+        assert "one" in out["text"] and "two" in out["text"]
+        assert "evil" not in out["text"]
+        assert out["status"] == 200
+
+    def test_private_address_rejected(self):
+        for url in ("http://127.0.0.1/admin", "http://169.254.1.1/x", "http://localhost:8000/api"):
+            with pytest.raises(ToolError) as ei:
+                web.fetch_url(_ctx(), web.FetchUrlInput(url=url))
+            assert ei.value.code in ("ssrf_blocked", "dns_error", "invalid_scheme"), url
+
+    def test_non_http_scheme_rejected(self):
+        with pytest.raises(ToolError) as ei:
+            web.fetch_url(_ctx(), web.FetchUrlInput(url="file:///etc/passwd"))
+        assert ei.value.code == "invalid_scheme"
+
+    def test_binary_content_rejected(self, monkeypatch):
+        monkeypatch.setattr(web.httpx, "get",
+                            lambda *a, **k: FakeResponse(content=b"%PDF-1.4...",
+                                                         headers={"content-type": "application/pdf"}))
+        with pytest.raises(ToolError) as ei:
+            web.fetch_url(_ctx(), web.FetchUrlInput(url="https://example.com/f.pdf"))
+        assert ei.value.code == "unsupported_content_type"