Photo-Agent¶
The chat sidebar on every page is answered by Claude Code in headless
mode. For each message, the Control Center runs claude -p and streams the
answer into the chat.
- Conversations: the Control Center picks each conversation's id
(--session-id), and follow-ups continue it (--resume). The pencil icon
starts a new chat.
- If Claude Code no longer has a conversation, the answer starts a new one
under the same id and says so.
- Stopping: the Stop button (the square that replaces Send), or closing
the page, ends the claude process.
The toolbar under the message box:
- Paperclip: attach files. You can also paste a screenshot or drop files
onto the box.
- Types: images (PNG, JPG, GIF, WebP), which the model sees; PDFs; and
text files such as CSV, TSV, TXT, JSON, Python or notebooks.
- Limits: up to 5 files and 20 MB each per message.
- Not supported: Excel and binary formats (HDF5, .npy). Export them
to CSV first.
- Model: --model, from LABNET_AGENT_MODELS. The menu shows the
version (for example "Sonnet 5.5") once Claude Code has used that model.
- Effort: --effort (Low, Medium, High, Extra high, Max), meaning how
much the model thinks before answering. Haiku has no effort levels, so the
menu is greyed out for it.
Each viewer's model and effort are remembered in their browser.
Attached files are saved in uploads/<conversation id>/ in the
workspace.
- Only that conversation's answers can read them: each run gets --add-dir
for its own folder only.
- They are deleted 14 days after the conversation was last used.
Account. It uses the Claude Code login of the Windows account that runs
the Control Center. No API key is used. Any ANTHROPIC_API_KEY,
ANTHROPIC_AUTH_TOKEN or CLAUDE_CODE_OAUTH_TOKEN is removed from its
environment.
- Set up once, on that computer and as that account: install Claude Code,
run claude, and use /login.
- The sidebar badge shows Ready, Sign-in needed, or Not
connected (Claude Code not found). Hover over it for the reason.
- Anthropic's terms limit sharing one person's subscription with other
people through an app. Before the whole lab uses it, confirm that your
plan covers this (for example a Team plan).
What it can do. It can read, but not act:
- Tools: only Read, Glob and Grep. --permission-mode dontAsk denies
everything else instead of asking. It has no shell, no file edits, no
device control, none of the account's own MCP connectors, and no memory
between chats (CLAUDE_CODE_DISABLE_AUTO_MEMORY).
- Folders:
- its working folder home/ in the workspace (LABNET_AGENT_WORKSPACE);
- LABNET_AGENT_SOURCES, the LabNet code, docs and examples;
- LABNET_AGENT_LITERATURE_DIR, papers, which it can read as PDFs;
- the conversation's own uploads.
- Blocked even inside those folders: *.env files, *.key files,
databases and logs/. The Control Center's own Central key is never in its
environment.
- Live network: before every message the Control Center writes
home/live/network.md in the workspace: Lab PCs, devices, states, who
holds them, and methods. Questions about what is online are answered from
it.
- An open Automation project: the chat shows "Photo-Agent sees:
It gets copies, not the folder itself, because an added folder would also
load any Claude Code settings someone put in it.
- Briefing: its role and the lab's script style (from_env(),
try/except/finally, release in reverse order) are in
agent_briefing.md.
It writes scripts for people to run. It never runs them.
In development, labnet dev seed points it at this checkout's packages/
and examples/, and at literature/ at the repository root (git-ignored;
drop papers there). Its workspace is .devnet/agent.