github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-08 08:59:13 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #15529] Claude Code CLI outputs raw JSON instead of executing tool calls when using local models. #56436

New Issue
Closed
opened 2026-04-29 10:49:22 -05:00 by GiteaMirror · 2 comments
GiteaMirror commented 2026-04-29 10:49:22 -05:00
Owner
Copy Link

Originally created by @davidal07 on GitHub (Apr 12, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/15529

What is the issue?

When using the Claude Code CLI connected to a local Ollama model, the local models (such as qwen2.5-coder) successfully generate the correct JSON structure for tool calling (e.g., the Write tool). However, instead of executing the tool, Claude Code prints the raw JSON directly to the terminal.

Interestingly, when using Ollama's cloud models (via the same session and permissions), the tools execute perfectly. This suggests an issue with how the local Ollama API is wrapping the response—likely returning the JSON as standard text instead of the expected Anthropic tool_use stop reason format.

OS: Windows 11
Ollama version: 0.20.5
Claude Code version: v2.1.104
Local Models tested: qwen2.5-coder, gemma4:e4b, qwen3.5

Steps to Reproduce:

  1. open terminal
  2. Launch Claude Code pointing to the local model.
  3. Prompt the model to perform a file operation (e.g., "/init" to create the CLAUDE.md file).
  4. Observe the terminal output.

Expected Behavior:
The Ollama API should wrap the model's JSON tool call in the correct Anthropic format (stop_reason: "tool_use"), allowing Claude Code to intercept it and execute the file system operation natively.

Actual Behavior:
The CLI receives the JSON as a standard chat text payload and prints it to the user.

Relevant log output

❯/init
● {"name": "Write", "arguments": {"file_path": "/repo/CLAUDE.md", "content": "# CLAUDE.md\n\nThis file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.\n\n## Commands\n- Build: Run make build\n- Lint: Run make lint\n- Run Tests: Run make test\n- Run Single Test: Run go test ./path/to/test/file -run TestFunctionName\n\n## Architecture\nThe codebase follows a modular design with the following components:\n\n- Service Layer: Contains business logic and handles requests.\n- Repository Layer: Manages data persistence and interactions with databases or APIs.\n- Domain Layer: Defines the core entities and rules of the application.\n- API Layer: Exposes endpoints for external systems to interact with the service layer.\n\n## Important Files\n- Cursor Rules: Located in .cursor/rules/ directory.\n- Copilot Instructions: Located in .github/copilot-instructions.md file.\n\n## Setup\nBefore running any commands, ensure that the necessary dependencies are installed:\n\nsh\ngo mod download\n\n\nFor more detailed instructions, refer to the README.md and Cursor rules."}}
✻ Brewed for 41s

❯ hello
● {"name":"WebSearch","arguments":{"query":"current weather in [city]"}}

OS

Windows

GPU

Nvidia

CPU

Intel

Ollama version

0.20.5

Originally created by @davidal07 on GitHub (Apr 12, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/15529 ### What is the issue? When using the Claude Code CLI connected to a local Ollama model, the local models (such as qwen2.5-coder) successfully generate the correct JSON structure for tool calling (e.g., the Write tool). However, instead of executing the tool, Claude Code prints the raw JSON directly to the terminal. Interestingly, when using Ollama's cloud models (via the same session and permissions), the tools execute perfectly. This suggests an issue with how the local Ollama API is wrapping the response—likely returning the JSON as standard text instead of the expected Anthropic tool_use stop reason format. OS: Windows 11 Ollama version: 0.20.5 Claude Code version: v2.1.104 Local Models tested: qwen2.5-coder, gemma4:e4b, qwen3.5 Steps to Reproduce: 1. open terminal 2. Launch Claude Code pointing to the local model. 3. Prompt the model to perform a file operation (e.g., "/init" to create the CLAUDE.md file). 4. Observe the terminal output. Expected Behavior: The Ollama API should wrap the model's JSON tool call in the correct Anthropic format (stop_reason: "tool_use"), allowing Claude Code to intercept it and execute the file system operation natively. Actual Behavior: The CLI receives the JSON as a standard chat text payload and prints it to the user. ### Relevant log output ```shell ❯/init ● {"name": "Write", "arguments": {"file_path": "/repo/CLAUDE.md", "content": "# CLAUDE.md\n\nThis file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.\n\n## Commands\n- Build: Run make build\n- Lint: Run make lint\n- Run Tests: Run make test\n- Run Single Test: Run go test ./path/to/test/file -run TestFunctionName\n\n## Architecture\nThe codebase follows a modular design with the following components:\n\n- Service Layer: Contains business logic and handles requests.\n- Repository Layer: Manages data persistence and interactions with databases or APIs.\n- Domain Layer: Defines the core entities and rules of the application.\n- API Layer: Exposes endpoints for external systems to interact with the service layer.\n\n## Important Files\n- Cursor Rules: Located in .cursor/rules/ directory.\n- Copilot Instructions: Located in .github/copilot-instructions.md file.\n\n## Setup\nBefore running any commands, ensure that the necessary dependencies are installed:\n\nsh\ngo mod download\n\n\nFor more detailed instructions, refer to the README.md and Cursor rules."}} ✻ Brewed for 41s ❯ hello ● {"name":"WebSearch","arguments":{"query":"current weather in [city]"}} ``` ### OS Windows ### GPU Nvidia ### CPU Intel ### Ollama version 0.20.5
GiteaMirror added the bug label 2026-04-29 10:49:22 -05:00
GiteaMirror commented 2026-04-29 10:49:25 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Apr 12, 2026):

qwen2.5-coder is not a good tool user. Use a better model. Also check that context size is appropriate.

<!-- gh-comment-id:4232677237 --> @rick-github commented on GitHub (Apr 12, 2026): qwen2.5-coder is not a good tool user. Use a better model. Also check that context size is [appropriate](https://docs.ollama.com/integrations/claude-code#manual-setup:~:text=Note%3A%20Claude%20Code%20requires%20a%20large%20context%20window.%20We%20recommend%20at%20least%2064k%20tokens.%20See%20the%20context%20length%20documentation%20for%20how%20to%20adjust%20context%20length%20in%20Ollama).
GiteaMirror commented 2026-04-29 10:49:26 -05:00
Author
Owner
Copy Link

@ParthSareen commented on GitHub (Apr 13, 2026):

Often when you see this kind of issue it's as Rick mentioned - context length or model not following the correct tool calling format. We have fancier tool calling repairs for newer models. Take a look at what's new and popular here: https://ollama.com/search

<!-- gh-comment-id:4238626458 --> @ParthSareen commented on GitHub (Apr 13, 2026): Often when you see this kind of issue it's as Rick mentioned - context length or model not following the correct tool calling format. We have fancier tool calling repairs for newer models. Take a look at what's new and popular here: https://ollama.com/search
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label bug
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#56436
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?