diff --git a/autofit/mcp/__init__.py b/autofit/mcp/__init__.py
new file mode 100644
index 000000000..62056c242
--- /dev/null
+++ b/autofit/mcp/__init__.py
@@ -0,0 +1,38 @@
+"""
+The read-only results-inspector MCP core, distributed as the optional
+``autofit[mcp]`` extra.
+
+It exposes PyAutoFit output directories to chat harnesses that cannot execute
+code (Claude Desktop, ChatGPT) as a small set of read-only tools — list fits
+ranked by evidence, read model/posterior/result summaries, view result images
+inline. It is deliberately *glue, not code*: every tool is one existing public
+aggregator call plus serialization; composing models and running searches stay
+Python-first.
+
+Layout:
+
+- ``tools`` — the plain aggregator-wrapper functions (no ``mcp`` dependency; the
+  test suite exercises them directly).
+- ``server.core_server`` — registers those functions on a FastMCP stdio server.
+- ``bootstrap`` — the launch-time config-pin + stdout protections, kept free of
+  any autofit import so a launcher can call them *before* importing the
+  autofit-backed modules.
+
+Run standalone with ``python -m autofit.mcp`` (needs the extra:
+``pip install autofit[mcp]``). Downstream servers — e.g. the PyAutoLens
+results-inspector, which adds a lens image/FITS layer — call ``core_server()``
+and register their own tools on top.
+"""
+
+
+def __getattr__(name):
+    # Lazy re-exports (PEP 562): ``autofit.mcp.core_server`` / ``autofit.mcp.png``
+    # resolve without importing ``autofit.mcp.server`` (and therefore autofit and
+    # the ``mcp`` package) at ``import autofit.mcp`` time — keeping this package
+    # import-light so a launcher can pin config before the heavy import.
+    if name in ("core_server", "png", "_png"):
+        from autofit.mcp import server
+
+        return server.core_server if name == "core_server" else server._png
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+
diff --git a/autofit/mcp/__main__.py b/autofit/mcp/__main__.py
new file mode 100644
index 000000000..55ea8672f
--- /dev/null
+++ b/autofit/mcp/__main__.py
@@ -0,0 +1,28 @@
+"""
+Standalone runner for the ``autofit[mcp]`` extra: ``python -m autofit.mcp``.
+
+Serves the seven read-only core tools over stdio. Downstream servers that add
+their own tools (e.g. the PyAutoLens lens layer) provide their own launcher and
+call :func:`autofit.mcp.server.core_server` directly.
+
+The ordering here is load-bearing: force JAX onto CPU and pin config to autofit's
+own bundled ``config/`` (so the server does not depend on the launch directory)
+*before* importing the autofit-backed server, and guard that import so no
+import-time stdout reaches the JSON-RPC channel.
+"""
+
+import contextlib
+import os
+import sys
+from pathlib import Path
+
+os.environ.setdefault("JAX_PLATFORMS", "cpu")
+
+from autofit.mcp.bootstrap import pin_config
+
+pin_config(Path(__file__).resolve().parents[1] / "config")
+
+with contextlib.redirect_stdout(sys.stderr):
+    from autofit.mcp.server import core_server
+
+core_server().run()
diff --git a/autofit/mcp/bootstrap.py b/autofit/mcp/bootstrap.py
new file mode 100644
index 000000000..f15ab4b30
--- /dev/null
+++ b/autofit/mcp/bootstrap.py
@@ -0,0 +1,53 @@
+"""
+Launch-time hardening for the MCP stdio server.
+
+This module is deliberately kept free of any ``autofit`` import so a launcher can
+call it *before* importing the autofit-backed tool modules — autofit reads its
+configuration at import time, and jax's backend probe logs to stdout during that
+import, both of which must be handled first. It imports only the standard library
+and ``autonerves``.
+"""
+
+import logging
+import sys
+from pathlib import Path
+from typing import Union
+
+
+def pin_config(config_path: Union[str, Path]) -> None:
+    """
+    Pin ``autonerves`` config to ``config_path`` instead of letting it default to
+    ``os.getcwd()/config``.
+
+    autofit reads config at import time; if the server is launched from a foreign
+    working directory — e.g. a chat client spawning ``wsl.exe`` in a Windows
+    folder — the default would scan unrelated files and crash on import (a
+    ``desktop.ini`` trips configparser interpolation). Call this before importing
+    ``autofit.mcp.server`` (or anything else that pulls in autofit).
+    """
+    from autonerves import conf
+
+    config_path = Path(config_path)
+    conf.instance = conf.Config(
+        str(config_path), output_path=str(config_path.parent / "output")
+    )
+
+
+def route_logging_to_stderr() -> None:
+    """
+    Rebind every stdout-bound logging handler to stderr.
+
+    stdout carries the JSON-RPC channel, but autofit's logging config (loaded on
+    import) attaches stdout stream handlers — one stray log line corrupts the
+    protocol. Call this after autofit has been imported.
+    """
+    loggers = [logging.getLogger()] + [
+        logging.getLogger(name) for name in logging.root.manager.loggerDict
+    ]
+    for logger in loggers:
+        for handler in getattr(logger, "handlers", []):
+            if (
+                isinstance(handler, logging.StreamHandler)
+                and getattr(handler, "stream", None) is sys.stdout
+            ):
+                handler.setStream(sys.stderr)
diff --git a/autofit/mcp/server.py b/autofit/mcp/server.py
new file mode 100644
index 000000000..940c22d89
--- /dev/null
+++ b/autofit/mcp/server.py
@@ -0,0 +1,113 @@
+"""
+The read-only results-inspector MCP stdio server core.
+
+``core_server()`` returns a FastMCP with the seven read-only tools registered
+over ``autofit.mcp.tools``. Callers — the assistants, or ``python -m autofit.mcp``
+— register any extra tools on the returned object and call ``.run()``.
+
+Importing this module imports autofit. A launcher must therefore, *before*
+importing it: pin config with ``autofit.mcp.bootstrap.pin_config`` (so config
+does not depend on the working directory), set ``JAX_PLATFORMS=cpu`` (skips the
+jax backend probe, which logs to stdout), and guard the import with
+``contextlib.redirect_stdout(sys.stderr)`` — stdout carries the JSON-RPC channel.
+``__main__`` does exactly this for the standalone case.
+"""
+
+import io
+
+from mcp.server.fastmcp import FastMCP, Image
+
+from autofit.mcp import tools
+from autofit.mcp.bootstrap import route_logging_to_stderr
+
+
+def _png(image) -> Image:
+    buffer = io.BytesIO()
+    image.save(buffer, format="PNG")
+    return Image(data=buffer.getvalue(), format="png")
+
+
+def core_server(name: str = "pyauto-results-inspector") -> FastMCP:
+    """
+    Build a FastMCP server with the seven read-only results-inspector tools
+    registered. Downstream servers add their own tools on the returned object.
+    """
+    route_logging_to_stderr()
+
+    mcp = FastMCP(name)
+
+    @mcp.tool()
+    def list_searches(
+        directory: str,
+        sort_by: str = "log_evidence",
+        limit: int = 20,
+        completed_only: bool = False,
+    ) -> list:
+        """
+        List every model-fit found under `directory` (searched recursively): one
+        row per fit with its name, unique tag, output directory, completion state,
+        log evidence, maximum log likelihood and free-parameter count.
+
+        Rows are sorted by `sort_by` (descending; fits without that value last) —
+        use "log_evidence" for nested samplers or "max_log_likelihood" generally —
+        and truncated to `limit` (pass 0 for all). The returned `directory` of a
+        row is what the other tools take as their `directory` argument.
+        """
+        return tools.list_searches(
+            directory, sort_by=sort_by, limit=limit, completed_only=completed_only
+        )
+
+    @mcp.tool()
+    def get_model(directory: str) -> dict:
+        """
+        The model that was fitted in one search-output directory: a human-readable
+        `info` block (component classes, priors) and the full model as a dict.
+        """
+        return tools.get_model(directory)
+
+    @mcp.tool()
+    def get_result_summary(directory: str) -> str:
+        """
+        The `model.results` text for one search-output directory: the fit's own
+        summary of the maximum-likelihood model and (when the search produces
+        them) parameter estimates with errors.
+        """
+        return tools.get_result_summary(directory)
+
+    @mcp.tool()
+    def get_samples_summary(directory: str) -> dict:
+        """
+        Posterior summary for one search-output directory: log evidence (None for
+        MLE/MCMC searches without one), maximum log likelihood, the model's
+        parameter paths, and the maximum-likelihood and median-PDF parameter
+        vectors (`median_pdf_parameters` is None for MLE searches, which have no
+        PDF).
+        """
+        return tools.get_samples_summary(directory)
+
+    @mcp.tool()
+    def get_search_info(directory: str) -> dict:
+        """
+        The non-linear search used in one search-output directory: name, unique
+        tag, completion state, and the search's serialized settings.
+        """
+        return tools.get_search_info(directory)
+
+    @mcp.tool()
+    def list_images(directory: str) -> list:
+        """
+        Names of the visualization images (`image/*.png`) available in one
+        search-output directory — pass a name (without `.png`) to `fetch_image`.
+        """
+        return tools.list_images(directory)
+
+    @mcp.tool()
+    def fetch_image(directory: str, name: str = "subplot_fit") -> Image:
+        """
+        One visualization image from a search-output directory (e.g.
+        "subplot_fit"), returned inline so it renders directly in chat. Use
+        `list_images` to see what is available.
+        """
+        return _png(tools.fetch_image(directory, name=name))
+
+    return mcp
diff --git a/autofit/mcp/tools.py b/autofit/mcp/tools.py
new file mode 100644
index 000000000..21f0e2376
--- /dev/null
+++ b/autofit/mcp/tools.py
@@ -0,0 +1,147 @@
+"""
+Read-only results-inspector tools over PyAutoFit output directories.
+
+Each function is a thin wrapper over an existing public PyAutoFit aggregator
+API (`autofit.aggregator.Aggregator`, `af.SearchOutput`): argument parsing, one call, and
+JSON-friendly serialization — nothing more. Any behaviour beyond that belongs
+in PyAutoFit itself, not here ("glue, not code").
+
+This module deliberately has no MCP dependency: `server.py` registers these
+functions as MCP tools, and the test suite exercises them without the `mcp`
+package installed.
+"""
+
+import contextlib
+import json
+import sys
+from pathlib import Path
+
+from autonerves.dictable import to_dict
+
+import autofit as af
+
+# The directory-backed aggregator — af.Aggregator is the database-backed one,
+# and the alias also keeps the API audit from resolving it there.
+from autofit.aggregator import Aggregator as DirectoryAggregator
+
+
+@contextlib.contextmanager
+def _stdout_to_stderr():
+    """
+    An MCP stdio server must keep stdout clean — it carries the JSON-RPC
+    channel — but the directory aggregator prints progress to stdout,
+    so every autofit call runs with stdout redirected to stderr.
+    """
+    with contextlib.redirect_stdout(sys.stderr):
+        yield
+
+
+def _float_or_none(value):
+    try:
+        return None if value is None else float(value)
+    except (TypeError, ValueError):
+        return None
+
+
+def _search_row(search) -> dict:
+    summary = search.samples_summary
+    max_lh_sample = getattr(summary, "max_log_likelihood_sample", None)
+    return dict(
+        name=search.name,
+        unique_tag=search.unique_tag,
+        directory=str(search.directory),
+        is_complete=search.is_complete,
+        log_evidence=_float_or_none(getattr(summary, "log_evidence", None)),
+        max_log_likelihood=_float_or_none(
+            getattr(max_lh_sample, "log_likelihood", None)
+        ),
+        model_free_parameters=getattr(search.model, "prior_count", None),
+    )
+
+
+def list_searches(
+    directory: str,
+    sort_by: str = "log_evidence",
+    limit: int = 20,
+    completed_only: bool = False,
+) -> list:
+    with _stdout_to_stderr():
+        aggregator = DirectoryAggregator.from_directory(
+            directory, completed_only=completed_only
+        )
+        rows = [_search_row(search) for search in aggregator]
+    rows.sort(
+        key=lambda row: (row.get(sort_by) is not None, row.get(sort_by)),
+        reverse=True,
+    )
+    return rows[:limit] if limit else rows
+
+
+def get_model(directory: str) -> dict:
+    with _stdout_to_stderr():
+        model = af.SearchOutput(Path(directory)).model
+    return dict(info=model.info, model=to_dict(model))
+
+
+def get_result_summary(directory: str) -> str:
+    with _stdout_to_stderr():
+        return af.SearchOutput(Path(directory)).model_results
+
+
+def get_samples_summary(directory: str) -> dict:
+    with _stdout_to_stderr():
+        search_output = af.SearchOutput(Path(directory))
+        summary = search_output.samples_summary
+        if summary is None:
+            raise FileNotFoundError(
+                f"No samples summary found under {directory}/files/ — "
+                "is this a completed search output directory?"
+            )
+        model = search_output.model
+        return dict(
+            log_evidence=_float_or_none(summary.log_evidence),
+            max_log_likelihood=_float_or_none(
+                summary.max_log_likelihood_sample.log_likelihood
+            ),
+            parameter_paths=[".".join(path) for path in model.paths],
+            max_log_likelihood_parameters=[
+                float(value)
+                for value in summary.max_log_likelihood_sample.parameter_lists_for_model(
+                    model
+                )
+            ],
+            # MLE searches (e.g. LBFGS) have no PDF, so no median-PDF sample.
+            median_pdf_parameters=None
+            if summary.median_pdf_sample is None
+            else [
+                float(value)
+                for value in summary.median_pdf_sample.parameter_lists_for_model(
+                    model
+                )
+            ],
+        )
+
+
+def get_search_info(directory: str) -> dict:
+    search_json = Path(directory) / "files" / "search.json"
+    with _stdout_to_stderr():
+        search_output = af.SearchOutput(Path(directory))
+        return dict(
+            name=search_output.name,
+            unique_tag=search_output.unique_tag,
+            is_complete=search_output.is_complete,
+            search=json.loads(search_json.read_text())
+            if search_json.exists()
+            else None,
+        )
+
+
+def list_images(directory: str) -> list:
+    # Visualization outputs live under <directory>/image/ (SearchOutput.image
+    # reads there, despite its docstring saying `files/`).
+    return sorted(path.name for path in (Path(directory) / "image").glob("*.png"))
+
+
+def fetch_image(directory: str, name: str = "subplot_fit"):
+    with _stdout_to_stderr():
+        return af.SearchOutput(Path(directory)).image(name)
diff --git a/pyproject.toml b/pyproject.toml
index d55d03e09..0f1aba552 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -67,6 +67,7 @@ local_scheme = "no-local-version"
 
 [project.optional-dependencies]
 jax = ["autonerves[jax]", "optax>=0.2.5"]
+mcp = ["mcp"]
 optional = [
     "autofit[jax]",
     "astropy>=5.0",
diff --git a/test_autofit/mcp/__init__.py b/test_autofit/mcp/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/test_autofit/mcp/test_mcp_tools.py b/test_autofit/mcp/test_mcp_tools.py
new file mode 100644
index 000000000..5f1e1dc5a
--- /dev/null
+++ b/test_autofit/mcp/test_mcp_tools.py
@@ -0,0 +1,130 @@
+"""
+Tests for the results-inspector MCP tool functions (`autofit/mcp/tools.py`, the
+``autofit[mcp]`` extra).
+
+The fixture runs a real (tiny) `af.LBFGS` fit of the `af.ex.Gaussian` toy so the
+output directory always matches the current on-disk format — a frozen fixture
+directory would silently drift. LBFGS is an MLE search, so `log_evidence` is
+legitimately absent; tests assert that shape rather than skipping it.
+"""
+
+from pathlib import Path
+
+import numpy as np
+import pytest
+from PIL import Image
+
+import autofit as af
+from autonerves import conf
+
+from autofit.mcp import tools
+
+# The shipped config (autofit/config), not test_autofit/config: it keeps search
+# output files on disk (remove_files: false), whereas the test config zips them
+# away — and the read-only tools operate on real, unzipped output directories,
+# exactly as an end user's `pip install autofit` config leaves them.
+CONFIG = Path(__file__).resolve().parents[2] / "autofit" / "config"
+
+
+@pytest.fixture(scope="module")
+def output_root(tmp_path_factory):
+    root = tmp_path_factory.mktemp("output")
+    conf.instance.push(new_path=str(CONFIG), output_path=str(root))
+
+    xvalues = np.arange(100.0)
+    gaussian = af.ex.Gaussian(centre=50.0, normalization=25.0, sigma=10.0)
+    data = gaussian.model_data_from(xvalues=xvalues)
+    noise_map = np.full(fill_value=1.0, shape=data.shape)
+    analysis = af.ex.Analysis(data=data, noise_map=noise_map)
+
+    for name in ("fit_0", "fit_1"):
+        search = af.LBFGS(name=name, path_prefix="mcp_fixture")
+        search.fit(model=af.Model(af.ex.Gaussian), analysis=analysis)
+
+    return root
+
+
+@pytest.fixture(scope="module")
+def fit_directory(output_root):
+    directory = sorted(
+        marker.parent for marker in output_root.rglob(".completed")
+    )[0]
+    (directory / "image").mkdir(exist_ok=True)
+    Image.new("RGB", (32, 16), color=(200, 40, 40)).save(
+        directory / "image" / "subplot_fit.png"
+    )
+    return directory
+
+
+def test_list_searches(output_root, fit_directory):
+    rows = tools.list_searches(str(output_root), sort_by="max_log_likelihood")
+
+    assert len(rows) == 2
+    assert {row["name"] for row in rows} == {"fit_0", "fit_1"}
+    for row in rows:
+        assert row["is_complete"] is True
+        assert isinstance(row["max_log_likelihood"], float)
+        assert row["log_evidence"] is None
+        assert row["model_free_parameters"] == 3
+
+
+def test_list_searches_limit(output_root):
+    assert len(tools.list_searches(str(output_root), limit=1)) == 1
+
+
+def test_list_searches_sorts_none_last(output_root):
+    rows = tools.list_searches(str(output_root), sort_by="log_evidence")
+    assert len(rows) == 2
+
+
+def test_get_model(fit_directory):
+    result = tools.get_model(str(fit_directory))
+
+    assert "Gaussian" in result["info"]
+    assert result["model"]["class_path"].endswith("Gaussian")
+
+
+def test_get_result_summary(fit_directory):
+    text = tools.get_result_summary(str(fit_directory))
+
+    assert "Maximum Log Likelihood" in text
+
+
+def test_get_samples_summary(fit_directory):
+    summary = tools.get_samples_summary(str(fit_directory))
+
+    assert summary["log_evidence"] is None
+    assert isinstance(summary["max_log_likelihood"], float)
+    assert summary["parameter_paths"] == [
+        "centre",
+        "normalization",
+        "sigma",
+    ]
+    assert len(summary["max_log_likelihood_parameters"]) == 3
+    assert summary["median_pdf_parameters"] is None
+    assert summary["max_log_likelihood_parameters"][0] == pytest.approx(
+        50.0, abs=1.0
+    )
+
+
+def test_get_search_info(fit_directory):
+    info = tools.get_search_info(str(fit_directory))
+
+    assert info["name"] == "fit_0"
+    assert info["is_complete"] is True
+    assert info["search"] is not None
+
+
+def test_list_images(fit_directory):
+    assert "subplot_fit.png" in tools.list_images(str(fit_directory))
+
+
+def test_fetch_image(fit_directory):
+    image = tools.fetch_image(str(fit_directory), name="subplot_fit")
+
+    assert image.size == (32, 16)
+
+
+def test_fetch_image_missing(fit_directory):
+    with pytest.raises(Exception):
+        tools.fetch_image(str(fit_directory), name="no_such_image")