A complete implementation of Claude in the terminal.
Unlike other tools that aim to support all kinds of LLMs, tclaude is designed specifically for Claude. As such, Claude-specific features like caching, Claude-native web search or code execution are implemented correctly and fully.
- Interactive chat with resumable sessions, extended thinking, and tool use
- Built-in grounded web search, code execution, and file analysis
- MCP server support (both remote and local)
- Implement any custom tool in just a few lines of Python
- Automatic caching (makes Claude up to 10x cheaper!)
pip install tclaudeThen set the ANTHROPIC_API_KEY environment variable to your Claude API key and you are good to go.
Running tclaude opens a new chat session. You can also directly pass a prompt to start a session.
tclaude "How do I make great pasta?"
# or: echo "How do I make great pasta?" | tclaude
> Great pasta starts with quality ingredients and proper technique. ...Or use an outward pipe to integrate tclaude into unix workflows
git diff --staged | tclaude "Write a commit message for this diff." | xargs -0 git commit -mUpload files with -f
tclaude -f paper.pdf "Summarize this paper."
tclaude -f cat.png "Is this a dog?"Claude will use web search and server-side code execution when the request demands it:
tclaude "Tell me the factorials from 1 through 20."
> [Uses Python to compute the answer.]
tclaude "What is the state of the art in physically based rendering?"
> [Uses web search and responds with citations.]Once you're done chatting, the session will be automatically named and saved as <session-name>.json in the working directory.
You can resume the session with tclaude -s <session-name>.json or browse past sessions with fuzzy finding via tclaude -s.
Customize where sessions are saved by passing --sessions-dir <dir> or by setting sessions_dir in the configuration file.
Enable thinking with --thinking or by setting thinking = true in the configuration file.
tclaude --thinking "Write a quine in C++."
> [Claude thinks about how to write a quine before responding.]Several commands are available to do other things than chatting with Claude, such as /download to download files previously created by Claude. Use /help to see a list of available commands.
If you'd like to customize the behavior of Claude (e.g. tell it to be brief, or give it background information), create ~/.configs/tclaude/roles/default.md.
The content of this file will be prepended as system prompt to all conversations.
If you'd like to load different system prompts on a case-by-case basis, you can pass them as command line argument or set role in the configuration file.
tclaude --role pirate.md "How do I make great pasta?"
> Ahoy there, matey! Ye be seekin' the secrets of craftin' the finest pasta this side of the Mediterranean, eh? ...Simply implement your tool as a function in src/tclaude/tools.py and it will be callable by Claude.
Make sure to document the tools' function thoroughly such that Claude uses it optimally.
tclaude can connect to any Anthropic-compatible endpoint.
By default tclaude uses Anthropic's official API endpoint, but you can set up your own endpoint in the configuration file and then use it with tclaude --endpoint <name>.
tclaude supports two kinds of MCP servers. See the the configuration section for details on how to set up the servers.
-
Remote servers (e.g. remote-mcp-servers)
- Claude will connect directly to the server and use the tools it provides. The connection is not made by your machine.
- Remote servers are useful for tools that require a lot of resources or need to be run in a server environment.
- If the server needs authentication, it can be done via OAuth2 or a custom token.
-
Local servers (running on your machine or in an internal network)
- tclaude will connect to the MCP server via your machine and forward the tools to Claude.
- Local servers are useful for tools that require access to local resources (e.g. files on your machine).
- Two protocols are supported: STDIN (tclaude starts the server and pipes the input to it) and HTTPS (tclaude connects to the server via a URL).
You can configure tclaude by creating the ~/.configs/tclaude/tclaude.toml file.
If you have $XDG_CONFIG_HOME set, the file should be placed at $XDG_CONFIG_HOME/tclaude/tclaude.toml.
The default configuration file looks as follows. (Or use tclaude --print-default-config to print it to stdout.)
You can copy it as a starting point for your own configuration or you can only set the settings you want to differ from the defaults.
Commenta indicate optional settings such as custom endpoints or MCP servers.
Nearly all settings can be overridden via command line arguments, e.g. tclaude --thinking or tclaude --no-web-search.
Run tclaude --help to see all available arguments.
# Default configuration for the tclaude CLI tool.
# All settings can be omitted in your own config, in which case the defaults from this file will be used.
model = "claude-sonnet-4-20250514"
max_tokens = 16384
role = "default.md" # Custom system prompt file. Should be placed ~/.config/tclaude/roles/
code_execution = true
web_search = true
thinking = false
thinking_budget = "auto" # In tokens. (default: max_tokens / 2)
sessions_dir = "." # Directory for storing session data.
endpoint = "anthropic"
[endpoints.anthropic]
kind = "anthropic"
url = "https://api.anthropic.com"
api_key = "$ANTHROPIC_API_KEY"
# Example vertex endpoint
# [endpoints.custom]
# kind = "vertex"
# url = "https://aiplatform.googleapis.com/v1/projects/<VERTEX_API_PROJECT>/locations/global/publishers/anthropic/models/<MODEL>:streamRawPredict"
# api_key = "$(gcloud auth print-access-token)"
# code_execution = false # This endpoint does not support code execution
# web_search = false # This endpoint does not support web search
# Example configuration for remote Model Context Protocol (MCP) servers
# [[mcp.local_servers]]
# name = "filesystem"
# command = "npx" # command and arguments to start the MCP server
# args = [
# "-y",
# "@modelcontextprotocol/server-filesystem",
# "~", # access to the home directory
# ]
# or: url = "http://localhost:3000" # if the server is already running
# [[mcp.remote_servers]]
# name = "example-mcp"
# url = "https://example-server.modelcontextprotocol.io/sse"
# authentication = "oauth2" # opens a browser window to authenticate on first use
# or: authentication = "none"
# or: authentication = "token", authorization_token = "<your-authorization-token>"
# Optional: restrict the tools that can be used with this MCP server
# authorization_token = "<your-authorization-token>"
# tool_configuration.enabled = true
# tool_configuration.allowed_tools = [
# "example_tool_1",
# "example_tool_2",
# ]
#
# [[mcp.remote_servers]]
# name = "another-mcp"
# ...GPLv3; see LICENSE for details.