ollama - 💡(How to fix) Fix Claude Code & Ollama Integration - Invalid tool parameters & CPU Fallback [8 comments, 3 participants]

What is the issue?

Description

When using Claude Code (CLI) with a local Ollama instance, the agent consistently fails during tool execution (e.g., entering "Plan Mode" or reading files). The model generates invalid JSON for the tool calls, leading to a loop of "Invalid tool parameters" errors.

Additionally, specific configurations cause extreme CPU spikes (100%+) and slow response times (50s+), which seems to be related to an unintended vision-processing overhead and a Flash Attention fallback.

System Environment

• OS: Linux (Docker Deployment)
• GPU: 2x NVIDIA GeForce RTX 3060 (12GB VRAM each)
• Ollama Version: 0.20.3
• Model: gemma4 (Local blob: sha256-4c27e0f5...)
• Claude Code Command: ollama launch cloude

Docker Configuration (docker-compose.yml)

    environment:
      - OLLAMA_SCHED_SPREAD=true
      - OLLAMA_NUM_CTX=32768
      - OLLAMA_FLASH_ATTENTION=0  # Setting to 1 causes 100% CPU load instead of GPU boost
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

Steps to Reproduce

Connect Claude Code to the local Ollama instance.

Provide a complex coding task (e.g., "Fix my project architecture and MQTT connection").

The agent attempts to initialize its internal "Plan Mode" tool.

The CLI returns: ⎿ Invalid tool parameters.

The model enters a loop: It apologizes for the wrong parameters and retries with the same (or similarly broken) JSON schema until the process is aborted.

Suspected Causes

Tool Parameter Formatting: The gemma4 model (likely due to its template or architecture) does not produce the exact JSON schema required by Claude Code's tool definitions.

Vision Encoder Overhead: The runner executes vision-related code (vision: encoded) for code-only prompts, which increases latency significantly and might interfere with the attention mechanism.

Flash Attention Regression: OLLAMA_FLASH_ATTENTION=1 results in a massive CPU spike. This suggests that the presence of the vision projector forces a fallback to a CPU-based attention implementation that is not optimized for large contexts.

Context Management: There is a discrepancy between the requested NUM_CTX=32768 and the actual prompt processing speed/stability wenn tools are involved.

Additional Context

The CPU load remains normal only wenn OLLAMA_FLASH_ATTENTION is disabled. However, the tool-calling issue persists regardless of this setting, preventing the agent from completing multi-step tasks.

Relevant log output

time=2026-04-07T12:23:41.133Z level=INFO source=runner.go:1290 msg=load request="{Operation:alloc ... FlashAttention:Disabled KvSize:32768 ...}"
time=2026-04-07T12:23:41.392Z level=INFO source=model.go:156 msg="vision: encoded" elapsed=147.336691ms shape="[2560 256]"
time=2026-04-07T12:23:41.579Z level=INFO source=ggml.go:494 msg="offloaded 43/43 layers to GPU"
[GIN] 2026/04/07 - 12:24:32 | 200 | 53.364869091s | 192.168.66.36 | POST "/v1/messages?beta=true"

OS

No response

GPU

No response

CPU

No response

Ollama version

No response