Claude

Metadata	Value
Category	ai
Capabilities	http, secrets, shell
Website	https://claude.ai

Returns shapes

• conversation — from get_conversation, read_conversation_cli
• conversation[] — from list_conversations, search_conversations, list_conversations_cli
• message[] — from import_conversation
• model[] — from list_models, list_models_cli

Connections

• api — Claude API — inference via the Messages API
• code — Claude Code — local CLI, uses the user’s existing auth (no API key)
• web — claude.ai — web chat history via session cookies

Readme

One skill for everything Claude. Four access modalities, one product.

Connection	File	What it does
api	claude_api.py	Inference via the Claude API (Messages endpoint)
code	claude_code.py	Inference via the local claude CLI, plus reads local Claude Code state
web	claude_web.py	Browse/search/import claude.ai chat history

Models are never hardcoded. All operations accept a model parameter that is resolved through the graph (list_models on the relevant connection populates it). See docs/specs/done/no-hardcoded-models.md for rationale.

Usage

api connection — Claude API inference

Tool	Description
list_models	Fetch the current model catalog from api.anthropic.com/v1/models
chat	Send a single Messages API request. Supports tools, system prompts, temperature. Returns raw tool_use blocks for the caller to process.

code connection — Claude Code

Uses the user’s logged-in claude binary (no API key required) AND reads local on-disk state under ~/.claude/projects/.

Tool	Description
agent	Run Claude as an agent via claude -p. Full agent loop — tool use + structured output via --mcp-config / --json-schema.
list_models_cli	List Claude models via the keychain OAuth token (no API key). Same endpoint as list_models, different auth.
list_projects	List every project directory under ~/.claude/projects/ with conversation counts and last activity.
list_conversations_cli	List local Claude Code conversations (one per JSONL transcript) as shape-native conversation[]. Optional project scope, optional limit.
read_conversation_cli	Read a full conversation transcript — returns one conversation with a nested message[] relation (content, blocks, author, published, tool calls).

Note: The code connection uses agent rather than chat because it behaves fundamentally differently from the API — it loops internally over tool calls. Both still @provides(llm) so capability routing can pick either.

The _cli suffix on list_models_cli / list_conversations_cli / read_conversation_cli is because skill tool names share a flat namespace across all .py files in the skill, and the same names are already taken by the api and web connections (for the Anthropic API model list and claude.ai web chats, respectively).

web connection — claude.ai chat history

Tool	Description
list_conversations	Browse conversations, most recent first
get_conversation	Full conversation with all messages
search_conversations	Search by title (client-side filter)
import_conversation	Import messages into graph for FTS
list_orgs	Discover orgs and capabilities
check_session	Verify cookies are valid, return identity
extract_magic_link	Parse magic link from raw email (used during login)

Setup

api connection

1. Get an API key from https://console.anthropic.com/settings/keys
2. Add credential in AgentOS Settings → Skills → Claude (API Key)

code connection

Install Claude Code and log in:

curl -fsSL https://claude.ai/install.sh | bash    # or: brew install claude-code
claude auth login                                   # opens browser for OAuth

Works with Pro/Max/Team/Enterprise subscriptions. Once logged in, claude_code.py uses that auth state directly — no key exchange with agentOS.

web connection

No setup needed if you’re logged in to claude.ai in a supported browser (Brave/Firefox). Cookie provider matchmaking extracts sessionKey automatically.