View locally
Read one explicit Claude Code or Codex JSONL file. Tool argument and output bodies are omitted; useful user and assistant text becomes reviewable.
Claude Code + Codex · JSONL transcript review
Review a Claude Code transcript.
Share the messages you choose.
Need to turn an AI coding conversation into a bug report, pull-request explanation or portfolio example? AgentShare creates a local HTML viewer for a specified JSONL file. Review sensitive-text replacements, edit messages and export a focused selection as HTML or Markdown.
Edit and select
Nothing is selected by default. Search the candidates, change their text and choose what to include in the exported document.
Export a file
The new standalone HTML or Markdown contains your selected, edited events. The browser export does not upload them.
Open your own transcript without installing anything
Use the browser tool to choose a local UTF-8 JSONL file, select Claude Code, Codex or automatic detection, and optionally add private literal terms. Processing and exports stay in your browser. Inspect the replacements, edit remaining private details, select the messages to share, and download HTML or Markdown.
Nothing is selected initially. Browser limits are 20 MiB, 100,000 Unicode code points per visible message, and 5,000 review candidates. Use the Python CLI for more candidates. Automatic rules cannot find every secret; review the final export before sharing.
Turn a local Claude Code JSONL file into a review page
Use an existing transcript you have explicitly chosen. These commands clone the source and run it without installing a Python package:
git clone https://github.com/fuxing0910-hue/agentshare.git
cd agentshare
python -m agentshare build session.jsonl --format claude --output review.htmlReplace session.jsonl with your input path. Open review.html locally, inspect and edit messages, then select them and use Export Markdown or Export HTML. Inspect the downloaded file before sharing it yourself.
For Codex, use --format codex; the default --format auto accepts supported records from either adapter. To try the workflow first with invented data:
python -m agentshare demo --output demo-review.htmlAdd your own literal replacements
Put one private, case-sensitive literal phrase per line in a local UTF-8 file. Keep that file private and use:
python -m agentshare build session.jsonl --output review.html --term-file private-terms.txtThe review page is a private intermediate file. Common-pattern replacements are best effort. They do not certify that the candidate text or export is safe to disclose.
Use this workflow from a coding agent
An agent with a shell can run the CLI above. A coding agent that supports Agent Skills can install the bundled workflow into the current project with the optional Skills CLI:
npx skills add fuxing0910-hue/agentshare --skill share-ai-transcriptsNode.js is required for this installer. The installed skill includes its Python source and template; running it requires Python 3.10+, without pip or a repository checkout.
A task you can give after installation:
“Prepare a local review of this Claude Code JSONL file. Replace common sensitive text, let me edit and select messages, and export my selection as Markdown.”
The agent prepares the candidate page; browser review and the final sharing decision remain deliberate steps. See the skill instructions or CLI, Python and function-tool integration guide. A public repository alone does not add this tool to every chat application.
Common transcript-sharing questions
How can I share only part of a Claude Code conversation?
Build a local review page from the specified JSONL file. Open it in a browser, edit the candidate text and select the messages to include. Export a new HTML or Markdown file. Nothing is selected initially, and excluded messages are not embedded in the selected export.
Can AgentShare remove API keys, emails and private paths automatically?
It applies rules for common credential, email and home-path patterns. A private UTF-8 term file can add case-sensitive literal replacements. Rules cannot find every secret or proprietary passage; inspect and edit the candidate review and final export before sharing.
Is the initial review.html ready to publish?
No. The initial review contains all supported candidate events after replacements and should stay private. Only the newly exported file contains the selected, edited events. Inspect that final file before sharing it yourself.
Does this also work with Codex transcripts or every JSONL file?
AgentShare supports documented Claude Code and Codex JSONL record shapes. Unknown records are omitted and counted; invalid JSONL fails the build. This is not a general JSONL viewer or a promise of compatibility with every past or future transcript version.
Input limits: UTF-8 JSONL, up to 20 MiB; visible user or assistant messages above 100,000 characters fail the build. Thinking, images and source metadata are omitted. See supported record shapes and the technical reference.
如何把 Claude Code 会话脱敏、精选并导出？
如果你需要把 Claude Code 或 Codex 的 JSONL 会话用于问题反馈、PR 说明或求职项目展示，AgentShare 可以先生成本地审阅页面。常见敏感模式会被替换；你逐条检查、修改文字并选择消息，最终导出新的 HTML 或 Markdown。
免安装浏览器工具
直接打开自己的 JSONL，选择 Claude Code、Codex 或自动识别，可选添加私人词条。文件在浏览器本地处理，不上传；默认选中 0 条消息。检查替换、编辑并选择后，再导出 HTML 或 Markdown。
浏览器上限为 20 MiB、每条可见消息 100,000 个 Unicode 码点、5,000 条候选消息；更多候选请使用 Python CLI。自动规则无法识别所有秘密，分享前仍需人工检查。
Python 本地流程
git clone https://github.com/fuxing0910-hue/agentshare.git
cd agentshare
python -m agentshare build session.jsonl --format claude --output review.html把输入路径替换为你明确选定的会话文件，然后在浏览器打开 review.html。初始状态不会选中任何消息。编辑并选择后，点击 HTML 或 Markdown 导出按钮，再检查下载的最终文件。
让支持 Skill 的编程代理使用
npx skills add fuxing0910-hue/agentshare --skill share-ai-transcripts安装器需要 Node.js，Skill 本身运行只需要 Python 3.10+。安装后可描述需求：“请把这个 Claude Code JSONL 会话整理成可编辑、可精选的本地审阅页，替换常见敏感内容，之后让我导出 Markdown。”
初始审阅页可以直接公开吗？
不要直接公开。初始页面含全部支持的候选消息，应留在本地；新导出文件才只包含选中且编辑后的内容。自动规则无法识别全部密钥、隐私或内部文字，分享前仍需人工检查。
支持哪些格式？
支持文档说明的 Claude Code 与 Codex JSONL 记录形状，未知记录会省略并计数；无效 JSONL 会终止生成。不处理任意 JSONL，不重放命令，也不自动搜索会话目录或上传文件。可用 --term-file private-terms.txt 追加自定义字面替换，每行一个，区分大小写。