Actual Windows-MCP and Code Graph acceptance

Recorded October 23, 2026, after publication of Koder CLI 0.4.0-preview.28. The published installer bytes remain unchanged. This report replaces the earlier “upstream programs not run” status only for the combinations below. Presets remain disabled until the user enables them.

Tested combinations

Integration	Environment	Actual checks	Result
Windows-MCP 0.8.8	Disposable GitHub-hosted Windows Server 2025, x64, Python 3.14.8; public preview.28 package	Stdio initialization and discovery; PowerShell writes an owned file and an independent reader verifies it; Snapshot sees an owned WinForms window; Click and Type enter a marker verified by the window's text-change handler; the window closes normally	Passed
Koder with Windows-MCP	The same runner and unmodified public preview.28 package	Explicit license acceptance, preset add/enable, connection test, Agent discovery, actual PowerShell call, independent file readback and model continuation	Passed
CodeGraphContext 0.6.13 with Ladybug 0.19.1	Isolated Python 3.12.14 venv on Apple silicon, macOS 27.0; OpenSSL 3 libraries in the test directory; 256 MiB database buffer	Index three Python files, complete the real background job, reopen the index in another process, inspect stats, find a function, identify its caller, verify an absent symbol and reject an outside-root indexing request	Passed
Koder with Code Graph	Existing Koder host, SDK and Agent against the actual upstream process	Discover each required schema, run three real graph calls under three approvals, read the source and continue in the same task	Passed

Windows evidence: successful hosted run, revision 47a60fe524417cd628d89549125746e0621d32fa. The test uses the public package with SHA-256 309d6c159ca432538ff8b90afd4e77a01cf28d4b38aba9eb4f8e535aeb9e8672.

Both Agent tests use scripted model responses to exercise tool dispatch and continuation. They do not call a paid model API or establish a real model's planning ability. The Mac graph test uses an owned encrypted credential test port to avoid the user's keychain; it does not certify macOS native credential storage. The Windows test uses the published CLI and its normal host.

Windows setup and a first check

Use the same project directory for every command, or repeat the same -w argument. Start with an installed Python 3.14+ executable matching the OS architecture. A user-local environment avoids administrator installation:

$runtime = Join-Path $env:LOCALAPPDATA 'Koder Tools\windows-mcp-0.8.8'
py -3.14 -m venv $runtime
& "$runtime\Scripts\python.exe" -m pip install 'windows-mcp==0.8.8'
koder mcp add windows --preset windows-mcp --command "$runtime\Scripts\windows-mcp.exe"
koder mcp test windows
koder mcp enable windows

The equivalent entry is /setup → Applications and MCP → Windows-MCP. If running an extracted archive directly, read koder eula and explicitly run koder eula accept only after agreeing. The installer's KODER_EULA_ACCEPT=1 setting does not silently accept the CLI's runtime agreement.

First ask Koder to inspect its owned test window, then perform a small action there. The upstream tool names verified here are PowerShell, Snapshot, Click and Type; discover the actual schemas instead of relying on old examples named Shell or State. No administrator elevation, persistent PowerShell policy change or login task was tested or added.

Mac graph native dependency and resource limits

The first real startup failed before MCP initialization: the Ladybug wheel linked to OpenSSL 3 libraries under /opt/homebrew/opt/openssl@3/lib. The visible “C API shared library missing” fallback hid the original native import error. A fresh Ladybug-only process confirmed that the missing OpenSSL libraries caused this failure; a Kuzu/Ladybug type collision was not the cause of this run.

Provide a compatible OpenSSL 3 installation before testing this combination. With Homebrew already installed, its normal dependency command is brew install openssl@3; this is a separate user action, not part of Koder installation. A custom, trusted library directory can instead be passed to the graph child explicitly. The test used SHA-256-verified official Homebrew libraries in its own directory and changed no system library paths.

export KODER_CGC_LIBDIR='/absolute/path/to/trusted/openssl3/lib'
export KODER_CGC_BUFFER_MB=256
koder mcp add graph --preset codegraph --command /absolute/path/to/venv/bin/codegraphcontext --env-ref DYLD_LIBRARY_PATH=KODER_CGC_LIBDIR --env-ref CGC_EMBEDDED_BUFFER_POOL_MB=KODER_CGC_BUFFER_MB
koder mcp test graph
koder mcp enable graph

Use DYLD_LIBRARY_PATH only when this native dependency needs it. The graph target is still the pinned CodeGraphContext 0.6.13 program. The 256 MiB setting was sufficient for this small acceptance project; it is not a large-repository performance guarantee. A matching native-library path must reach the actual Python process; an intervening macOS system launcher can strip DYLD variables. The tested route launches the venv executable directly.

Indexing is an explicit, approved action. check_job_status reports its state under job.status; wait for actual completion. Use the actual find_code schema and analyze_code_relationships with query_type: find_callers or find_callees. Query app_tools for the specific tool name rather than loading the entire catalog. Large results return an artifact reference and app_result_read supports bounded pages. Read the source after following graph relationships.

Boundaries that remain open

Windows 10/11 user PCs, Chinese IME composition, UAC/admin actions, Code Graph on Windows/Linux, uvx cold downloads, real provider accounts, autonomous model planning, large repositories and long-running use have not been certified by this run. The static acceptance fields in the preview.28 catalog describe its publication snapshot, not a blanket certificate for every platform.

The seven held model routes remain unavailable. No test scanned user projects, used a real mailbox or API key, changed billing, modified global dependencies, interrupted another task, or silently enabled a preset. Earlier dependency failures and test-driver mistakes remain in the acceptance record; only the corrected, observed runs are marked passed.