Delta Code

AI-powered coding agent for the terminal. Any model. Every tool. Zero lock-in.

Delta Code connects to any LLM provider — local models via LM Studio/Ollama, OpenAI, Anthropic, Google, Azure, or any OpenAI-compatible endpoint — and gives the model a rich set of coding tools: file editing, search, shell execution, linting, testing, codebase navigation, and more.

Switch providers mid-conversation. Plan with a cloud model, implement with a local one. Your context carries over.

Install

Requires Node.js >= 20.

git clone https://github.com/tjdoomer/Delta-code.git
cd Delta-code
npm ci && npm run build
npm install -g .

Quick Start

# First run — add your provider
delta
> /model add lmstudio openai-compatible http://localhost:1234/v1

# Every run after — auto-connects to your saved provider
delta

Or use environment variables:

OPENAI_API_KEY="lm-studio" OPENAI_BASE_URL="http://localhost:1234/v1" delta

Provider Registry

Save connections once, switch between them anytime. Credentials persist in ~/.delta/providers.json.

# Add connections
/model add lmstudio openai-compatible http://localhost:1234/v1
/model add ollama openai-compatible http://localhost:11434/v1
/model add openai openai https://api.openai.com/v1 sk-your-key
/model add anthropic anthropic https://api.anthropic.com/v1 sk-ant-your-key

# Discover local models
/model refresh

# List everything
/model list

# Switch provider mid-conversation (history preserved)
/switch lmstudio
/switch anthropic

Tools

Delta ships with 20+ tools the model can invoke:

Tool	What it does
read_file	Read file contents with pagination
write_file	Create or overwrite files
replace	Precise text replacement with context matching
multi_edit	Atomic edits across multiple files
diff_preview	Preview an edit as a unified diff before applying
glob	Find files by pattern
search_file_content	Regex search across the codebase
read_file_summary	Structural skeleton — signatures only, ~95% fewer tokens
repo_map	Ranked codebase map with dependency graph and PageRank
run_shell_command	Execute shell commands
lint	Auto-detect and run project linter with structured output
run_tests	Auto-detect and run project tests
think	Structured reasoning before acting (no side effects)
checkpoint_save	Save workspace state before risky changes
checkpoint_restore	Revert to a saved checkpoint
checkpoint_list	List saved checkpoints
save_memory	Persist facts across sessions
todo_write	Track tasks within a session
web_search	Search the web (requires Tavily API key)
web_fetch	Fetch and parse URL content
sub_agent	Spawn autonomous sub-agents for complex tasks

Commands

Command	What it does
/model list	Show all available models across connections
/model set <name>	Change the active model
/model add	Save a new provider connection
/model refresh	Re-discover models from local servers
/switch <id>	Swap provider mid-conversation
/doctor	Provider health check — env vars, connectivity, model
/persona set <id>	Change writing style (see Personas below)
/checkpoint save	Save a workspace checkpoint
/checkpoint restore	Restore a checkpoint
/chat save	Save conversation
/chat list	List saved conversations
/theme	Change color theme
/help	Show all commands

Personas

Easter egg — change Delta's personality without affecting technical capability.

/persona list                # Show all personas
/persona set noir            # Hardboiled detective investigating bugs
/persona set salaryman       # Dedicated Japanese office worker
/persona set 1950s           # Mid-century company man
/persona set pirate          # Seafaring programmer
/persona set shakespeare     # The Bard writes your pull requests
/persona set drill-sergeant  # WILL NOT tolerate sloppy code
/persona set corporate       # Synergize the codebase
/persona set default         # Back to normal

Custom personas: drop a .md file in ~/.delta/personas/ — filename becomes the ID, content becomes the voice.

Diagnostics

/doctor

Validates your setup: API keys (detects placeholders), local server reachability (LM Studio, Ollama), saved connection health, and active model configuration.

Supported Providers

Provider	Connection Type	Local
LM Studio	openai-compatible	Yes
Ollama	openai-compatible	Yes
OpenAI	openai	No
Anthropic Claude	anthropic	No
Google Gemini	gemini	No
Azure OpenAI	openai-compatible	No
AWS Bedrock	openai-compatible	No
Any OpenAI-compatible	openai-compatible	Depends

Architecture

Monorepo with npm workspaces:

• packages/core — content generators, tools, indexing, provider registry
• packages/cli — Ink/React terminal UI, slash commands, auth flow
• packages/vscode-ide-companion — VS Code extension

Key internals:

• Provider adapters convert between a canonical type system and each provider's API (OpenAI, Anthropic, Google)
• Schema normalization adapts tool declarations per-provider (strict mode for OpenAI, permissive for Gemini)
• Repo indexing uses regex-based symbol extraction, import dependency graph, and PageRank for codebase navigation
• Think block passthrough surfaces reasoning traces from models like DeepSeek-R1, QwQ, Qwen3 instead of stripping them

Docs

• Commands
• Authentication
• Configuration
• Tools
• Architecture
• Troubleshooting

License

Apache-2.0. See LICENSE.

Based on Gemini CLI via QwenLM/qwen-code.