Collama

AI-powered code completion and editing for VS Code using local LLM backends

Overview • Features • Installation • Configuration • Models • Usage • Contributing

---

Overview

Collama is a VS Code extension that provides code completions, refactoring suggestions, and documentation generation. It supports multiple backends:

• Ollama - local
• OpenAI compatible — local / cloud

Status: Heavy active development — output may occasionally be unexpected.

Features

Code Completion

• Inline, multiline, and multiblock suggestions
• Uses currently opened tabs as context

Code Edits

• Generate docstrings, extract functions, refactor code
• Simplify complex code, fix syntax errors
• Manual instructions for custom edits

Chat Interface

• Multiple chat sessions with custom titles
• Temporary chat sessions — create unlisted, non-persisted sessions for quick experiments
• Send selected code/files as context with file references
• Search and attach files/folders directly from the chat input
• Send to chat from file tree — right-click files/folders in VSCode explorer
• Real-time context usage bar with automatic trimming
• Edit messages, copy sessions, scroll navigation
• Export chat history to JSON for backup and sharing
• Auto-accept all toggle for edits and creates to streamline workflow
• Summarize individual turns or entire conversations to reduce context usage
• AGENTS.md support — define custom agent rules by placing an AGENTS.md file in your project root

AI Agent with Tool Calling

• Tools for filesystem, git, and code diagnostics (see Available Tools)
• Workspace-boundary and .gitignore protection; optional read-only mode
• Confirmation flow with Accept / Accept All / Cancel-with-reason, plus a duration counter

Commit Messages

• AI-generated conventional commits from staged changes
• Accessible via command palette or Source Control view

Current Context Management

• Smart pruning of editor tabs in autocomplete to fit context
• Chat history optimization (tool results removed)
• Token counter visualization in agent-loop / total usage

Installation

Prerequisites: VS Code 1.109.0+, an Ollama or OpenAI-compatible endpoint (local or remote), and a supported code model (see Models).

Install the extension from the marketplace or build the vsix yourself, then configure an endpoint in settings. For authentication, set your API key as a bearer token — see Bearer Tokens.

Ollama (local, remote)
See Ollama installation instructions or the Docker image. Point apiEndpointCompletion / apiEndpointInstruct to your Ollama host (default: http://127.0.0.1:11434).

OpenAI / OpenAI-compatible (local, remote, cloud)
Point the endpoint settings to your server (e.g. vLLM, LiteLLM) or to https://api.openai.com for the OpenAI cloud API.

Configuration

Configure Collama via VS Code Settings (Preferences → Settings, search "collama"):

Setting	Type	Default	Description
collama.apiEndpointCompletion	string	http://127.0.0.1:11434	Endpoint for code auto-completion
collama.apiEndpointInstruct	string	http://127.0.0.1:11434	Endpoint for code edits/chat
collama.apiModelCompletion	string	qwen2.5-coder:3b	Model for code completions
collama.apiModelInstruct	string	gpt-oss:20b	Model for code edits (use instruct/base variant)
collama.apiTokenContextLenCompletion	number	4096	Context window size (tokens) for the completion model
collama.apiTokenContextLenInstruct	number	4096	Context window size (tokens) for the instruct/chat model
collama.apiTokenPredictCompletion	number	400	Max tokens to generate per completion request
collama.apiTokenPredictInstruct	number	4096	Max tokens to generate per instruct/chat request
collama.autoComplete	boolean	true	Enable auto-suggestions
collama.suggestMode	string	inline	Suggestion style: inline, multiline, or multiblock
collama.suggestDelay	number	1500	Delay (ms) before requesting completion
collama.verbosityMode	string	medium	Chat response detail: compact, medium, or detailed
collama.agenticMode	string	plain	Agent mode: plain, lite, or default
collama.enableEditTools	boolean	true	Enable edit tools (read-only mode when disabled)
collama.enableShellTool	boolean	false	Enable shell tool usage
collama.tlsRejectUnauthorized	boolean	true	Verify TLS certificates for HTTPS endpoints

Manual Token Settings

Set token limits to match your model or server configuration. Values are tokens, not characters.

• apiTokenContextLen* - available context window
• apiTokenPredict* - maximum generated tokens per request

Context length is shared by input tokens and the LLM answer, so a higher predict limit leaves less room for input/context.

Note

Check your model's maximum context window online and keep memory reservation in mind.

Bearer Tokens (API Key)

If an endpoint needs authentication, store the token via Command Palette:

• collama: Set Bearer Token (Completion)
• collama: Set Bearer Token (Instruct)

Tokens are sent as Authorization: Bearer <token> and stored in VS Code's encrypted credential storage. Run the same command with an empty value to clear one.

Models

Recommended Models

Collama is tested primarily with the Qwen Coder for Completion and liteLLM frontier models for Instructions. Small models may struggle with verbosity settings. Test it and keep the setting for the specific model.

For Code Completion (FIM - Fill In Middle)

• any qwen coder > 3b recommended

For Code Edits (Instruct/Base Models)

• any instruct model with thinking capabilities (e.g. gpt-oss:20b)
• Mid-level MoE models with dynamic quantization (e.g. unsloth's dynamic GGUF quants) gave noticeably better results than dense models of similar size
• Dense mid-size models (e.g. qwen3:14b) can struggle here — prefer a MoE/dynamic-quant variant if results are poor

Agentic Chat

• Prefer frontier models (e.g. gpt-oss:120b, glm-4.7, minimax2.7) for agentic mode.
• Mid-level MoE models with dynamic quantization also handled agentic editing well (e.g. gpt-oss:20b bf16, qwen3.6:35b, qwen3:30b — all dynamic quants)

Pure Chat

• Any chat model can be used for regular chat
• Keep Agentic mode OFF for small models — small models lack the reasoning for reliable tool calling

Model Completion Compatibility Table

Model	Tested Sizes	FIM Support	Status	Notes
codeqwen	—	⚠️	Untested	May work
qwen2.5-coder	1.5B, 3B, 7B	✅	Stable	Recommended
qwen3-coder	30B	✅	Stable	Recommended
starcoder	—	⚠️	Untested	May work
starcoder2	3B	✅	Stable	Like qwen2.5-coder

Note: tested primarily at q4 quantization (results may vary with others), and ChatML-format models are not supported — only true FIM models will work for autocomplete.

Usage

• Completions: trigger automatically after suggestDelay; press Tab to accept or Esc to dismiss.
• Code edits: select code and use collama (on selection) for docstrings, refactors, fixes, or manual edits.
• Chat: attach files/folders, manage sessions, summarize context, and toggle auto-accept for edits/creates.
• Commit messages: stage changes, then run collama: Generate Commit Message.

AI Agent Usage

If a chat already contains tool calls, switch to a fresh chat after turning Agentic OFF. Edit Tools can be turned OFF for read-only exploration. Shell Tool is OFF by default and can be enabled from the status bar menu.

• Agentic mode has been tested on vLLM (NVIDIA H200) with:
  • gpt-oss:120b
  • glm-4.7-fp8
  • minimax2.5, minimax2.7
  • deepseek-v4-flash

Warning

Small models should use chat with Agentic OFF.

Available Tools:

• Explore

  • read - Read a workspace file, optionally by line range
  • grep - Search workspace files with a regex pattern
  • glob - Find files and folders by glob pattern

• Git

  • gitLog - List commits or branches with optional filters
  • gitDiff - Show working tree, staged, or commit/branch diffs

• Code Analysis

  • diagnostics - Language-server analysis. Returns errors/warnings/hints for a file

• Decision

  • decision - Ask the user to choose between options when the right next step is ambiguous

• Edit Tools

  • edit - Replace an exact string in a workspace file with preview and confirmation
  • create - Create a new file or folder with confirmation
  • delete - Delete a file or folder with confirmation
  • notebook - Edit Jupyter notebook cells with rich diff preview support

• Shell Tool

  • shell - Run shell commands with confirmation when collama.enableShellTool is enabled; large output is written to a temp file instead of truncating

Contributing

Contributions are welcome! Here's how you can help:

1. Report Issues Open an issue
2. Submit PRs: Fork, create a feature branch, and submit a pull request