goosecode

A claude code like thingy, Vibed by GPT-5.4 with C, just for fun.

goosecode is a local AI coding agent with a Go-based TUI (terminal user interface) and C backend. It features an OpenAI-compatible API client, tool loop, slash commands, sessions, subagents, MCP support, and a terminal-first workflow.

It works with:

• local OpenAI-compatible servers such as Ollama, vLLM, llama.cpp, ik-llama, LM Studio, text-generation-webui proxies, or custom gateways
• hosted OpenAI-compatible providers such as OpenAI, Together, Fireworks, Groq, or self-hosted proxies

What It Can Do

• TUI Mode (default): Interactive terminal UI with bubbletea
  • Color-coded Plan/Build mode toggle (Tab key)
  • Scrollable chat history
  • Visible cursor with mode-specific colors
• REPL Mode (fallback): Legacy interactive REPL with multiline input
• one-shot prompt mode from the shell
• file editing and shell execution
• task tracking and plan mode
• resumable subagents and optional git worktrees
• MCP resource listing/reading
• LSP queries
• local git workflow commands like /branch, /commit, and /review
• provider presets and first-run provider setup

Current Surface Area

• Tools: 29
• Slash commands: 17

Main tools include:

• bash
• read_file, write_file, edit_file
• glob_search, grep_search
• web_fetch, web_search
• todo_write, task_create, task_get, task_list, task_update
• ask_user_question
• enter_plan_mode, exit_plan_mode
• agent
• list_mcp_resources, read_mcp_resource
• lsp
• repl, powershell

Main slash commands include:

• /help
• /model
• /provider
• /session
• /compact
• /plan
• /config
• /tasks
• /branch
• /commit
• /review
• /subagents
• /permissions
• /tools
• /exit

Build

Requirements:

• gcc with C11 support
• libcurl
• pthread support from libc

The project vendors cJSON, so you do not need to install it separately.

Debian / Ubuntu

sudo apt update
sudo apt install build-essential libcurl4-openssl-dev

macOS

brew install curl

If Homebrew curl is not on your default compiler path, set the include/library flags yourself before building.

Build Commands

make        # Build goosecode (TUI) and goosecode-backend
make tui    # Build only the TUI
make clean  # Clean build artifacts
make install
make uninstall

Build output:

• ./goosecode - TUI launcher (symlink to goosecode-tui)
• ./goosecode-tui - TUI binary (Go)
• ./goosecode-backend - C backend

User-local install path by default:

~/.local/bin/goosecode

If ~/.local/bin is on your PATH, you can then start goosecode from anywhere with:

goosecode

Custom install path example:

make install INSTALL_BINDIR=/usr/local/bin

Usage

Interactive TUI (default)

./goosecode

TUI features:

• Press Tab to toggle between PLAN and BUILD mode
• Cursor color changes: yellow (PLAN), green (BUILD)
• Separator line color matches mode
• Type messages or use slash commands (e.g., /help, /exit)

Legacy REPL mode

./goosecode --repl

One-shot Prompt

./goosecode "explain the architecture of this project"

Common CLI Flags

--provider <provider>
--model <model>
--base-url <url>
--permission <mode>
--max-turns <n>
--session <id>
--help

Example Invocations

# interactive
./goosecode

# one-shot
./goosecode "write a fibonacci function in C"

# override model for one run
./goosecode --model gpt-4o-mini "summarize this repository"

# choose a provider preset for one run
./goosecode --provider ollama

# resume a saved session
./goosecode --session 1775092052_390207107

# set permissive mode for a local sandbox session
./goosecode --permission allow

Connecting To Providers

goosecode talks to any server that exposes an OpenAI-compatible /v1 API.

Built-in Provider Presets

You can switch providers directly inside the REPL:

/provider list
/provider set openai
/provider set ollama
/provider set vllm
/provider set llama.cpp
/provider set ik-llama
/provider test
/model list
/model set <name>

/provider set ... now prompts for:

• base URL
• model name
• API key when needed or optional

Provider settings are saved per provider, so switching back to a previous provider restores its last saved base_url, model, and api_key instead of overwriting everything.

First-run behavior:

• if no environment variables or settings files are present, goosecode opens a guided provider setup flow before the REPL starts
• the REPL banner also shows the active provider, model, and base URL so it is obvious what you are connected to

Environment Variables

export OPENAI_BASE_URL=...
export OPENAI_MODEL=...
export OPENAI_API_KEY=...

OPENAI_API_KEY is optional for many local servers.

Local Providers

Use this when your model server is on the same machine or LAN.

Examples:

# local gateway on port 8083
export OPENAI_BASE_URL=http://localhost:8083/v1
export OPENAI_MODEL=cyankiwi/Qwen3.5-122B-A10B-AWQ-8bit
./goosecode

# Ollama
export OPENAI_BASE_URL=http://localhost:11434/v1
export OPENAI_MODEL=llama3
./goosecode

# vLLM
export OPENAI_BASE_URL=http://localhost:8000/v1
export OPENAI_MODEL=your-model-name
./goosecode

# llama.cpp or ik-llama
export OPENAI_BASE_URL=http://localhost:8080/v1
export OPENAI_MODEL=your-model-name
./goosecode

# or use the built-in provider presets interactively
./goosecode
# then run /provider set ollama or /provider set vllm

Notes:

• many local servers ignore OPENAI_API_KEY
• the base URL should usually end in /v1
• model names must match what your server exposes

External Providers

Use this when talking to hosted services.

Examples:

# OpenAI
export OPENAI_BASE_URL=https://api.openai.com/v1
export OPENAI_API_KEY=sk-...
export OPENAI_MODEL=gpt-4o
./goosecode

# Together / Fireworks / Groq / any compatible host
export OPENAI_BASE_URL=https://your-provider.example/v1
export OPENAI_API_KEY=...
export OPENAI_MODEL=provider-model-name
./goosecode

Notes:

• if requests fail, first confirm the endpoint is OpenAI-compatible
• if streaming behaves oddly, test a simple non-tool prompt first
• hosted providers usually require OPENAI_API_KEY

Settings Files

Configuration is loaded from:

• ~/.goosecode/settings.json
• .goosecode/settings.json

Project settings override user settings where applicable.

Example project config:

{
  "provider": "vllm",
  "base_url": "http://localhost:8083/v1",
  "model": "cyankiwi/Qwen3.5-122B-A10B-AWQ-8bit",
  "permission_mode": "allow",
  "max_turns": 64
}

User settings also keep a provider_profiles map internally so each provider can remember its own last-used values.

Permissions

Supported permission modes:

• read-only
• workspace-write
• danger-full-access
• prompt
• allow

Examples:

./goosecode --permission read-only
./goosecode --permission allow

Environment override:

export GOOSECODE_PERMS=allow

Useful REPL Notes

Editor controls:

• Left / Right: move cursor
• Up / Down: history recall
• Tab: complete slash commands
• Ctrl+A: start of line
• Ctrl+E: end of line
• Ctrl+J: insert newline into the current prompt

Examples:

/provider list
/provider set ollama
/provider test
/model list
/model set llama3
/tasks create investigate parser failure
/plan set
1. Reproduce
2. Fix
.
/review

Bash Tool Timeout

The bash tool supports a configurable timeout in seconds.

That matters for commands like:

• make
• cargo build
• docker build
• long-running test suites

Example tool call shape:

{
  "command": "docker build -t app .",
  "timeout": 1800
}

Notes:

• default timeout: 120 seconds
• maximum timeout: 7200 seconds
• both numeric and numeric-string timeout values are accepted

Sessions, Tasks, and Subagents

Examples:

/session
/tasks
/tasks create add logging around API failures
/subagents

Stored state lives under:

• ~/.goosecode/sessions
• ~/.goosecode/subagents
• ~/.goosecode/worktrees
• ~/.goosecode/todos.json

MCP and LSP

MCP

Configure MCP servers in settings:

{
  "mcp_servers": [
    {
      "name": "test",
      "command": "/usr/bin/python3",
      "args": ["/path/to/mcp_server.py"]
    }
  ]
}

Supported MCP tools today:

• list_mcp_resources
• read_mcp_resource

LSP

Supported LSP actions today:

• hover
• definition
• document_symbols

The lsp tool can use:

• default server selection for supported file types
• explicit server_command and server_args

Architecture

src/
├── main.c              # Entry point and CLI setup
├── agent.c/h           # REPL + turn loop + tool execution
├── api.c/h             # OpenAI-compatible API client
├── config.c/h          # Env + settings file loading
├── session.c/h         # Session persistence
├── permissions.c/h     # Permission checks
├── prompt.c/h          # System prompt assembly
├── commands/           # Slash commands
├── tools/              # Tool implementations
└── util/               # JSON, SSE, terminal, markdown, buffers, HTTP

Development Notes

Helpful commands:

make test
./goosecode --help
./goosecode --permission allow

When testing interactively, prefer a sandbox working directory instead of the main source tree.

License

MIT