OPTIONAL LOCAL MCP · WINDOWS

A private second opinion for Codex.

Thalen adds a read-only local reviewer to Codex. It chooses the best eligible model for the task and your hardware through Ollama or LM Studio, runs the review on your PC, and leaves Codex in charge.

Download from GitHub

Open source Zero helper telemetry Windows x64 Unsigned beta

Thalen AI

READY

LOCAL REVIEWSOn

Nothing is loaded until Codex asks.

AUTOMATIC ROUTINGTask-aware

Normal & deepQwythos 9BLM Studio

Quick / GPU busyQwen3 8BOllama

GPU STATUSNo model loadedLOW IMPACT

01 Paste one prompt02 Codex verifies the release03 You approve the install

OPENAI BUILD WEEK 2026

See the full Thalen story.

A 1:43 narrated walkthrough of the guarded install, protected Codex configuration, task-aware Ollama and LM Studio routing, passive planning, and the limits that keep local review advisory.

NarratedCaptioned1:43

Explore the verified public release →

FROM ZERO TO LOCAL REVIEW

No terminal expertise required.

Paste one guarded prompt into Codex. It verifies the exact public release before the installer runs, preserves your existing configuration, and asks before every model action.

Paste into Codex

Copy the prompt above into any Codex task.

Verify, then approve

Codex checks the owner, release, checksum, and build attestation.

Choose your local route

Use a compatible model you already have or separately approve a guided Ollama download.

AUTOMATIC MODEL ROUTING

The right local model for the moment.

Thalen selects only installed, validated models that fit the task, available GPU memory, Windows pressure, and active workloads. It plans passively before inference and never treats a catalog listing as permission to download.

Qwythos for deeper reviewA supported LM Studio route for normal and deep work when eligible.

Smaller Ollama routes for quick workUseful when the task is light or the GPU is busy.

Automatic refusal under pressureOptional review skips instead of interrupting emulators or graphics workloads.

REGISTERED LOCAL ROUTESUPPORTED

LM Studio + Qwythos

Connect an existing LM Studio installation and the catalog-supported Qwythos 9B BF16 GGUF. Thalen verifies the exact file, process ownership, and loopback endpoint before routing.

Thalen does not install LM Studio or silently acquire the GGUF.

GUIDED LOCAL ROUTESUPPORTED

Ollama + Qwen

Use an eligible audited model already installed, or approve a model download after Thalen shows its exact size, storage path, and hardware fit.

Loopback-only · one review at a time · immediate unload by default.

See the current supported model catalog +

Ollama · guided download after approval
Qwen2.5 Coder 0.5B, 1.5B, 3B, 7B, 14B, and 32B; Qwen3 8B and 14B; Qwen3 Coder 30B. Automatic eligibility still depends on the exact catalog entry, validation, hardware, and live resource pressure.

LM Studio · existing-file registration only
Qwythos 9B Claude Mythos BF16 GGUF. Thalen verifies the exact catalogued file; it does not install LM Studio or download that GGUF.

View the audited model catalog on GitHub →

SIMPLE ON PURPOSE

One switch. One route. No mystery.

The Control Center shows what is ready, which route normal work will use, whether the GPU is free, and one clear switch to pause local processing. Less-common controls stay under Advanced settings.

See the actual route

Automatic mode names Qwythos, Qwen, LM Studio, or Ollama instead of showing a stale fallback.

Stay out of the way

Cross-task locking, memory guards, low-impact mode, and immediate unload protect active work.

Preserve what exists

Backups, dry-run diffs, managed markers, idempotent updates, and rollback protect Codex files.

HARDWARE-AWARE BY DEFAULT

Built for the GPU you actually have.

Thalen measures available VRAM, system memory, storage, and workload pressure before it offers or routes a model. Entry-level systems get conservative choices; high-end local AI workstations can unlock deeper routes.

The stationary background uses a representative NVIDIA RTX 6000 Ada visual. Thalen does not require or include a specific GPU.

LOCAL ADVICE, LIMITED AUTHORITY

Helpful by design.
Unable to act.

Only review text explicitly supplied by Codex is sent to the selected loopback provider. Thalen has no helper telemetry and cannot open files, run commands, edit code, or publish anything. Codex's own OpenAI traffic remains separate.

CAN

Return bounded review suggestions
Report passive health and routing
Use verified loopback providers

CANNOT

Read repositories or credentials
Run shell, Git, or deployment tools
Apply changes or send messages

READY TO TRY IT?

Paste this into Codex.

Codex handles the technical verification and stops for your approval before anything runs.

Show the complete installation prompt +

Install the highest valid semantic prerelease of Codex GPU Thalen Helper from this exact repository:

https://github.com/rennavationstudios-bit/codex-gpu-thalen-helper

Use GitHub Releases as the only installer source. Consider only releases from that exact repository where isDraft=false and isPrerelease=true; choose the highest valid semantic prerelease with the complete required asset set. Do not use /releases/latest. Read INSTALL-WITH-CODEX.md and docs/CODEX-HANDOFF.md from that exact release tag. Verify the exact owner and repository, the installer SHA-256 against SHA256SUMS.txt, and its GitHub artifact attestation. If any proof fails, stop without running anything.

Before launching the installer, tell me the exact version, explain that this beta is not Authenticode-signed and may trigger Windows SmartScreen, summarize what you verified, and ask for my confirmation.

Preserve my Codex configuration and instructions. Use only the helper's protected backup, dry-run, managed merge, repair, and rollback paths. Never replace or hand-edit my Codex files. Preserve any existing unmarked local_gpu_reviewer unless I separately approve a reviewed migration.

The base install must not download, load, validate, move, or delete a model. Before each model action, show the exact model, approximate size, storage path, provider, and whether inference will run, then wait for my explicit approval. For Ollama, use only helper-guided catalog actions after that approval. For LM Studio, never install LM Studio or download, copy, move, or substitute a GGUF; register only an already-existing catalog-supported exact file through the helper-managed route. Keep every provider loopback-only. If exact file identity, signed process ownership, loopback safety, or unload proof fails, leave that route and the helper disabled and report why.

After installation, follow the packaged handoff checklist, tell me when Codex must restart, verify the MCP after restart, prove setup is idempotent, and report any remaining manual step. Never expose credentials, private instructions, prompts, responses, or backups.

STRAIGHT ANSWERS

Before you install.

Thalen is an independent open-source project. Current prereleases are unsigned, so verify the release and expect a possible SmartScreen warning.

What is an MCP?+

MCP is the connection Codex uses to call Thalen's narrow read-only review tools. It does not make Thalen an official OpenAI product or give the helper access to your files or shell.

Does Thalen replace Codex?+

No. Codex remains the primary agent. Thalen returns optional local advice that Codex must independently verify.

Will installation download a model?+

No. The base installer downloads and loads no model. Every later model action shows the provider, size, storage location, and whether inference will run before asking for approval.

Can I use models I already have?+

Yes, when the exact model is in the audited catalog and passes identity, validation, hardware, and loopback checks. Supported routes include Ollama models and a registered LM Studio Qwythos GGUF.

What can the reviewer access?+

Only text explicitly supplied by Codex. It has no filesystem, shell, Git, deployment, publishing, email, or remote-worker tools.

Why does SmartScreen appear?+

Current public betas are not Authenticode-signed. Codex verifies the release checksum and GitHub artifact attestation first, but Windows may still show Unknown publisher.