diff --git a/agentdescent/cli.py b/agentdescent/cli.py
index 378cea62..6684ff9c 100644
--- a/agentdescent/cli.py
+++ b/agentdescent/cli.py
@@ -2125,6 +2125,7 @@ def cmd_run(a: argparse.Namespace) -> int:
 def status_payload(run_id: Optional[str], *, store: Optional[str] = None,
                    recent_rounds: int = 3) -> Any:
     if run_id:
+        run_id = resolve_run_id(run_id, store=store)
         rd = runstore.get(run_id, store=store)
         st = rd.status().to_dict()
         st["recent_rounds"] = rd.rounds()[-recent_rounds:]
@@ -2387,6 +2388,7 @@ def _partial_show(rd: "runstore.RunDir", st: Any) -> Dict[str, Any]:
 
 def show_payload(run_id: str, *, store: Optional[str] = None, diff: bool = True,
                  max_chars: int = 40_000) -> Dict[str, Any]:
+    run_id = resolve_run_id(run_id, store=store) or run_id
     rd = runstore.get(run_id, store=store)
     st = rd.status()
     result = rd.result()
@@ -2551,6 +2553,7 @@ def resolve_apply_dest(to: Optional[str], install_to: Optional[str]) -> Optional
 
 def apply_payload(run_id: str, *, to: Optional[str] = None, store: Optional[str] = None,
                   dry_run: bool = False, backup: bool = True) -> Dict[str, Any]:
+    run_id = resolve_run_id(run_id, store=store) or run_id
     rd = runstore.get(run_id, store=store)
     result = rd.result()
     if result is None:
@@ -5426,10 +5429,33 @@ def _accept_json_after_verb(sub: "argparse._SubParsersAction") -> None:
                             help="machine-readable output")
 
 
-#: Words that stand for "the newest run in the store" wherever a run id goes (#892).
+#: Words that stand for "the newest run in the store" wherever a run id goes (#892, #896).
 LATEST_RUN_ALIASES = ("latest", "last")
 
 
+def resolve_run_id(run_id: Optional[str], *, store: Optional[str] = None
+                   ) -> Optional[str]:
+    """Map ``latest`` / ``last`` to the newest run id.
+
+    A real run directory named ``latest`` or ``last`` still wins. An empty
+    store raises :class:`~agentdescent.runstore.RunStoreError` with the same
+    wording the CLI uses (#892), so MCP tools and payload helpers match
+    ``agentdescent show latest`` instead of looking up a folder named
+    ``latest`` (#896). Any other value, including ``None``, is returned
+    unchanged.
+    """
+    if not isinstance(run_id, str) or run_id.lower() not in LATEST_RUN_ALIASES:
+        return run_id
+    if os.path.isdir(os.path.join(runstore.store_path(store), run_id)):
+        return run_id
+    runs = runstore.list_runs(store=store)
+    if not runs:
+        raise runstore.RunStoreError(
+            f"no runs under {runstore.store_path(store)} yet, so {run_id!r} has nothing "
+            "to point at; start one with `agentdescent demo` or `agentdescent evolve`")
+    return runs[0].run_id
+
+
 def _resolve_latest_run(a: argparse.Namespace) -> None:
     """Swap ``latest`` / ``last`` for the newest run id on run-id commands.
 
@@ -5440,17 +5466,9 @@ def _resolve_latest_run(a: argparse.Namespace) -> None:
                                       cmd_cancel, cmd_resume, cmd_diagnose):
         return
     rid = getattr(a, "run_id", None)
-    if not isinstance(rid, str) or rid.lower() not in LATEST_RUN_ALIASES:
-        return
-    store = getattr(a, "store", None)
-    if os.path.isdir(os.path.join(runstore.store_path(store), rid)):
-        return
-    runs = runstore.list_runs(store=store)
-    if not runs:
-        raise runstore.RunStoreError(
-            f"no runs under {runstore.store_path(store)} yet, so {rid!r} has nothing "
-            "to point at; start one with `agentdescent demo` or `agentdescent evolve`")
-    a.run_id = runs[0].run_id
+    resolved = resolve_run_id(rid, store=getattr(a, "store", None))
+    if resolved is not None:
+        a.run_id = resolved
 
 
 def main(argv: Optional[Sequence[str]] = None) -> int:
diff --git a/agentdescent/diagnose.py b/agentdescent/diagnose.py
index fa0f271d..29f6df6c 100644
--- a/agentdescent/diagnose.py
+++ b/agentdescent/diagnose.py
@@ -183,6 +183,8 @@ def diagnose_payload(run_id: str, *, store: Optional[str] = None) -> DiagnosisRe
     from .cli import show_payload
 
     payload = show_payload(run_id, store=store, diff=False)
+    # show_payload maps latest/last (#896); report the resolved id.
+    run_id = str(payload.get("run_id") or run_id)
     state = str(payload.get("state") or "")
     if state == "cancelled":
         return DiagnosisReport(
diff --git a/agentdescent/mcp.py b/agentdescent/mcp.py
index 1ed85d9c..25994fae 100644
--- a/agentdescent/mcp.py
+++ b/agentdescent/mcp.py
@@ -37,7 +37,7 @@
 from . import __version__, runstore
 from .cli import (NESTED_ENV, apply_from_payload, apply_payload, doctor_report,
                  plan_payload, require_cli_on_path, require_resume_host_on_path,
-                 show_payload, status_payload)
+                 resolve_run_id, show_payload, status_payload)
 from .evolvespec import (EvolveSpec, SpecError, budget_round_overshoot_warning,
                          compose, estimate)
 
@@ -147,24 +147,27 @@ def advanced_from_env(environ: Optional[Dict[str, str]] = None) -> bool:
         "Progress of one run (round, best held-out reward, calls, dollars if priced, "
         "state, the last few rounds, and `audit_store` when the run has one) or, with no "
         "run_id, {store, runs} where `runs` is "
-        "every run newest first and is empty when there are none. Cheap; safe to poll. "
+        "every run newest first and is empty when there are none. `run_id` may be "
+        "`latest` or `last` for the newest run (same as the CLI). Cheap; safe to poll. "
         "Summarise round-to-round deltas for the user rather than pasting JSON."),
     "show": (
         "The evolved artifact when a run is done: for a directory, the list of files, a "
         "unified diff against the original, and the plan `apply` would execute (files "
         "written, extra files, backups); for text, the evolved instruction. Also the "
-        "outcomes() histogram that says why proposals were committed or refused. Explain "
-        "what changed and why; do not paste the whole tree."),
+        "outcomes() histogram that says why proposals were committed or refused. Pass the "
+        "run_id from start, or `latest`/`last` for the newest run (same as the CLI). "
+        "Explain what changed and why; do not paste the whole tree."),
     "diagnose": (
         "Explain why a run ended where it did: solved, seed-already-solved, "
         "partial-improvement, running, "
         "cancelled, run-failed, reflector-error, scoring-mismatch, agent-failures, "
         "gate-rejected, no-improvement or unknown, with bounded evidence and the next "
-        "step. Call this after `show` when the run learned little, instead of guessing "
+        "step. `run_id` may be `latest` or `last` for the newest run (same as the CLI). "
+        "Call this after `show` when the run learned little, instead of guessing "
         "from the outcome histogram."),
     "apply": (
         "DESTRUCTIVE: write an evolved artifact back over a target. Pass `run_id` to "
-        "install a finished run's result (optional `to` / `install_to` overrides the run's target; `install_to` is the host-skills-root form for skill_library), OR "
+        "install a finished run's result (`latest`/`last` = newest run, same as the CLI; optional `to` / `install_to` overrides the run's target; `install_to` is the host-skills-root form for skill_library), OR "
         "pass `from` + `to` to install an on-disk artifact such as a campaign "
         "`final_target` -- same rules as CLI `apply --from` (from requires to; mutually "
         "exclusive with run_id; symlinks refused). Backs the original up first unless "
@@ -173,11 +176,13 @@ def advanced_from_env(environ: Optional[Dict[str, str]] = None) -> bool:
         "already matched, so nothing was written and no backup exists -- tell the user "
         "nothing changed. Otherwise tell them the backup path afterwards."),
     "cancel": (
-        "Stop a running evolution and every worker agent it started. The run keeps its "
+        "Stop a running evolution and every worker agent it started. `run_id` may be "
+        "`latest` or `last` for the newest run (same as the CLI). The run keeps its "
         "ledger and can be resumed."),
     "resume": (
         "Continue a stopped, failed or cancelled run on its existing ledger (the engine "
-        "picks up where it left off). Returns the new status."),
+        "picks up where it left off). `run_id` may be `latest` or `last` for the newest "
+        "run (same as the CLI). Returns the new status."),
     "audit_status": (
         "Read a sparse-audit JSONL: how biased the run's cheap verifier is against ground "
         "truth (`delta_hat`), how *scattered* that error is (`resid_sd` -- the number the "
@@ -680,6 +685,7 @@ def apply(self, run_id: Optional[str] = None, to: Optional[str] = None,
 
     def cancel(self, run_id: str) -> Dict[str, Any]:
         try:
+            run_id = resolve_run_id(run_id, store=self.store) or run_id
             return runstore.cancel(run_id, store=self.store).to_dict()
         except runstore.RunStoreError as e:
             return {"error": str(e)}
@@ -689,6 +695,7 @@ def resume(self, run_id: str, budget_usd: Optional[float] = None,
         try:
             # A resumed run gets the *current* bridge: the session that started
             # it is likely gone, and its address with it.
+            run_id = resolve_run_id(run_id, store=self.store) or run_id
             runstore.require_budget_pair(budget_usd, usd_per_call)
             # Same host gate as CLI resume: an absent plugin host must not
             # relaunch and report running (#370).
diff --git a/changelog.d/896-mcp-latest-run-alias.md b/changelog.d/896-mcp-latest-run-alias.md
new file mode 100644
index 00000000..f0d652d3
--- /dev/null
+++ b/changelog.d/896-mcp-latest-run-alias.md
@@ -0,0 +1,2 @@
+Fixed
+`show_payload` / `status_payload` / `apply_payload` and the MCP `show` / `status` / `apply` / `diagnose` / `cancel` / `resume` tools accept `latest` (or `last`) the same way the CLI does, so a host that asks to show the newest run no longer gets `no run 'latest'` (#896).
diff --git a/tests/test_latest_run_alias.py b/tests/test_latest_run_alias.py
index ee1fc704..9e8ed677 100644
--- a/tests/test_latest_run_alias.py
+++ b/tests/test_latest_run_alias.py
@@ -1,8 +1,10 @@
-"""`latest` / `last` stand for the newest run id on run-id commands (#892)."""
+"""`latest` / `last` stand for the newest run id on run-id commands (#892, #896)."""
 
 import json
 import os
 
+import pytest
+
 from agentdescent import cli, runstore
 
 
@@ -42,3 +44,48 @@ def test_other_ids_untouched(tmp_path, capsys):
     _mk(store, "20260101-000000-aaaaaa")
     assert cli.main(["--store", store, "status", "nope"]) == 2
     assert "no run 'nope'" in capsys.readouterr().err
+
+
+def test_payload_helpers_resolve_latest(tmp_path):
+    store = str(tmp_path / "runs")
+    _mk(store, "20260101-000000-aaaaaa")
+    _mk(store, "20260202-000000-bbbbbb")
+    shown = cli.show_payload("latest", store=store)
+    assert shown["run_id"] == "20260202-000000-bbbbbb"
+    st = cli.status_payload("LAST", store=store)
+    assert st["run_id"] == "20260202-000000-bbbbbb"
+    with pytest.raises(runstore.RunStoreError, match="20260202-000000-bbbbbb"):
+        cli.apply_payload("latest", store=store)
+
+
+def test_mcp_tools_resolve_latest(tmp_path):
+    from agentdescent.mcp import Tools
+
+    store = str(tmp_path / "runs")
+    _mk(store, "20260101-000000-aaaaaa")
+    _mk(store, "20260202-000000-bbbbbb")
+    t = Tools(store)
+    assert t.show("latest")["run_id"] == "20260202-000000-bbbbbb"
+    assert t.status("last")["run_id"] == "20260202-000000-bbbbbb"
+    assert t.diagnose("LATEST")["run_id"] == "20260202-000000-bbbbbb"
+    err = t.apply("latest")
+    assert "error" in err
+    assert "20260202-000000-bbbbbb" in err["error"]
+    assert "no run 'latest'" not in err["error"]
+    cancelled = t.cancel("latest")
+    assert cancelled.get("run_id") == "20260202-000000-bbbbbb" or (
+        "error" in cancelled and "no run 'latest'" not in cancelled["error"])
+
+
+def test_mcp_empty_store_latest_is_clear(tmp_path):
+    from agentdescent.mcp import Tools
+
+    store = str(tmp_path / "empty")
+    t = Tools(store)
+    for fn in (t.show, t.diagnose, t.cancel, t.resume, t.apply):
+        out = fn("latest")
+        assert "error" in out
+        assert "no runs under" in out["error"] and "agentdescent demo" in out["error"]
+    out = t.status("latest")
+    assert "error" in out
+    assert "no runs under" in out["error"]