gpt2099.nu

创建者

cableheada year ago

a Nushell cross.stream extension to interact with LLMs and MCP servers

# gemini

# code-generation

概览内容工具评论

内容

gpt2099.nu

A Nushell scriptable MCP client with editable context threads stored in cross.stream

Features

Consistent API Across Models: Connect to Gemini + Search and Anthropic + Search through a single, simple interface. (Add providers easily.)
Persistent, Editable Conversations: Conversation threads are saved across sessions. Review, edit, and control your own context window — no black-box history.
Flexible Tool Integration: Connect to MCP servers to extend functionality. gpt2099.nu already rivals Claude Code for local file editing, but with full provider independence and deeper flexibility.

Built on cross.stream for event-driven processing, gpt2099.nu brings modern AI directly into your Nushell workflow — fully scriptable, fully inspectable, all in the terminal.

Getting started

Step 1.

First, install and configure cross.stream. This may take a little effort to get right, but once it’s set up, you’ll have the full cross.stream ecosystem of tools for editing and working with your context windows.

https://cablehead.github.io/xs/getting-started/installation/

After this step you should be able to run:

"as easy as" | .append abc123
.head abc123 | .cas

Step 2.

It really is easy from here.

overlay use -pr ./gpt

Step 3.

Enable your preferred provider. This stores the API key for later use:

gpt provider enable

Step 4.

Set up a milli alias for a lightweight model (try OpenAI's gpt-4.1-mini or Anthropic's claude-3-5-haiku-20241022):

gpt provider ptr milli --set

Step 5.

Give it a spin:

"hola" | gpt -p milli

The default alias scheme ranks models by relative weight: nano < milli < kilo < giga. Reasoning-optimized variants use .r. See docs/configure-providers.md for details. For more commands see docs/commands.md.

Conversation Turn Schema (`gpt.turn`)

Each turn in a thread is stored as a gpt.turn frame, with these top-level attributes in its meta record:

role
: speaker for this turn ("user", "assistant", "system")

inherited  (currently named "options")
: attributes auto-inherited down the thread (deep-merged at each turn)
    servers
    : list of MCP server names to use
    search
    : enable LLM-side search
    tool_mode
    : provider-specific tool mode
    provider_ptr (required)
    : the provider ptr to use for this turn, ("nano", "milli", "mega")

head
: thread bookmark; must be explicitly carried forward if continuity is needed

continues
: id or list of ids, links to previous turn(s) for context

cache
: ephemeral cache flag for this turn

content_type
: MIME type for content (e.g., "application/json")