Input, Network, and LLM Capabilities for Each Capture Adapter in Claude Obsidian

Claude Obsidian defines seven distinct capture adapters in config/adapters.json that declare specific input types alongside Boolean flags indicating whether network access and LLM processing are required for each ingestion method.

The claude-obsidian repository implements a modular capture system where every adapter is annotated with capability metadata. These flags—network and llm—determine whether an adapter can operate offline and whether it invokes language-model-driven pipelines for content extraction or summarization.

Capture Adapter Capability Overview

In config/adapters.json, each adapter entry specifies three core properties: the input type it handles, its network requirement, and its LLM dependency. The manifest serves as the declarative registry that claude_obsidian/capture.py consults before dispatching operations. According to the source code, adapters with "network": true require external connectivity to retrieve resources, while those with "llm": true pass content to AI services for OCR, summarization, or analysis.

Adapter-by-Adapter Capability Breakdown

Filesystem Adapter (Local File Paths)

The filesystem adapter processes documents from the visible inbox using purely local operations. As recorded at line 15 of config/adapters.json, this adapter sets both "network": false and "llm": false. It handles standard file ingestion without external API calls or AI processing.

URL Adapter (Remote HTTPS Resources)

The url adapter ingests content from remote hosts via HTTPS. The manifest at line 47 declares "network": true because the adapter must contact external servers to retrieve resources, but "llm": false since it only fetches raw data without AI transformation.

Image Adapter (Visual Content with LLM OCR)

Unlike the filesystem adapter, the image adapter (line 57) operates on local image files but requires LLM capabilities. With "network": false and "llm": true, this adapter passes PNG and JPEG files to an LLM-based OCR pipeline for text extraction while remaining fully offline for the initial file read.

PDF Adapter (Document Analysis)

The pdf adapter handles local PDF files with AI-driven processing. According to line 68 of the configuration, it sets "network": false and "llm": true. The adapter extracts text from PDFs and feeds the content to Claude for summarization and analysis without requiring internet access.

YouTube Adapter (Media Retrieval)

For video ingestion, the youtube adapter (line 79) requires network access to download audio and video streams, declaring "network": true. However, it sets "llm": false because the adapter's responsibility ends at media retrieval; any downstream transcription or analysis is handled by separate skills.

EPUB Adapter (Ebook Processing)

The epub adapter processes local ebook files with language model assistance. At line 90, the manifest shows "network": false and "llm": true, indicating that while the file is read locally, the parsed contents are handed to an LLM for structured processing and summarization.

OCR Adapter (Dedicated Text Extraction)

Specifically designed for optical character recognition, the ocr adapter (line 101) processes image files locally ("network": false) but relies entirely on LLM-backed OCR services ("llm": true) to extract text from visual data.

Implementation in the Capture Module

The claude_obsidian/capture.py file implements the capability validation logic. Before executing any capture operation, the dispatcher reads the adapter manifest and verifies that the requested operation meets the declared constraints. For example, attempting to use the url adapter without network permissions raises a validation error based on the "network": true flag defined in the JSON registry.

The module exposes capture_filesystem and capture_filesystem_batch functions that route inputs according to their detected type. While the function names suggest filesystem operations, the underlying dispatcher uses the manifest's input type definitions to select the appropriate adapter—whether that involves local file handling, remote HTTP requests, or LLM-powered document analysis.

Practical Usage Examples

The following patterns demonstrate how different capabilities affect capture operations:

from claude_obsidian.capture import capture_filesystem
from pathlib import Path

vault = Path("/path/to/vault")

# Local file: network=false, llm=false (filesystem adapter)

local_file = Path("/path/to/vault/inbox/example.txt")
result = capture_filesystem(vault, local_file)
print(result["stored_path"])   # → .raw/captured/example.txt

# Remote URL: network=true, llm=false (url adapter)

# Requires network capability enabled

result = capture_filesystem(vault, "https://example.com/data.json")

# PDF with AI analysis: network=false, llm=true (pdf adapter)

# Processes locally but sends text to LLM

pdf_path = Path("/path/to/vault/inbox/report.pdf")
result = capture_filesystem(vault, pdf_path)

# Internally extracts text and requests Claude summarization

Unit tests in tests/test_capture.py verify that each adapter respects its declared capabilities, ensuring that LLM-dependent adapters like image, pdf, epub, and ocr properly initialize AI pipelines, while network-dependent adapters like url and youtube validate connectivity before execution.

Summary

• Seven adapters are defined in config/adapters.json with explicit capability flags.
• Network capabilities are required only by the url (line 47) and youtube (line 79) adapters for remote resource retrieval.
• LLM capabilities are enabled for image (line 57), pdf (line 68), epub (line 90), and ocr (line 101) adapters to support AI-driven text extraction and analysis.
• Offline operation is possible with filesystem, image, pdf, epub, and ocr adapters since they declare "network": false.
• The capture.py module validates these flags before dispatching operations to ensure secure, capability-appropriate content ingestion.

Frequently Asked Questions

Which capture adapters require an internet connection to function?

Only the url and youtube adapters require network access according to the manifest at config/adapters.json. The url adapter needs connectivity to fetch remote HTTPS resources (line 47), while the youtube adapter downloads video and audio streams from external servers (line 79). All other adapters—including filesystem, image, pdf, epub, and ocr—operate entirely on local files with "network": false.

Do any adapters process content without using an LLM?

Yes, three adapters function without LLM processing. The filesystem adapter handles raw file moves and copies locally (line 15), the url adapter fetches remote data as-is (line 47), and the youtube adapter retrieves media streams without AI intervention (line 79). These adapters set "llm": false in the configuration registry.

How does the capture module determine which adapter to use?

The claude_obsidian/capture.py dispatcher inspects the input type and consults config/adapters.json to match the resource against adapter capabilities. For example, when receiving a local PDF path, it selects the pdf adapter (line 68), verifies that "llm": true is available, and then invokes the LLM pipeline for text extraction. The module validates both network and LLM permissions against the manifest flags before execution.

Can the OCR adapter work offline?

While the ocr adapter reads image files locally and declares "network": false (line 101), it requires "llm": true. This means it depends on a local or externally configured LLM service for text recognition. If the LLM service runs locally, the adapter operates offline; if it relies on cloud-based models, network access is effectively required despite the adapter's network flag being false.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}