/Human generated text: Simple to use just ask Claude to use the Bionic LLM Subagent to help during tasks and it will find it a job to do. Tips: Pick a small fast model. If it fails be sure to look at what model Claude is trying to talk to in Bionic you can always tell it you’re using a different one. Output will be slow unless you have a fast machine Claude will determine a good task. /end of human generated text.

Bionic Claude

Lets Claude Code delegate coding subtasks to a local LLM running on your machine via LM Studio (the engine behind the “Bionic” desktop app), so routine work — boilerplate, first drafts, mechanical refactors, routine tests — can run offline and in parallel with Claude’s own work.

This package contains everything needed to install the skill and the configuration app on a fresh machine, or to reinstall/back up your current setup.

---

What’s in this package

BionicClaude/
├── Install.ps1        <- run this to install
├── Uninstall.ps1       <- run this to remove everything
├── README.md           <- this file
└── skill/
    ├── SKILL.md                       <- the skill Claude Code reads
    ├── assets/
    │   └── BionicClaude.ico           <- app icon (shortcut + window titlebar)
    └── scripts/
        ├── Subagent-BionicClaude.ps1  <- the config/control GUI
        ├── bionic-ask.ps1             <- what Claude Code actually calls
        ├── BionicConfig.psm1          <- shared settings module
        ├── Launch-BionicClaude.vbs    <- launches the GUI without a console window
        └── Generate-Icon.ps1          <- regenerates BionicClaude.ico if you want to tweak it

---

Prerequisites

1. Windows with PowerShell 5.1+ (built in on Windows 10/11).
2. LM Studio installed — this provides the lms CLI and the local model server that actually runs the LLM. You don’t need LM Studio’s own window open; its background server does the work.
3. Claude Code installed and set up on this machine.

You don’t need any model downloaded yet — the config app can download one for you. Git is optional — only needed if you turn on Staged-diff mode (see below), which uses it to generate diffs; without git, that mode falls back to saving the full proposed file without a diff.

---

Installing

1. Open a PowerShell window in this folder (or right-click Install.ps1 and choose “Run with PowerShell”).

2. If double-clicking doesn’t work due to execution policy, run:

   powershell -ExecutionPolicy Bypass -File Install.ps1
   

3. The installer will:

   • Check that LM Studio’s lms CLI is present (warns, but doesn’t block, if it isn’t installed yet — you can install LM Studio afterward and everything will just work).
   • Copy the skill into %USERPROFILE%\.claude\skills\bionic, which is where Claude Code looks for skills.
   • Create a config.json and transcript.log in that folder if they don’t already exist (existing ones from a previous install are never overwritten, so your saved defaults and activity history survive a reinstall).
   • Create a “Subagent BionicClaude” shortcut on your Desktop.

Re-running the installer later (e.g. after you update the files in this package) is safe — it refreshes the scripts and SKILL.md but leaves your settings and transcript alone.

---

First-time setup

1. Double-click the Subagent BionicClaude shortcut on your Desktop.
2. Click Start Server (top section). The status label turns green once LM Studio’s server is up.
3. In the Installed Models list:
   • If you already have models downloaded in LM Studio, select one and click Load Selected.
   • Otherwise click Download New… and enter a model name or search term (e.g. qwen coder or a full identifier like qwen/qwen2.5-coder-7b-instruct). The download streams progress into the log and can be cancelled mid-way.
4. In the Defaults section, optionally set a default model, port, temperature, and max tokens, then click Save Settings. Claude Code’s calls to bionic-ask.ps1 pick these up automatically for anything you don’t explicitly override.
5. Leave the window open (or reopen it any time) to watch the Bionic Activity panel at the bottom — a black-background, green-text terminal that live-updates with every task Claude sends to the local model and what it replies. Click Clear to wipe it.

---

Using it from Claude Code

You don’t need to do anything special — the skill’s description tells Claude Code to use it proactively for well-scoped, mechanical coding subtasks (boilerplate, first drafts, routine tests, mechanical refactors), and any time you explicitly say things like “ask Bionic to …” or “use the local model for …”.

Under the hood, Claude Code runs:

powershell -File "%USERPROFILE%\.claude\skills\bionic\scripts\bionic-ask.ps1" -Task "" -Files @("path\to\file.py")

The local model has no filesystem access — it only sees whatever file contents are passed in -Files, and can only propose changes back as text (Claude reviews and applies them, it never writes to disk on the local model’s say-so).

---

The config app, section by section

LM Studio Server — Start/Stop the local model server, and Refresh Status to see if it’s running and on which port. Clicking Start Server also auto-reloads whichever model was last used (by either this panel or a Claude Code task), if nothing’s already loaded and that model is still installed — you don’t need to manually reload your usual model every time.

Installed Models — Everything you’ve downloaded via LM Studio.

• Load Selected / Unload Selected — manage what’s resident in memory. Only one model is ever kept loaded at a time; loading a new one automatically unloads whatever else was loaded first (loading a model that’s already loaded is a no-op, not a duplicate).
• Unload All — safety-net button to clear everything out of memory at once.
• Download New… — fetch a new model by name or search term, with live progress and a cancel button.

Defaults used by the Bionic skill — the model, port, temperature, and max-token settings bionic-ask.ps1 falls back to when Claude Code doesn’t pass them explicitly. Also has a Staged-diff mode checkbox (off by default): when on, if the local model proposes changes to a file that already exists, Claude sees a diff instead of the full new file content, and applies it with a plain file copy instead of retyping the whole file — worth turning on if you’re doing a lot of edits to large files and want to save Claude’s tokens. New files aren’t affected either way, since there’s nothing to diff against and no doubling-up cost to avoid.

Bionic Activity — the black/green terminal panel showing a live transcript of every task sent to the local model and its response, tailing transcript.log. Purely read-only and for visibility; it doesn’t affect anything. Use Clear to wipe it.

---

Troubleshooting

“Model not found” when loading — LM Studio sometimes lists a model with a (N variant) suffix (e.g. qwen/qwen3.8-27b (1 variant)). The installed scripts already strip this automatically; if you ever see this error, it means the scripts in your live install are out of date — re-run Install.ps1 from this package.

Multiple copies of the same model showing as loaded — this used to happen because loading an already-loaded model created a second in-memory instance instead of reusing it. Fixed: the scripts in this package always check what’s loaded first and unload anything else before loading something new, so there’s only ever one model resident. If you still see duplicates from before this fix, click Unload All once to clear them out.

Server won’t start — check LM Studio is actually installed (%USERPROFILE%\.cache\lm-studio\bin\lms.exe should exist) and that nothing else is using port 1234 (change the port in Defaults if it is).

Claude says it’s sending a task but nothing ever comes back — this happened when bionic-ask.ps1 was told (via a stale hardcoded suggestion in SKILL.md) to use a model that had since been deleted; it would try to load it, fail silently, then send the actual request to a model that was never loaded and just hang. Fixed on two fronts: SKILL.md no longer tells Claude to guess or hardcode a model name (it should omit -Model unless you’ve told it which one to use), and bionic-ask.ps1 now checks that a requested model is actually installed and actually finishes loading before sending anything — if either check fails, you get an immediate, clear error instead of a silent hang.

Model name showed up truncated to one letter (e.g. “q” instead of the full name) — a PowerShell quirk: a function returning an array with exactly one item gets “unwrapped” to a plain string on return, so indexing it with [0] grabbed the first character instead of the first array element. Fixed in the model-listing helper functions across both scripts.

Claude Code isn’t using the skill — confirm the skill is in %USERPROFILE%\.claude\skills\bionic\SKILL.md (re-run Install.ps1 if not) and that Claude Code has been restarted since installing.

---

Uninstalling

Run Uninstall.ps1 the same way you ran the installer. It removes the skill folder (including your saved settings and activity log) and the desktop shortcut. It does not touch LM Studio itself or any models you’ve downloaded — those are managed independently by LM Studio.

---

Why this exists instead of automating the “Bionic” app directly

The Bionic desktop app itself is a plain Electron GUI with no CLI, API, or IPC surface — there’s nothing to script against it. The model it runs sits on top of LM Studio’s own runtime, which does expose a scriptable lms CLI and a local OpenAI-compatible REST API once its server is started. This skill talks to that server directly, so you don’t even need the Bionic app window open — LM Studio’s background server is the actual engine.