Give your agent a URL. Let x-reader handle the source.
x-reader is a universal content reader for AI agents. It reads X/Twitter, articles, YouTube, Bilibili, podcasts, WeChat, Xiaohongshu, Telegram, RSS, and the open web, then normalizes the result into structured source content for downstream research, summarization, and automation.
Universal content reader MCP Server for 10+ platforms
星數
988
工具
—
可用率
—
Its differentiator is evidence-aware reading: a search snippet is not the source, a tweet caption is not the attached video, and a video description is not a transcript. When a material source layer is still missing, x-reader keeps that gap explicit instead of pretending the read is complete.
Your agent installs the skill and the CLI, runs the read-only x-reader doctor self-check, and reads a first link with you. The steps it follows are in install.md; to do it by hand, see below.
Evidence Receipt web utility
The public beta is live for public X, YouTube, and generic web URLs: Product · 中文 · Open app. Paste a URL and it returns the same canonical evidence states used by x-reader: PASS, PARTIAL, FAIL, or UNKNOWN. Results can be copied as a human-readable receipt or JSON, and shared without upgrading missing evidence into success.
Then give your agent a URL and ask it to read the original source. For example:
Read the original source. If it contains attached video/audio that matters to the answer, consume that too. Tell me what you actually read and what is still missing.
The canonical skill uses an explicit evidence contract:
PASS original source + all material media needed for the answer were read
PARTIAL some material layer is still missing
FAIL the original source could not be fetched reliably
UNKNOWN retrieved material is too ambiguous to verify
Evidence Receipt web preview (experimental)
Run the same evidence contract through a small local web surface:
The current hosted-MVP policy accepts public X/Twitter, YouTube, and generic web URLs only. It uses a fresh temporary HOME/inbox per request and does not inherit saved browser sessions, Telegram credentials, Groq keys, Obsidian/output paths, or external X cookies.
The web layer does not add a second reader or evidence state machine: it invokes the existing x-reader URL --json path in the isolated environment and projects the result into PASS / PARTIAL / FAIL / UNKNOWN.
See Evidence Receipt for API examples, the isolation boundary, and the mixed-media receipt contract.
v0.3.1: verified before / after
The source-fetch hardening has a live, reproducible delta against public v0.3.0:
https://example.com: v0.3.0 hit Jina HTTP 401 and failed; v0.3.1 safely falls back to direct_html_pinned and returns READ.
Mixed-media X: v0.3.0 returned the post text without structured media coverage; v0.3.1 returns media_status=present, so post text can be READ while unconsumed attached media stays PARTIAL.
A public X benchmark was run end-to-end after the Agent Skill merged to main:
Source:
https://x.com/dontbesilent/status/2103875422522077377
Post text PASS fetched from the original status via oEmbed
Attached video PASS public MP4 resolved and downloaded
Video duration ~178.3 seconds
Spoken content PASS native zh subtitle track retrieved and read
Broken URL FAIL non-zero exit + machine-readable JSON error
The public owner/repo install path was also verified from a fresh directory:
The x-reader-mcp console entrypoint is packaged inside x_reader; the root mcp_server.py remains as a backward-compatible source-checkout shim.
Tools exposed:
read_url(url)
read_batch(urls)
list_inbox()
detect_platform(url)
For Claude Code, add the server with the CLI instead of editing a Claude Desktop config:
claude mcp add x-reader -- python /absolute/path/to/x-reader/mcp_server.py
Use --scope user if you want the server available beyond the current project.
The MCP server currently targets FastMCP 1.x; the mcp and all extras intentionally pin mcp<2.
Podcast reading: metadata versus spoken audio
Podcast episode show notes are not an audio transcript. x-reader classifies supported podcast episode URLs as source_type=podcast and marks metadata-only reads PARTIAL.
For a local, explicitly requested and bounded 12-second sample using free on-device Whisper (still PARTIAL, never a full episode):
# Local CLI only (not on the hosted web reader); experimental.
python -m pip install "x-reader[media] @ git+https://github.com/runesleo/x-reader.git"
# Also install ffmpeg and yt-dlp to PATH
x-reader 'https://www.xiaoyuzhoufm.com/episode/6abb9b69195d838e2aeb9c2b' --media-preview-seconds 12 --json
For an explicit complete short-audio read (public Xiaoyuzhou only, ≤2 MiB encoded / ≤120 seconds) (experimental, local CLI only):
Short mode returns PASS only after exact HTTP 206 source bytes, SHA-256, audio/PCM durations and full-input local ASR processing are verified. Episodes exceeding the short-mode limits remain PARTIAL in that mode.
Experimental resumable long audio: The local CLI supports a separately opted-in, single public Xiaoyuzhou episode up to 64 MiB encoded and 90 minutes. It uses verified 1-MiB HTTP 206 chunks, strong-ETag source versioning, local 60-second PCM/ASR segments, temporal coverage checks, and private on-disk checkpoints:
# Local CLI only; not available on the hosted web reader.
x-reader 'https://www.xiaoyuzhoufm.com/episode/6aab4896051af796b9e966c5' --media-full-long --json
One real 277.9-second episode passed full-source proof and resumed all 5/5 source chunks plus 5/5 ASR checkpoints on repeat. Very long or inaccessible audio remains PARTIAL, not a false PASS. The cached audio chunks and ASR text persist locally in a private directory; use an isolated X_READER_LONG_CACHE_DIR for sensitive workflows. Machine transcripts can still be wrong. No background work or paid API is enabled. See media coverage, short-source proof, and the 2026-10-10 long-audio verification.