LM Studio Bionic Agent Tools Stuck on Working in 2026: Local Model Compatibility & Troubleshooting

If LM Studio Bionic can generate normal text but gets stuck on Working... when a task needs a tool, the model is not necessarily broken. Agentic work adds several compatibility layers: the selected model must support tool calling, Bionic must be able to execute the requested tool, the local model runtime must handle the model correctly, and the task must pass through Bionic's permission and coding workflow.

This guide focuses on a practical gap in the current Bionic ecosystem: diagnosing local-model tool failures without immediately replacing the model, increasing context, or blaming the GPU. It is especially useful for coding projects, file operations, shell commands and other agent workflows.

Why normal chat can work while agent tools fail

A useful first principle is therefore: separate text generation from tool execution. If the same model can answer a normal question but cannot complete a simple file-listing task, investigate the agent/tool path before changing quantization.

What Bionic actually supports

LM Studio describes Bionic as a separate agent application for coding, research and document/file work. Local models can run directly on the computer, while remote models can be reached through LM Link. The official Bionic documentation also warns that model choice should account for whether the model supports the tools and image inputs required by the task.

For coding projects, Bionic can search a repository, edit files, use Git and execute shell commands in the selected working directory. The documented workflow is to create a project, enable Allow coding, choose a folder and then select a suitable model.

Before troubleshooting: verify the model and runtime

Start with a model that is known to support tool use in your particular workflow. Do not assume that a model being labeled as an instruction or coding model automatically means every agent tool path will work.

Check the model's details in Bionic and confirm:

• The model is actually downloaded and available locally.
• The selected model fits your available GPU/unified memory or system RAM.
• The model format and runtime are supported by your installation.
• The task's required tool capabilities are appropriate for that model.
• You are testing with a current Bionic release rather than an old tutorial's configuration.

Bionic 1.1.6, released September 23, 2026, includes llama.cpp 2.43.0 extension packs and several agent-workspace improvements. LM Studio's changelog is the right place to check later compatibility fixes before diagnosing an issue against an older build.

Build the smallest possible reproduction

Do not begin with a complex coding request involving dozens of files, MCP servers and a large context. Create a small project containing one text file and ask Bionic for a simple operation.

mkdir bionic-tool-test
cd bionic-tool-test
printf "hello from bionic
" > test.txt

On Windows PowerShell:

New-Item -ItemType Directory bionic-tool-test
Set-Location bionic-tool-test
"hello from bionic" | Set-Content test.txt

In Bionic:

1. Create a new Project.
2. Enable Allow coding.
3. Select the bionic-tool-test folder as the working directory.
4. Choose your local model.
5. Ask: List the files in this project and tell me what is inside test.txt.

The expected behavior is not merely an answer claiming that the file exists. Bionic should actually inspect the working directory using its available project tools.

How to tell whether a tool call actually happened

When a request needs a tool, look at the session activity rather than only the final text. A successful agent cycle normally has a pattern similar to:

user request
  ↓
model decides to use a tool
  ↓
Bionic executes the tool
  ↓
tool result returns
  ↓
model continues reasoning
  ↓
final response

If the interface stays on Working... with no visible tool execution, the failure may be before the actual tool invocation. If a tool runs and then fails, focus on the tool result, permission boundary or target application instead.

Check the simplest tool before blaming MCP

Bionic's built-in coding workflow is a better first test than an external MCP server. If basic repository inspection works but an MCP server fails, the local model and core Bionic agent path are probably not the first suspects.

Model compatibility matters more than model size

A larger model is not automatically a better Bionic model. Agentic workflows need reliable tool selection and structured tool-call behavior in addition to general language quality.

If one local model can complete the same tool task while another gets stuck, keep the reproduction identical and compare only the model. This gives you a useful compatibility signal without claiming a universal benchmark.

A September 22, 2026 community report in the LM Studio bug tracker described a Bionic 1.1.4+3 session where a MiMo-V2.6-Distill-Qwen-9B GGUF model could generate normal text but remained on Working... when an agent tool was required. The reporter said Qwen3.5-9B GGUF completed the same file-listing task in the same project. This is a user report for a specific model/runtime combination, not evidence that the model universally fails with Bionic.

That kind of A/B test is exactly what you want: same project, same prompt, same machine, different model.

Do not increase context length as the first fix

Large context can be useful for coding, but it also increases memory pressure and can make an already marginal local setup harder to diagnose.

For a first reproduction:

• Use a small project.
• Use a short prompt.
• Use one local model.
• Avoid unnecessary MCP servers.
• Do not attach large documents.
• Keep concurrent sessions low.

Once the basic tool path works, increase context and project complexity gradually. If the problem appears only after a large context or many files are introduced, you have a much narrower diagnostic target.

Check RAM and VRAM before changing quantization

Local agent workloads consume more than the model weights alone. Context, KV cache, runtime buffers and concurrent work all affect memory.

Do not interpret a tool failure as proof of insufficient VRAM. First establish whether the model actually loaded and whether a simple tool call can be executed.

Use the LM Studio CLI to isolate the server path

LM Studio also provides a local server and CLI, which are useful when you want to determine whether an issue belongs to Bionic's agent layer or the underlying model/runtime.

npx lmstudio install-cli
lms server start

The official LM Studio developer documentation exposes OpenAI-compatible endpoints and tool calling. A simple API test can confirm that the selected model is serving responses independently of Bionic's project agent.

curl http://localhost:1234/v1/models

If the server responds and the model can generate normally through the API but Bionic's agent workflow fails, the problem is less likely to be basic model loading.

Test tool calling outside the full project

For developers, LM Studio documents tool use through its OpenAI-compatible API. You can use this to distinguish model-level tool-call behavior from Bionic's higher-level agent orchestration.

The important diagnostic question is not whether the model can produce JSON-looking text. It is whether the API returns a structured tool call that the client can execute.

Keep the test deliberately small: one tool, one argument, one deterministic response. Once that works, add another tool or move back into Bionic.

Shell approvals are a separate security boundary

A tool not running can be a safety feature rather than a failure. Bionic includes command approval mechanisms, and its Auto Review workflow uses deterministic analysis of shell commands before execution. If your request requires a command, check whether Bionic is waiting for an approval decision rather than assuming the model is frozen.

This distinction matters because automatically granting every shell command defeats one of the important protections in an agentic coding environment.

Skills can change the behavior of an agent

Bionic supports the Agent Skills format through SKILL.md files. Skills can be installed, created and used automatically when relevant, or explicitly referenced with @.

When debugging a tool problem, temporarily reduce the number of active skills and test the core task again. A skill can change instructions, introduce additional context or cause an agent to choose a different workflow. If the minimal task succeeds without a skill but fails after adding one, compare the skill's instructions and supporting files.

MCP: add it only after built-in tools work

MCP adds another layer: client configuration, server startup, transport, protocol negotiation, tool schemas and permissions. If Bionic cannot complete a simple file-listing task, adding an MCP server makes diagnosis harder.

A sensible sequence is:

1. Normal local generation.
2. Bionic project creation.
3. Built-in file inspection.
4. Built-in shell operation with approval.
5. One MCP server.
6. One MCP tool.
7. More complex agent workflows.

This layered approach lets you identify the first failing boundary rather than debugging five systems simultaneously.

Privacy: local inference is not the same as zero data movement

Bionic supports local models, remote models through LM Link and optional cloud models. If a task must remain on your own machine, explicitly choose a local model and understand which integrations are enabled.

LM Studio states that its local models run on-device, while cloud models are processed in LM Studio Secure Cloud under its stated Zero Data Retention policy. Those are different execution paths. Web search, remote model access, MCP servers and external APIs can also introduce data flows outside the local model process.

For sensitive source code, credentials and private documents:

• Prefer local inference when the task requires it.
• Audit MCP servers before connecting them.
• Do not place secrets in prompts or project files unnecessarily.
• Review shell commands before allowing them.
• Check whether the selected model is local, remote or cloud before starting a sensitive task.

When to switch models

Switch models when you have evidence that the selected model is the failing variable. A good A/B test keeps everything else constant.

A compact troubleshooting checklist

1. Update Bionic and note the exact version.
2. Confirm the model is downloaded and actually selected.
3. Run a normal chat request.
4. Create a tiny project with one file.
5. Enable Allow coding.
6. Ask Bionic to list and read that file.
7. Watch for actual tool execution rather than only final text.
8. Try a second known-compatible local model.
9. Check RAM/VRAM while reproducing the problem.
10. Test the LM Studio local API separately if needed.
11. Add skills and MCP servers only after the built-in path works.
12. Review permissions and shell approvals before changing security settings.

FAQ

Why does Bionic answer normally but stay on Working when using tools?

Normal generation and tool execution are different paths. The model may be able to produce text while its tool-calling behavior, runtime support, Bionic integration or permissions are incompatible with the requested operation.

Does Bionic require a special model for coding?

There is no single universal model requirement. Bionic's documentation recommends considering whether the selected model supports the tools required by the task, along with available memory and other capabilities. Test the model against a minimal real task rather than relying only on its marketing label.

Should I use a larger model if tool calls fail?

Not automatically. First determine whether the failure is model-specific. A smaller compatible model can be more useful than a larger model that cannot reliably complete the required tool path.

Could low VRAM cause a tool to appear stuck?

Memory pressure can cause slowdowns, load failures or unstable behavior, but a tool failure should not automatically be attributed to VRAM. Check model loading, memory use and a minimal reproduction first.

Should I add MCP when Bionic's built-in coding tools fail?

No. MCP adds another diagnostic layer. Get basic project inspection and a simple built-in tool operation working first.

Where should I look for current Bionic fixes?

Use the official Bionic changelog and LM Studio documentation first. For unresolved edge cases, the LM Studio bug tracker can provide useful community reports, but treat individual reports as evidence about a specific environment rather than universal behavior.

Official sources

• LM Studio Bionic documentation
• Bionic model selection documentation
• Bionic coding documentation
• Bionic Skills documentation
• LM Studio Developer Docs
• Bionic 1.1.6 release notes
• LM Studio bug report: MiMo-V2.6-Distill-Qwen-9B agent-tool issue

Related GyanAangan guides

• LM Studio Bionic Canvas in 2026
• LM Studio Bionic Session References & Introspection
• LM Studio Responses API: Local Models, Tools & MCP
• LM Studio API Setup 2026
• Ollama vs llama.cpp vs LM Studio vs vLLM