diff --git a/src/clare/config.py b/src/clare/config.py
index cd3fdac..60df837 100644
--- a/src/clare/config.py
+++ b/src/clare/config.py
@@ -54,6 +54,11 @@ class OrchestratorConfig(_Strict):
     # the fleet, identifies blockers, and surfaces next-step recommendations
     # — closing the loop on "reduce required user interaction." 0 disables.
     rounds_interval_s: int = Field(default=0, ge=0, le=86400)
+    # When False (default), Clare's rounds loop only *proposes* dispatch —
+    # spawning a session is a real external action, so it waits for the
+    # user. Set True to let Clare dispatch eligible work autonomously
+    # (still gated by the budget envelope + per-host caps).
+    autonomous_dispatch: bool = False
 
 
 class BudgetConfig(_Strict):
@@ -155,6 +160,8 @@ def _serialize(cfg: ClareConfig) -> str:
         lines.append(f"reply_timeout_s = {orch.reply_timeout_s}")
         if orch.rounds_interval_s != 0:
             lines.append(f"rounds_interval_s = {orch.rounds_interval_s}")
+        if orch.autonomous_dispatch:
+            lines.append("autonomous_dispatch = true")
     # Emit [budget] / [limits] only when non-default.
     bud = cfg.budget
     if bud.daily_token_cap != 0 or bud.low_priority_floor != 0.8:
diff --git a/src/clare/orchestrator/mcp_server.py b/src/clare/orchestrator/mcp_server.py
index 481b4be..1b7606c 100644
--- a/src/clare/orchestrator/mcp_server.py
+++ b/src/clare/orchestrator/mcp_server.py
@@ -479,6 +479,21 @@ def build_server() -> FastMCP:
             lambda c, _g: tools.fleet_load(c),
         )
 
+    @mcp.tool()
+    async def dispatch_task(task_id: str, host: str, cwd: str) -> dict[str, Any]:
+        """Spawn a Claude session to work a task on `host` at `cwd`.
+
+        Gated: refuses (with a `reason`) when over the daily token cap,
+        past the low-priority floor for non-P0/P1 work, or the host is at
+        its parallelism cap. Check `budget_status` + `fleet_load` first.
+        On success, creates an Assignment linking the task to the new
+        session. Spawning a session is a real action — only call this
+        when the user has opted into autonomous dispatch or confirmed."""
+        return await _call_tool(
+            "dispatch_task", {"task_id": task_id, "host": host, "cwd": cwd},
+            lambda c, g: tools.dispatch_task(c, g, task_id=task_id, host=host, cwd=cwd),
+        )
+
     # Send tool -------------------------------------------------------------
     @mcp.tool()
     async def send_to_session(session_ref: str, text: str) -> dict[str, Any]:
diff --git a/src/clare/orchestrator/tools.py b/src/clare/orchestrator/tools.py
index cff200f..f89d5ef 100644
--- a/src/clare/orchestrator/tools.py
+++ b/src/clare/orchestrator/tools.py
@@ -240,6 +240,7 @@ def help_text() -> dict[str, Any]:
             {"name": "set_project_status", "summary": "active/paused/done/archived — archive hides stale projects."},
             {"name": "budget_status", "summary": "Today's token burn vs daily cap — check before low-priority dispatch."},
             {"name": "fleet_load", "summary": "Live sessions per host vs cap + rate headroom — check before dispatch."},
+            {"name": "dispatch_task", "summary": "Spawn a session for a task (gated by budget + host caps)."},
         ],
     }
 
@@ -404,6 +405,35 @@ def suggest_assignments(conn: sqlite3.Connection) -> dict[str, Any]:
     return scheduler.suggest_assignments(conn, per_host_max=cap)
 
 
+def dispatch_task(
+    conn: sqlite3.Connection,
+    gen: HLCGenerator,
+    *,
+    task_id: str,
+    host: str,
+    cwd: str,
+) -> dict[str, Any]:
+    """Spawn a Claude session for a task, gated by budget + host caps.
+
+    Refuses (with a reason) when over the daily token cap, past the
+    low-priority floor for a non-P0/P1 task, or the host is at its
+    parallelism cap. On success spawns via rclaude, discovers the new
+    session, and creates the Assignment.
+    """
+    result = _wrap_service(
+        service.dispatch_task,
+        conn=conn, gen=gen, task_id=UUID(task_id), host=host, cwd=cwd,
+    )
+    return {
+        "dispatched": result.dispatched,
+        "task_id": result.task_id,
+        "reason": result.reason,
+        "session_uuid": result.session_uuid,
+        "host": result.host,
+        "assignment_id": result.assignment_id,
+    }
+
+
 def fleet_load(conn: sqlite3.Connection) -> dict[str, Any]:
     """Live-session count per host vs the per-host cap + rate headroom.
 
@@ -784,6 +814,7 @@ __all__ = [
     "create_tag",
     "get_session",
     "budget_status",
+    "dispatch_task",
     "fleet_load",
     "list_projects",
     "set_project_org",
diff --git a/src/clare/web/app.py b/src/clare/web/app.py
index 69cd72e..3cdad30 100644
--- a/src/clare/web/app.py
+++ b/src/clare/web/app.py
@@ -156,8 +156,22 @@ def create_app(
                 "   • `set_project_status name_or_id=<x> status=archived` for "
                 "anything that's clearly a sync/demo/test artifact — never "
                 "delete, just archive.\n"
-                "5. ALWAYS `report_status` with your own one-line summary.\n"
-                "6. ALWAYS Write the formatted output below to "
+                "5. CONSIDER DISPATCH — call `budget_status` and "
+                "`fleet_load`. If there is unassigned open work, pick the "
+                "top-priority task that fits the envelope (under the daily "
+                "token cap; P0/P1 only if past the low-priority floor; a "
+                "host with capacity). "
+                + (
+                    "Autonomous dispatch is ON — `dispatch_task(task_id, "
+                    "host, cwd)` it directly.\n"
+                    if cfg.orchestrator.autonomous_dispatch
+                    else "Autonomous dispatch is OFF — do NOT dispatch; "
+                    "instead surface the single best dispatch in NEEDS YOU "
+                    "as a proposal (task, host, cwd, why) for the user to "
+                    "approve.\n"
+                )
+                + "6. ALWAYS `report_status` with your own one-line summary.\n"
+                "7. ALWAYS Write the formatted output below to "
                 "`.claude/plans/clare-mobile-app-plan.md` in your cwd. This "
                 "file IS the user's HUD when they connect via "
                 "/remote-control clare from claude.ai/code or mobile.\n"
diff --git a/src/clare/web/service.py b/src/clare/web/service.py
index 4e82ef7..7ebbcde 100644
--- a/src/clare/web/service.py
+++ b/src/clare/web/service.py
@@ -990,3 +990,103 @@ def budget_status(conn: sqlite3.Connection, *, day: str | None = None) -> dict:
         "by_source": rollup["by_source"],
         "calls": rollup["calls"],
     }
+
+
+# ---------------------------------------------------------------------------
+# Dispatch — the rate/load-aware engine actually ACTS on work. Spawns a
+# Claude session for a task, gated by the budget envelope + host caps.
+# ---------------------------------------------------------------------------
+
+
+@dataclass(frozen=True)
+class DispatchResult:
+    dispatched: bool
+    task_id: str
+    reason: str                       # why refused, or "ok"
+    session_uuid: str | None = None   # the spawned session, when dispatched
+    host: str | None = None
+    assignment_id: str | None = None
+
+
+def dispatch_task(
+    conn: sqlite3.Connection,
+    gen: HLCGenerator,
+    *,
+    task_id: UUID,
+    host: str,
+    cwd: str,
+    rclaude: Rclaude | None = None,
+    discover_timeout_s: int = 30,
+) -> DispatchResult:
+    """Spawn a Claude session for `task_id` on `host` at `cwd`, gated by
+    the budget envelope and per-host load cap.
+
+    Refuses (dispatched=False, reason explains) when: the task is closed,
+    the daily token cap is hit, the task is low-priority and we're past
+    the low-priority floor, or the host is at its parallelism cap. On
+    success: spawns via rclaude, discovers the new session uuid, creates
+    the Assignment, returns it.
+    """
+    from ..config import load_or_init
+    from ..orchestrator.bootstrap import discover_session
+
+    task = read.get_task(conn, task_id)
+    if task is None:
+        raise NotFound(f"no such task: {task_id}")
+    if task.status in {TaskStatus.DONE}:
+        return DispatchResult(False, str(task_id), "task already done")
+
+    # Budget gate.
+    bs = budget_status(conn)
+    if bs["over_cap"]:
+        return DispatchResult(
+            False, str(task_id),
+            f"daily token cap reached ({bs['used_tokens']}/{bs['daily_token_cap']}) — dispatch refused",
+        )
+    if bs["over_low_priority_floor"] and task.priority > 1:
+        return DispatchResult(
+            False, str(task_id),
+            f"past low-priority floor ({bs['fraction_used']:.0%} of cap) — only P0/P1 dispatch allowed",
+        )
+
+    # Host-load gate.
+    cfg = load_or_init()
+    from .. import scheduler as _sched
+    if not _sched.host_has_capacity(conn, host, per_host_max=cfg.limits.per_host_max):
+        return DispatchResult(
+            False, str(task_id),
+            f"host {host!r} at capacity ({cfg.limits.per_host_max} live sessions)",
+        )
+
+    # Spawn + discover.
+    rcl = rclaude or Rclaude()
+    try:
+        pre_rows = rcl.list_sessions()
+    except RclaudeError:
+        pre_rows = []
+    pre_uuids = {
+        str(r.uuid) for r in pre_rows
+        if r.host == host and r.cwd and r.cwd.rstrip("/") == cwd.rstrip("/")
+    }
+    try:
+        rcl.spawn(host=host, cwd=cwd)
+    except RclaudeError as exc:
+        return DispatchResult(False, str(task_id), f"rclaude spawn failed: {exc}")
+
+    new_uuid = discover_session(
+        cwd=cwd, host=host, rclaude=rcl,
+        timeout_s=discover_timeout_s, ignore_uuids=pre_uuids,
+    )
+    if new_uuid is None:
+        return DispatchResult(
+            False, str(task_id),
+            "session spawned but not discovered within timeout",
+        )
+
+    assignment = create_assignment(
+        conn, gen, task_id=task_id, session_uuid=UUID(new_uuid),
+    )
+    return DispatchResult(
+        dispatched=True, task_id=str(task_id), reason="ok",
+        session_uuid=new_uuid, host=host, assignment_id=str(assignment.id),
+    )
diff --git a/tests/test_dispatch.py b/tests/test_dispatch.py
new file mode 100644
index 0000000..d097b04
--- /dev/null
+++ b/tests/test_dispatch.py
@@ -0,0 +1,191 @@
+"""Tests for the rate/load-gated dispatch engine."""
+
+from __future__ import annotations
+
+import uuid as _uuid
+from dataclasses import dataclass
+from uuid import UUID
+
+import pytest
+
+from clare import events as ev
+from clare.db import migrate, open_db
+from clare.hlc import HLCGenerator
+from clare import scheduler
+from clare.web import service
+
+
+def _setup() -> tuple:
+    conn = open_db(":memory:")
+    migrate(conn)
+    return conn, HLCGenerator("test-machine")
+
+
+# --- host load / caps ------------------------------------------------------
+
+
+def _alive_session(conn, gen, host: str, cwd: str) -> UUID:
+    sid = _uuid.uuid4()
+    ev.append(conn, gen, ev.SessionObserved(session_uuid=sid, host=host, cwd=cwd))
+    conn.execute("UPDATE sessions SET liveness = 'alive' WHERE uuid = ?", (str(sid),))
+    return sid
+
+
+def test_host_load_counts_only_alive() -> None:
+    conn, gen = _setup()
+    _alive_session(conn, gen, "apricot", "/a")
+    _alive_session(conn, gen, "apricot", "/b")
+    # An observed-but-not-alive session must not count.
+    dead = _uuid.uuid4()
+    ev.append(conn, gen, ev.SessionObserved(session_uuid=dead, host="apricot", cwd="/c"))
+    load = scheduler.host_load(conn)
+    assert load.get("apricot") == 2
+
+
+def test_host_has_capacity() -> None:
+    conn, gen = _setup()
+    for i in range(3):
+        _alive_session(conn, gen, "apricot", f"/p{i}")
+    assert scheduler.host_has_capacity(conn, "apricot", per_host_max=3) is False
+    assert scheduler.host_has_capacity(conn, "apricot", per_host_max=4) is True
+    assert scheduler.host_has_capacity(conn, "plum", per_host_max=3) is True
+
+
+# --- fake rclaude for dispatch ---------------------------------------------
+
+
+@dataclass
+class _FakeSessionRow:
+    host: str
+    uuid: UUID
+    cwd: str
+    mtime_epoch: int
+    snippet: str = ""
+
+
+class _FakeRclaude:
+    """spawn() records the call and makes a new session discoverable."""
+
+    def __init__(self, *, spawn_ok: bool = True) -> None:
+        self._rows: list[_FakeSessionRow] = []
+        self.spawn_calls: list[dict] = []
+        self._spawn_ok = spawn_ok
+
+    def list_sessions(self) -> list[_FakeSessionRow]:
+        return list(self._rows)
+
+    def list_tmux(self) -> list:
+        return []
+
+    def spawn(self, *, host: str, cwd: str, mcp_config: str | None = None) -> str:
+        from clare.rclaude import RclaudeError
+        if not self._spawn_ok:
+            raise RclaudeError("spawn failed")
+        self.spawn_calls.append({"host": host, "cwd": cwd})
+        self._rows.append(_FakeSessionRow(
+            host=host, uuid=_uuid.uuid4(), cwd=cwd, mtime_epoch=999,
+        ))
+        return f"claude-tester-{len(self.spawn_calls)}"
+
+
+def _project_with_task(conn, gen, *, priority: int = 2):
+    proj_id = _uuid.uuid4()
+    task_id = _uuid.uuid4()
+    ev.append(conn, gen, ev.ProjectCreated(project_id=proj_id, name="p"))
+    ev.append(conn, gen, ev.TaskAdded(
+        task_id=task_id, project_id=proj_id, title="t", priority=priority,
+    ))
+    return task_id
+
+
+def test_dispatch_success() -> None:
+    conn, gen = _setup()
+    task_id = _project_with_task(conn, gen)
+    rcl = _FakeRclaude()
+    result = service.dispatch_task(
+        conn, gen, task_id=task_id, host="plum", cwd="/work",
+        rclaude=rcl, discover_timeout_s=2,
+    )
+    assert result.dispatched is True
+    assert result.reason == "ok"
+    assert result.session_uuid is not None
+    assert result.assignment_id is not None
+    assert len(rcl.spawn_calls) == 1
+    # Assignment row exists.
+    rows = conn.execute(
+        "SELECT COUNT(*) FROM assignments WHERE task_id = ?", (str(task_id),)
+    ).fetchone()
+    assert rows[0] == 1
+
+
+def test_dispatch_refused_host_at_cap() -> None:
+    conn, gen = _setup()
+    task_id = _project_with_task(conn, gen)
+    # Fill plum to the default cap of 3.
+    for i in range(3):
+        _alive_session(conn, gen, "plum", f"/filler{i}")
+    rcl = _FakeRclaude()
+    result = service.dispatch_task(
+        conn, gen, task_id=task_id, host="plum", cwd="/work",
+        rclaude=rcl, discover_timeout_s=2,
+    )
+    assert result.dispatched is False
+    assert "capacity" in result.reason
+    assert rcl.spawn_calls == []  # never spawned
+
+
+def test_dispatch_refused_over_budget(monkeypatch, tmp_path) -> None:
+    # Point config at an isolated file with a tiny daily cap.
+    cfg_path = tmp_path / "config" / "clare" / "clare.toml"
+    monkeypatch.setenv("XDG_CONFIG_HOME", str(tmp_path / "config"))
+    cfg_path.parent.mkdir(parents=True, exist_ok=True)
+    cfg_path.write_text(
+        'machine_id = "m"\n\n[web]\nhost = "127.0.0.1"\nport = 8765\n'
+        "\n[budget]\ndaily_token_cap = 100\nlow_priority_floor = 0.8\n",
+        encoding="utf-8",
+    )
+    conn, gen = _setup()
+    task_id = _project_with_task(conn, gen, priority=3)
+    # Burn past the cap.
+    service.record_usage(conn, gen, source="nl", model="haiku",
+                         input_tokens=120, output_tokens=10)
+    rcl = _FakeRclaude()
+    result = service.dispatch_task(
+        conn, gen, task_id=task_id, host="plum", cwd="/work",
+        rclaude=rcl, discover_timeout_s=2,
+    )
+    assert result.dispatched is False
+    assert "cap" in result.reason.lower()
+    assert rcl.spawn_calls == []
+
+
+def test_dispatch_low_priority_floor(monkeypatch, tmp_path) -> None:
+    monkeypatch.setenv("XDG_CONFIG_HOME", str(tmp_path / "config"))
+    cfg_path = tmp_path / "config" / "clare" / "clare.toml"
+    cfg_path.parent.mkdir(parents=True, exist_ok=True)
+    cfg_path.write_text(
+        'machine_id = "m"\n\n[web]\nhost = "127.0.0.1"\nport = 8765\n'
+        "\n[budget]\ndaily_token_cap = 1000\nlow_priority_floor = 0.5\n",
+        encoding="utf-8",
+    )
+    conn, gen = _setup()
+    # 600/1000 used = 60% — past the 50% floor.
+    service.record_usage(conn, gen, source="nl", model="haiku",
+                         input_tokens=600, output_tokens=0)
+    rcl = _FakeRclaude()
+    # A P3 task is refused past the floor...
+    low = _project_with_task(conn, gen, priority=3)
+    r_low = service.dispatch_task(conn, gen, task_id=low, host="plum",
+                                  cwd="/w", rclaude=rcl, discover_timeout_s=2)
+    assert r_low.dispatched is False
+    assert "low-priority" in r_low.reason
+    # ...but a P0 task still goes through.
+    proj_id = _uuid.uuid4()
+    p0 = _uuid.uuid4()
+    ev.append(conn, gen, ev.TaskAdded(task_id=p0, project_id=conn.execute(
+        "SELECT project_id FROM tasks LIMIT 1").fetchone()[0] and
+        UUID(conn.execute("SELECT project_id FROM tasks LIMIT 1").fetchone()[0]),
+        title="urgent", priority=0))
+    r_p0 = service.dispatch_task(conn, gen, task_id=p0, host="plum",
+                                 cwd="/w", rclaude=rcl, discover_timeout_s=2)
+    assert r_p0.dispatched is True
diff --git a/tests/test_usage.py b/tests/test_usage.py
new file mode 100644
index 0000000..04d4155
--- /dev/null
+++ b/tests/test_usage.py
@@ -0,0 +1,69 @@
+"""Tests for migration 0007_usage — token-budget accounting."""
+
+from __future__ import annotations
+
+from clare import events as ev
+from clare.db import migrate, open_db
+from clare.hlc import HLCGenerator
+from clare.web import service
+
+
+def _setup() -> tuple:
+    conn = open_db(":memory:")
+    migrate(conn)
+    return conn, HLCGenerator("test-machine")
+
+
+def test_usage_recorded_projects() -> None:
+    conn, gen = _setup()
+    ev.append(conn, gen, ev.UsageRecorded(
+        source="nl", model="haiku",
+        input_tokens=120, output_tokens=40, cache_read_tokens=2000,
+    ))
+    row = conn.execute("SELECT * FROM usage").fetchone()
+    assert row["source"] == "nl"
+    assert row["input_tokens"] == 120
+    assert row["output_tokens"] == 40
+    assert row["cache_read_tokens"] == 2000
+    assert len(row["day"]) == 10  # YYYY-MM-DD
+
+
+def test_record_usage_service() -> None:
+    conn, gen = _setup()
+    service.record_usage(conn, gen, source="triage", model="haiku",
+                         input_tokens=50, output_tokens=10)
+    service.record_usage(conn, gen, source="nl", model="haiku",
+                         input_tokens=80, output_tokens=20)
+    from clare import read
+    roll = read.usage_rollup(conn)
+    assert roll["total_tokens"] == 160
+    assert roll["calls"] == 2
+    sources = {s["source"] for s in roll["by_source"]}
+    assert sources == {"triage", "nl"}
+
+
+def test_budget_status_no_cap() -> None:
+    conn, gen = _setup()
+    service.record_usage(conn, gen, source="nl", model="haiku",
+                         input_tokens=100, output_tokens=100)
+    b = service.budget_status(conn)
+    assert b["used_tokens"] == 200
+    assert b["daily_token_cap"] == 0   # default — unlimited
+    assert b["over_cap"] is False
+    assert b["over_low_priority_floor"] is False
+
+
+def test_usage_replay_roundtrip() -> None:
+    conn, gen = _setup()
+    for i in range(5):
+        service.record_usage(conn, gen, source="nl", model="haiku",
+                             input_tokens=10 * i, output_tokens=i)
+    before = conn.execute(
+        "SELECT COALESCE(SUM(input_tokens+output_tokens),0) FROM usage"
+    ).fetchone()[0]
+    ev.replay(conn)
+    after = conn.execute(
+        "SELECT COALESCE(SUM(input_tokens+output_tokens),0) FROM usage"
+    ).fetchone()[0]
+    assert before == after
+    assert conn.execute("SELECT COUNT(*) FROM usage").fetchone()[0] == 5