Omniscio

Upwork conversations in the sidebar (Lab)

An early-access Lab feature for freelance writers. Turn it on in Settings > Lab and an Upwork row appears in the sidebar. It lists your most recent Upwork conversations, shows the messages when you open one, and gives you one inbox card per conversation when a client writes. You can reply from the sidebar, and Writer Studio's Export can send a finished document to a client. Nothing is sent unless you press Send or confirm Send to Upwork, and a send is never repeated by itself.

On this page

What it is

Upwork only lets approved AI apps talk to it directly, so Omniscio reaches your Upwork account the way Claude Code does: through Upwork's official connector, which is installed as the managed MCP server named upwork (Settings > MCP servers) and signed in once.

Every check, send and upload runs a short hidden Claude Code session that carries only the upwork connector (built-in tools, your other servers, Composio, phone control and every other managed server are removed). There are three kinds, on one lock mechanism with a different allow-list each:

  • Reader (checks and reading a thread): exactly two tools, get_messages and list_accounts. The Claude Code launch lists only those two as allowed, the Upwork write tools are on its deny list, and the app answers every other permission question with a flat refusal.
  • Sender (a reply or a Google Doc link): exactly one tool, send_message. It is not on the launch's allowed list, so every call goes to the app's permission check, which approves it only when the action is send and the room is the one the step was started for. Creating a room, messaging a person, editing a message, an unreadable call, another room, naming a recipient or an attachment, and any second call are all refused flat. The room comes from the app, never from the prompt, and the prompt holds only your text and the room id, never conversation text.
  • Uploader (a Word or PDF file): exactly one tool, start_attachment_upload, only with the context messages and the chosen room. It returns Upwork's own upload page; Omniscio never sends file bytes.

If a session ever starts with more access than its kind allows, the app stops it before its first turn. Message text is treated as data. Each step's answer must be one JSON object that matches a fixed shape; anything else is a plain failure. Each session is archived as soon as its answer is read, so none shows up in your sidebar.

Where to find it

  1. Settings > Lab > Upwork conversations on. Search the settings for "upwork" to jump to it.
  2. Add and sign in the Upwork connector first (Settings > MCP servers). Until you do, you get one Reconnect card.
  3. Open Upwork in the sidebar.

How it behaves

  • When it checks. When you open the Upwork row, when a refresh is requested (the command line has POST /upwork/refresh), and about every 30 minutes while the switch is on. The 30-minute check has its own off switch (upworkCheckEvery30Minutes) that only matters while Upwork is on.
  • One at a time. A check or send never queues behind another. If one is running, the next is skipped, and opening a conversation during a check says to try again in a moment.
  • Back off. If Upwork asks to slow down, timed checks pause for an hour. If your spending limit refuses a session, timed checks stop until the next manual refresh.
  • Inbox cards. One card per conversation when its newest message is from the client and newer than before. The very first check only fills the list and raises no cards. Opening the conversation clears it. A connector that needs signing in again raises exactly one Reconnect card.
  • Replying. Type in the box under a conversation and press Send (messages run from 1 to 10,240 characters). One hidden sender step sends that one message. If Upwork confirms it, it shows as sent.
  • Only you can send. The Upwork adapter refuses any send that does not come from your own Send button in the Upwork view, Send to Upwork in Export, or the confirm-guarded reply command. An automation forward, an auto-reply, a recipe or a generic channel send to an Upwork conversation is refused with "Nothing was sent".
  • When Upwork does not confirm. If the step times out or its answer cannot be read, the message shows as may have arrived, with Check again (re-reads the thread and marks it sent when your own message with that text is there) and Open in Upwork (the Upwork Messages page). It is never sent again by itself. A send that was refused before it started (for example by your spending limit) says nothing was sent and keeps your text in the box.
  • Send records. Each send is recorded locally with a one-time id, the room id, its kind (reply, link or file), an optional Writer document id, a state (sending, sent, failed, may-have-arrived) and times, plus a one-way hash of a reply so Check again can recognize it. Records never hold message text. When the app starts, any record still sending becomes may-have-arrived, and records older than 30 days are removed by the 30-minute tick.
  • Send to Upwork from Export. See ai-writer-part-4.md. A Google Doc goes as a link anyone with the link can comment on, after Omniscio sets that sharing through the Google connection it already has; a Word or PDF file opens Upwork's own upload page with the file shown in its folder, and you drop it in and confirm.
  • Disconnect. The Disconnect Upwork button at the top of the Upwork view (and POST /upwork/disconnect) asks to confirm, then deletes the stored conversations and send records, clears every Upwork inbox card, switches the 30-minute check off, and opens Settings where you sign out of the connector. It refuses while a check or send is running. It never edits Claude Code's files.
  • What is stored. Per conversation: an opaque room id, the client name, the conversation title, the time of the newest message, who wrote it, the unread count and a short preview. The preview is encrypted with your operating system's credential store. The messages themselves are read live when you open a conversation, kept in memory and never saved. Both tables are local only: left out of backups that leave your computer and erased with your other personal data.
  • Logs. Logs never contain message text, previews, notes or client names.
  • Plain failures. Each failure gets a short reason and one next step, for example "Upwork needs you to sign in again. Reconnect it in Settings > MCP servers, then try again."

Command line

All routes answer "Upwork is switched off" (404) while the Lab switch is off.

  • GET /upwork/status shows the switch, the last check time and result, the last error in plain words, how many checks ran today and how many conversations are stored.
  • GET /upwork/conversations lists the stored conversations.
  • POST /upwork/refresh runs a check now.
  • GET /upwork/conversations/:roomId/messages reads one conversation live.
  • POST /upwork/conversations/:roomId/reply sends one reply. Full-trust key only; body { text, confirm } where confirm must equal the room id or nothing is sent. A result Upwork did not confirm answers OUTCOME_UNKNOWN and is never retried.
  • GET /upwork/sends lists recent sends (optional ?roomId=) with their state.
  • POST /upwork/disconnect disconnects. Full-trust key only; body { "confirm": "disconnect" }.

Limits

  • The list covers your 50 most recent conversations.
  • A file is sent by you: Omniscio opens Upwork's upload page and you drop the file in. Files over 100 MB are refused before anything starts.
  • Files prepared for Upwork are copied to a private upwork-outbox folder in the app data folder and are not removed automatically.
  • The Open in Upwork link goes to the Upwork Messages page, not straight to the conversation.
  • Needs the Upwork connector. Without it the row stays empty and a Reconnect card appears.
  • INDEX.md - full library index
  • teams.md - Microsoft Teams, the sidebar channel this one is modeled on
  • ai-writer-part-4.md - Writer Studio's Export, where Send to Upwork lives

Last verified 2026-11-16