AgentCode

An open, multi-model agentic coding assistant — available as a CLI and VS Code extension. Inspired by Claude Code.

Works with Claude, GPT, Gemini, and any model supported by LiteLLM.

---

Choose your setup

Option A: CLI only

pip install agentcode-cli

Add your API key to a .env file in your project:

ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...        # optional
GEMINI_API_KEY=...           # optional

Run:

agentcode                          # interactive REPL
agentcode "fix the failing tests"  # one-shot mode
agentcode --model gpt-5.6-terra    # use a specific model

All flags:

Flag	Description
--model, -m	Model to use (overrides settings and AGENTCODE_MODEL)
--no-route	Disable cost-aware routing — always use the specified model
--auto-approve, -y	Skip permission prompts (use with caution)
--dir, -d	Project directory to operate in (default: current dir)
--init-settings	Create a starter .agentcode/settings.json and exit
--server	JSON stdio server mode (used by the VS Code extension)

---

Option B: VS Code Extension

The extension requires the CLI installed as a backend. Install it first.

Step 1 — Install the CLI:

pip install agentcode-cli

Step 2 — Install the extension:

• Search AgentCode in the VS Code Marketplace and install, or
• Download the latest .vsix from the GitHub releases page and install via Cmd+Shift+X → ... → Install from VSIX...

Step 3 — Add your API key: Cmd+, → search AgentCode → paste your key into agentcode.anthropicApiKey

Step 4 — Open the chat panel: Press Cmd+Shift+A

Extension settings

Setting	Description
agentcode.anthropicApiKey	Anthropic API key
agentcode.openaiApiKey	OpenAI API key
agentcode.geminiApiKey	Google Gemini API key
agentcode.model	Default model (e.g. claude-sonnet-5)
agentcode.executablePath	Path to agentcode if not on PATH
agentcode.inlineCompletions.enabled	Enable/disable inline completions (default: true)

Extension commands

Command	Shortcut	Description
AgentCode: Open	Cmd+Shift+A	Open the chat panel
AgentCode: Ask about selection	Right-click	Ask about highlighted code
AgentCode: Explain this file	Right-click	Explain the current file
AgentCode: Toggle inline completions	Command Palette	Enable or disable inline completions

Extension features

• Inline completions — AI-powered ghost-text suggestions as you type, powered by Claude Haiku. Press Tab to accept
• Live streaming — responses stream in real time
• Model picker — switch between Claude, GPT, and Gemini models from the dropdown in the header
• Active file context — your current file is sent automatically, no copy-pasting needed
• Diff viewer — edit tool calls open VS Code's native diff viewer so you can review changes before applying
• Right-click actions — ask about selected code or explain a file directly from the editor

---

Supported Models

Provider	Model	API Key
Anthropic	claude-sonnet-5 (default)	ANTHROPIC_API_KEY
Anthropic	claude-opus-5	ANTHROPIC_API_KEY
Anthropic	claude-fable-5 (opt-in, hardest long-horizon work)	ANTHROPIC_API_KEY
Anthropic	claude-haiku-4-5	ANTHROPIC_API_KEY
OpenAI	gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna	OPENAI_API_KEY
Google	gemini/gemini-3.1-pro-preview, gemini/gemini-3.6-flash, gemini/gemini-3.5-flash-lite	GEMINI_API_KEY

Any of LiteLLM's ~3,000 models works — the table above is just the routing default. Run /models <filter> in the REPL to browse what's available with current prices, then /model <name> to switch.

Keeping up with new releases

Pricing and model availability come from LiteLLM's model registry, not from hardcoded values in this repo. When a provider ships a new model:

pip install -U litellm

That's it — the new model becomes usable by name and its costs are tracked correctly, with no AgentCode release required. /models will list it.

What does not update automatically is which model serves each routing tier (light/medium/heavy) — that's a judgement call, so it stays explicit. Point a tier at a new model yourself in .agentcode/settings.json:

{
  "model": {
    "light":  "gpt-5.6-luna",
    "medium": "gpt-5.6-terra",
    "heavy":  "gpt-5.6-sol"
  }
}

If you set a model AgentCode can't price (a typo, or one newer than your LiteLLM version), it warns at startup and reports costs as $0.00 rather than guessing.

---

Cost-Aware Routing

AgentCode automatically picks the cheapest model that can handle the task:

Tier	Anthropic	OpenAI	Gemini
Light	Haiku 4.5	GPT-5.6 Luna	Gemini 3.5 Flash-Lite
Medium	Sonnet 5	GPT-5.6 Terra	Gemini 3.6 Flash
Heavy	Opus 5	GPT-5.6 Sol	Gemini 3.1 Pro

Simple questions go to cheap/fast models. Complex multi-file tasks go to powerful ones. Use --no-route to always use the specified model.

For the hardest long-horizon agent work, pick claude-fable-5 manually (CLI --model or the extension picker). It is not used by auto-routing so routine heavy tasks stay on Opus.

---

Tools

File & Shell

Tool	Description	Permission
read_file	Read file contents with line numbers	Auto
write_file	Create or overwrite a file	Ask
edit_file	Surgical find-and-replace edit	Ask
run_command	Execute a bash command	Ask
list_directory	Tree view of directory structure	Auto
search_files	Find files by glob pattern	Auto
search_text	Grep for text across files	Auto

Git

Tool	Description	Permission
git_status	Show working tree status	Auto
git_diff	Show staged or unstaged changes	Auto
git_log	Show recent commit history	Auto
git_commit	Stage files and create a commit	Ask
git_branch	List, create, or switch branches	Ask
git_push	Push commits to a remote	Ask

Subagents

Tool	Description	Permission
spawn_subagents	Run multiple agents in parallel on subtasks	Auto

Permission model: Read-only tools auto-approve. Write/execute tools and MCP tools ask before running (unless --auto-approve / -y is set). Tools on the permissions.deny list are always blocked, even with auto-approve on. Which tools fall into which bucket is configurable — see Settings below.

---

Settings

Create a settings file with:

agentcode --init-settings

This writes .agentcode/settings.json in your project. A global file at ~/.agentcode/settings.json is also read; precedence is CLI flags → project → global → built-in defaults. View the merged result anytime with /settings.

{
  "permissions": {
    "auto_approve_all": false,
    "auto_approve": ["read_file", "list_directory", "search_files",
                     "search_text", "git_status", "git_log", "git_diff",
                     "spawn_subagents"],
    "deny": []
  },
  "model": {
    "default": "claude-sonnet-5",
    "routing": true,
    "light": null, "medium": null, "heavy": null
  },
  "limits": {
    "max_file_size": 1000000,
    "max_output": 50000,
    "max_search_results": 100,
    "max_iterations": 25
  },
  "hooks": {}
}

Key	Meaning
permissions.auto_approve_all	true skips all permission prompts (like -y)
permissions.auto_approve	Tools that run without asking (default: read-only tools)
permissions.deny	Tools that are always blocked — beats every other setting
model.default	Default model string
model.routing	Enable cost-aware routing
model.light/medium/heavy	Override the model for a routing tier
limits.*	File-size, output, search, and iteration caps
hooks	Same format as hooks.json (see Hooks)

To gate an MCP tool or approve it permanently, use its full name, e.g. "deny": ["mcp__github__delete_repository"] or "auto_approve": ["mcp__filesystem__read_file"].

---

Slash Commands (CLI only)

Command	Description
/model <name>	Switch model on the fly
/models [filter]	Browse available models and live prices
/route	Show or toggle cost-aware routing
/cost	Show session cost breakdown
/mcp	Manage MCP server connections
/mcp list	Show connected servers
/mcp add <server>	Connect a server
/mcp remove <server>	Disconnect a server
/clear	Reset conversation and delete saved session
/compact	Force LLM-powered context compaction
/tokens	Show estimated token usage
/init	Create an AGENTCODE.md template
/settings	Show resolved settings
/help	Show help
/exit	Quit

---

Session Persistence

Conversations are automatically saved to .agentcode_session.json in your project directory and resumed on next launch. Use /clear to start fresh.

---

MCP Support (CLI only)

Connect to any MCP server using /mcp add:

/mcp add github       # prompts for GitHub token
/mcp add filesystem   # no credentials needed
/mcp add postgres     # prompts for connection string
/mcp add sqlite       # prompts for database path

Config is saved to .agentcode/mcp.json and reloaded on next launch.

Security notes:

• Credentials in mcp.json are stored in plaintext. AgentCode sets the file to mode 600 and adds .agentcode/ to your .gitignore automatically, but prefer scoped, revocable tokens where the service offers them.
• MCP tools ask for permission before running, like built-in write tools. Add specific tools to permissions.auto_approve or permissions.deny in settings.json to change that (see Settings).

Advanced — edit .agentcode/mcp.json directly for custom servers:

{
  "mcpServers": {
    "my-server": {
      "command": "npx",
      "args": ["-y", "@myorg/mcp-server"],
      "env": {"API_KEY": "..."}
    }
  }
}

Global config goes in ~/.agentcode/mcp.json.

---

Hooks

Run shell commands before or after any tool call. Create .agentcode/hooks.json:

{
  "post_edit_file": "prettier --write \"$AGENTCODE_PATH\"",
  "post_write_file": "prettier --write \"$AGENTCODE_PATH\"",
  "pre_run_command": "echo \"Running: $AGENTCODE_COMMAND\""
}

Supported keys: pre_<toolname>, post_<toolname>, pre_tool / post_tool (wildcard). Global hooks go in ~/.agentcode/hooks.json.

---

AGENTCODE.md

AgentCode loads project instructions from AGENTCODE.md in your project directory and injects them into the system prompt automatically. Run /init to generate a starter template.

---

Subagents

AgentCode can spawn parallel agents for independent subtasks:

"Analyze agent.py, router.py, and tools.py in parallel and summarize each one"

The agent calls spawn_subagents internally, runs up to 5 agents in parallel, and returns combined results.

Subagents inherit the session's permission settings, and their approval prompts are serialized so they never talk over each other. By default they cannot spawn subagents of their own — raise AGENTCODE_MAX_SUBAGENT_DEPTH if you want deeper nesting.

---

Architecture

┌────────────────────────┐      ┌────────────────────────┐
│      cli.py (UI)       │      │  server.py (VS Code)   │
│  REPL · slash commands │      │  JSON stdio protocol   │
│  Rich terminal UI      │      │  for the extension     │
└───────────┬────────────┘      └───────────┬────────────┘
            │                               │
┌───────────▼───────────────────────────────▼────────────┐
│                    agent.py (Brain)                    │
│  Agentic loop · context compaction · permissions       │
│  Hooks · subagents (depth-capped) · settings.py config │
│                                                        │
│   while needs_follow_up:                               │
│     1. router.py picks the model (cost-aware tiers)    │
│     2. Send messages + tools → LLM (via LiteLLM)       │
│     3. If tool_calls → execute, append, loop           │
│        (xml_tool_parser.py converts inline XML/JSON    │
│         tool calls from open-weight fine-tunes)        │
│     4. If text only  → done                            │
└───────────┬───────────────────────────────┬────────────┘
            │                               │
┌───────────▼─────────┐         ┌───────────▼────────────┐
│  tools.py (Hands)   │         │      mcp_client.py     │
│  read/write/edit    │         │  Connect to MCP        │
│  run_command        │         │  servers and expose    │
│  list/search        │         │  their tools to the    │
│  git operations     │         │  agent loop            │
│  spawn_subagents    │         └────────────────────────┘
└─────────────────────┘

---

Environment Variables

Variable	Description	Default
AGENTCODE_MODEL	Default model	claude-sonnet-5
AGENTCODE_MAX_ITERATIONS	Max tool-call iterations per turn	25
AGENTCODE_MAX_SUBAGENT_DEPTH	How deep subagents may nest	1
ANTHROPIC_API_KEY	Anthropic API key	—
OPENAI_API_KEY	OpenAI API key	—
GEMINI_API_KEY	Google Gemini API key	—

---

How to Extend

Add a new tool

1. Add the function schema to TOOL_DEFINITIONS in tools.py
2. Implement the function
3. Register it in TOOL_MAP
4. If it requires approval, omit it from permissions.auto_approve in .agentcode/settings.json

---

Publishing to PyPI

pip install build twine
python -m build
twine upload dist/*

---

License

MIT