Bots

A bot is an agent session that lasso launches and keeps running. Each bot has its own folder with its instructions, skills and environment, plus its own model and MCP servers.

A bot's runtime is one of two:

• omp (the default for a new bot): omp itself in a herdr pane, on any host lasso drives, kept running by lasso. It survives lasso restarts and updates mid-turn. It also gets a herdr pane, where you can watch it, type to it (Esc stops a turn, a number answers a question) and reach it with herdr's own agent commands. Its instructions are AGENTS.md, and the questions and approvals it asks wait in its chat for you to answer.
• Claude Code: Claude Code in a herdr pane, on any host lasso drives, with channels. Its instructions are CLAUDE.md.

The rest of this page describes both, and says where they differ. An assistant that answers your mail is a bot, and so is a watcher that runs on a schedule. The Bots view lists them like a messaging app: the bots down the left, and the selected one's conversation on the right.

Open it with ⌘. (press it again to go back to the terminal), from the footer's view menu (it names the view you are in) on a desktop, or from the view picker on a phone (see On a phone). Like Chat and the Grid, Bots covers the terminal without unmounting it. Leaving the view and coming back finds it on the same page, with the conversation's scroll and draft where you left them.

The address bar follows the view, so these links work and Back moves between them:

The list#

Each row shows the bot's avatar, its name, its newest message (prefixed You: when it was yours), how long ago that was, and its state:

A filled dot at the end of the row means the bot has said something since you last had its conversation open in this browser. Unread is kept per browser, so your phone and your desktop each keep their own.

Drag a bot by the grip at its left edge to reorder the list; the order is saved for everyone. New bots go to the end. The + in the header creates a bot, the bell beside it turns notifications on or off for this device, and Manage bots at the bottom opens the table.

When the view is narrow#

The list folds away when the Bots view itself is narrower than 760 px. That happens on a phone, in a narrow window, or beside a wide right sidebar. The list then becomes its own screen, and every other page gets a ‹ Bots button that goes back to it. In a conversation, the bot's name in the header becomes a switcher: it opens a menu of every bot with its state, plus All bots, Manage bots and New bot.

Talking to a bot#

The conversation is the chat view pointed at the bot (a Claude Code bot's pane, or an omp bot's session), with these differences:

• Messages from its channels appear as cards. A bot with channels (mail, Mattermost, iMessage and texts, its own scheduled loops) is mostly answering messages that arrive by themselves. Each one shows where it came from (Email, Mattermost, Scheduled and so on), who sent it, and its subject or thread. A long one is clamped, with Show more.
• The header is the bot's. It shows the bot's name and state and a gear that opens its settings. Pinning and ending the pane are not offered here: a bot is started and stopped from its settings, and lasso relaunches it if its pane is closed under it.

Everything else works as in Chat, including two composer features that apply to every conversation:

• Model, thinking and context in the header. A chip shows the model (with (default) when the bot names none) and opens a picker of your omp models plus free text; another shows the thinking level and lists the levels that model takes; a meter shows how full the context is (used / window and a share, amber from 70%, red from 90%; a Claude Code bot shows tokens only). An omp bot switches at once and keeps the change; a Claude Code bot saves it and offers Restart now. In a narrow view the three fold into one chip. In an omp bot's herdr pane, a status line shows the same, and /model and /thinking change them. Closed the pane's view by accident? mise run attach in the bot's folder brings it back.
• New conversation, or resume one. The New icon in the header starts an empty conversation (Undo in the toast goes back) or resumes any earlier one, newest first. An omp bot switches at once; a Claude Code bot restarts into it. Not while it is mid-turn. In the pane: /new (or /clear), /sessions, /resume <n>.
• An omp bot takes a message mid-turn. It goes in at once as a steer rather than waiting for the turn to end, and a question or approval it asks shows above the composer as buttons to answer (an approval is Approve or Deny).
• Stop. While the agent works, an empty composer's send button becomes Stop, which interrupts the turn (one Esc in a Claude Code bot's pane, an abort for an omp bot). If the agent was not working, or lasso could not tell whether the interrupt landed, it says so above the composer.
• Up next. A message sent while a Claude Code bot works is held, shown as Up next above the composer, and sent once the bot is idle, one at a time and in order. lasso holds the queue, so it goes out with the tab closed or the phone locked, and every device shows the same one. Edit, move up or remove a held message from its row. Pressing Stop puts the held messages on hold: nothing goes out until you press Send queued messages. When lasso cannot confirm that a held message was delivered, the queue pauses and says so, and nothing is resent automatically.

A stopped bot has no conversation to show. Its page offers Start, which carries on its last conversation, and Start a new conversation.

Settings#

The gear in the conversation's header opens /bots/<name>/settings. General and Connections are the bot's own settings and are saved together by the footer's Save. When the bot is running, Save & restart saves them and relaunches it; if it is in the middle of a turn, lasso asks first. The other tabs edit files in the bot's folder and save as you change them.

General#

Pictures#

Click the avatar under Avatar to upload a PNG, JPEG, WebP or GIF of up to 2 MB (a square one looks best). It shows in the list, the conversation and the bot's notifications. Click it again to replace the picture, or the small × on its corner to remove it. The picture is saved as soon as you pick it, not with the form. lasso keeps it in the bot's folder as .lasso/avatar.<ext>. Without a picture the name's first letter shows, and SVG images are not accepted.

Connections#

The MCP servers the bot gets, as a grid of cards. Each card shows the server's name, a CHANNEL badge or its type, its command or URL, and its environment-variable count or sign-in state. A dot marks a card with unsaved changes. The chips above filter the grid: All, Channels, or Tools (everything that is not a channel), each with its count, and the choice is remembered in this browser. Add connection under Channels starts the new server as a channel.

Click a card, or Add connection, to edit it in a dialog. A server is stdio (a command, its arguments one per line, and environment variables) or http/sse (a URL and headers). Tick Channel on a server that also delivers messages for the bot to answer. Done keeps the change in the page and the footer's Save writes it; Remove connection takes the server out. Closing the dialog on a new server that was never filled in drops it.

A dashed lasso card, added by lasso, is the server lasso gives every bot on its own machine for its settings tools. A dashed lasso-channel card is the channel that delivers its jobs. Both are shown so the grid is complete, and cannot be edited.

Only these servers starts the bot with these servers and nothing else, without your claude.ai connectors or user-level MCP servers.

Keep secrets out of this page. Write ${VAR} in a server's environment or headers, and set VAR under Environment, where it can be stored encrypted.

Signing in with OAuth. For an http or sse server that asks you to log in, tick Sign in with OAuth, save, then press Sign in.

• What happens: the server's login page opens in a new tab. Once you approve, it sends you back to lasso, which finishes the sign-in by itself. lasso discovers the server's sign-in settings and registers itself as a client automatically.
• When the redirect won't load: some servers will only send you back to localhost. When lasso runs on a VPS, that page fails to load. Copy the address from that tab's address bar (it holds ?code=…&state=…), paste it into the box under the server, and press Finish.
• Where the tokens go: lasso keeps the tokens in the bot's fnox.toml, encrypted like any other secret, and refreshes them before they expire. Claude reads the current one each time it connects. Each bot signs in separately, so two bots can use different accounts on the same service.
• A running bot needs a restart: the first sign-in, and a sign-out, change how the bot connects, and a running bot only reads that when it starts. The panel says so; restart it from Launch. Later token refreshes need no restart.
• Servers that don't register clients automatically: open Server without automatic client registration and enter the client ID, the redirect URI it was registered with, and the scope.

Jobs#

Scheduled prompts, watches and webhooks that lasso delivers into the bot's session, shown as a grid of square cards (as many to a row as fit). Each card shows the job's name, its schedule in plain words with the time zone ("Every day at 6 AM, noon and 6 PM · PT"), when it runs next and when it last ran (✓ delivered, ✗ dropped: hover for why), and the start of its message. A watch's card adds Checked: when its command last ran and what came of it ("3m ago, quiet", "1h ago, failed (exit 1) ×5"; hover a failure for its last stderr line). Badges mark a watch, a webhook, a paused job, and queued ×N when it fired while the bot was busy or not listening. Enabled jobs come first, by next run, then jobs with no schedule, then paused ones. Run now fires a job at once (a paused one too); for a watch it runs the command and says what came of it: delivered, no output (nothing delivered), or failed. Copy URL copies a webhook's address. The ⋯ menu pauses, resumes, duplicates or deletes. Changes take effect immediately, with no save of the bot and no restart.

Click a card, or New job, to edit it:

• Message: the instruction delivered each time the job fires. For a watch it is optional and is shown above the command's output.
• Command: optional. A job with a command is a watch: each time it fires, lasso runs the command (with sh -c, in the bot's folder) and delivers only what it prints. Exit 0 with no output delivers nothing, which is the point: the bot is woken only when something changed. Exit 0 with output delivers the message, then the output. A non-zero exit or a timeout (Timeout, 60 seconds unless set, up to 3600) is reported to the bot, damped to the 1st, 2nd, 4th, 8th… failure in a row, with one notice when it works again. The command runs on the bot's host: as lasso's user on lasso's own machine, or as the user lasso connects as on another host. Either way it gets a minimal environment, not a login shell: use absolute paths, and have the script fetch its own secrets (secret KEY). Last check shows the newest run. A watch for a stopped bot does not run at all.
• Schedule: Once (a date and time), Every… (minutes or hours), Daily, Weekdays, Weekly (pick days), Monthly (pick a day), each with one or more times, or Custom for a cron expression. The time zone defaults to your browser's. A box under it reads the schedule back as a sentence with its next three run times, or says why it is not valid. A one-time run fires once (late, if lasso was down at the time) and then has no next run.
• Webhook: turn it on and save to get a URL. Anything that POSTs to it fires the job, and the request body is delivered after the message, marked as coming from the caller. The URL's key is its only credential, so it is masked until Show, and Rotate key replaces it (the old URL stops working at once). A caller that sends headers can use Authorization: Bearer <key> instead. From outside the tailnet, the hostname's edge (Cloudflare Access) must let /hooks/ through.
• Recent: the job's last deliveries: when, what fired it, and whether it was delivered or dropped. A watch's quiet runs add nothing here; a reported failure is marked failed or timed out.

The line under the description says whether the bot is listening. Jobs reach a bot through lasso's own channel, which it gets from its next start: a bot that was running before lasso had jobs offers Restart. A job that fires while the bot is stopped is dropped, not saved for later, and one nobody picks up within a day is dropped too. Firings that pile up while the bot is busy arrive as one delivery with a count; for a watch, each run that printed stays in that delivery under its own header. A Claude Code bot on another host gets its jobs too, when that host has a lasso new enough to deliver them (the line says to update it when it is not): lasso writes each delivery into the bot's folder over SSH, and the host's own lasso hands it to the bot, so nothing on that host connects back to lasso. An omp bot gets its jobs on any host, as prompts lasso sends over the bot's socket, which it reaches over SSH.

Talk#

Talk gives someone a chat with the bot in an app on their phone that speaks the OpenAI API, such as Conduit. Their messages reach the bot's session and its replies go back to their app. Claude bots only.

Turn on Talk makes a key. The tab then shows:

• Base URL, API key (masked until Show) and Model (the bot's name), each with Copy. In Conduit: Add connection, provider type OpenAI-compatible, the Base URL, authentication Bearer token with the key, then pick the model.
• While it works:
  • Plain (the default): the app's thinking line shows only the kind of each step ("Reading a file", "Running a command").
  • Detailed: adds each step's subject, which shows file names and commands from the bot's machine.
• Whether the bot is taking messages. A bot started before Talk existed takes them after a restart, and the tab offers the button.
• Rotate replaces the key; the old one stops working at once. Turn off Talk retires the key.

Whoever holds the key can ask the bot for anything its instructions allow, so say in its CLAUDE.md what it may do for them. The bot answers through its channel's reply tool. A long job gets an "On it" first, and an answer that misses its request (the app was closed) arrives at the start of the next one. From outside the tailnet, the hostname's edge (Cloudflare Access) must let /talk/ through.

Skills#

Always available lists your user-level skills, from ~/.claude/skills on the bot's host, in a scrolling list. Claude Code loads them in every session, so every bot has them without adding anything.

This bot only is the bot's project skills, in .claude/skills in its folder, which no other session sees. To add one, paste anything into Add a skill: a URL, a GitHub repo, a path, a SKILL.md, or a description of what the skill should do. Then press Ask to install it. The request goes to the running bot as a chat message, and the bot fetches or writes the skill into its own .claude/skills; you can follow along in its chat. The list updates once the skill is there. Removing a skill asks for a second click.

Instructions#

The bot's CLAUDE.md (an omp bot's AGENTS.md): who it is and how it works. Edit it here and save it (or ⌘S). A running bot reads it again on its next restart.

Environment#

Variables in the bot's fnox.toml, which mise hands to the bot, and nothing else, when it launches. Tick Secret to keep a value in the file's default provider. Unless you changed it, that is lasso's own age key (~/.lasso/age.txt on the bot's host), and the value is decrypted only at launch. Afterwards it shows as ●●●●●●●● and can only be replaced, never read back. To keep secrets in 1Password, a vault or elsewhere, add a provider to the bot's fnox.toml with fnox provider add and make it default_provider. Changes reach a running bot on its next restart. Bots need mise 2026.10.4 and fnox 1.39 or newer on their host.

A host with no age key yet cannot store secrets. The page says so and offers Create encryption key, which asks before writing ~/.config/mise/age.txt on that host. Every bot on the host uses that one key, so back it up.

Launch#

Start, stop and restart the bot: Restart, same conversation carries on where it was, Restart, new conversation begins an empty one (the old one stays on disk); and Open terminal to switch to its pane in the terminal. The bot can also restart itself: ask it to, and it runs mise run restart, which relaunches it in the same pane on the same conversation once its turn ends. Below them is the launch script lasso generates from the settings. The bot runs as mise run bot in its folder, and lasso rewrites the script on every save, so edit the settings rather than the file. Run by hand in the folder (after updating Claude Code, say), a bare mise run bot resumes the bot's conversation; mise run bot -- --fresh starts a new one.

Launch task picks the mise task that starts the bot: bot, or a mode of your own in its folder. Start, keep-running, herdr's restore and the bot's own restart all go through it. A mode is a task that sets an environment and hands off to bot, so the channel grants and launch dialogs stay lasso's. For example, .mise/tasks/deepseek runs the bot on DeepSeek's Anthropic-compatible API:

#!/bin/sh
#MISE description="Run on DeepSeek"
#MISE secrets=["DEEPSEEK_API_KEY"]
#MISE interactive=true
export ANTHROPIC_BASE_URL="https://api.deepseek.com/anthropic"
export ANTHROPIC_AUTH_TOKEN="$DEEPSEEK_API_KEY"
unset ANTHROPIC_API_KEY DEEPSEEK_API_KEY
exec mise run bot -- --model sonnet "$@"

Add DEEPSEEK_API_KEY as a secret under Environment, then pick deepseek here. #MISE interactive=true is required: without it a task that receives secrets loses the terminal. Keep "$@" last, since a restore appends --resume <id>. Two limits to know:

• A conversation started on Anthropic may not resume on another provider. DeepSeek refuses some of the content Claude Code stores, so after switching, use Restart, new conversation (by hand, mise run deepseek -- --fresh).
• claude.ai connectors are off under another provider's token. Claude Code turns them off whenever ANTHROPIC_AUTH_TOKEN or ANTHROPIC_API_KEY is set, so put any connector the bot needs among its own MCP servers.

A launch task that the folder doesn't define is refused when you save.

Deleting a bot#

Delete bot… in the footer works in any state: a running bot is stopped first, which closes its pane. lasso forgets the bot, and its folder stays on the host with its instructions, skills and environment unless you tick Also delete its folder, which removes the folder and everything in it. lasso refuses that for a folder that holds your home directory or lasso's own.

Right-click a bot in the list (long-press on a touch screen) for the same actions without opening it: Open chat, Settings…, Start, or Restart and Stop while it runs, and Delete….

Asking a bot to change itself#

A bot can change its own settings when you ask it to in its conversation: "switch to Opus", "add the GitHub MCP server", "set API_URL to …", "use this picture". Every bot on lasso's own machine gets lasso's MCP server, named lasso, and its tools read and change that bot's settings:

• the General settings: model, effort, permission mode, extra args, keep running, avatar and workspace
• its MCP servers and channels
• its environment variables, plain or secret
• its picture, from an image file it downloaded or generated

It edits its CLAUDE.md and .claude/skills directly, as files in its folder. Most changes take effect after a restart, and the bot restarts itself to apply them. Some things stay with you:

• Approving a sign-in to an OAuth server. The bot can start it (begin_bot_oauth), but you open the address it gives you and approve it, or sign in from Connections.
• Bots on other hosts and gated servers. A bot on another host doesn't get the lasso server, and neither does any bot while lasso's /mcp requires OAuth (MCP_OAUTH). Add the server by hand under Connections in that case. A server you named lasso yourself is kept as is.

It is safer to type a secret into Environment yourself than to paste it into a conversation.

Notifications#

lasso can notify you each time a bot finishes a reply. The notification is titled with the bot's name, shows the start of its message and its picture, and opens that bot's conversation. A newer answer from the same bot replaces the older notification rather than stacking.

• Turn it on for each device. The bell in the list's header turns notifications on or off for the device you are on. It does the same as Settings → General → Notifications,.
• Turn it off for a bot. Untick Notify me when it answers in that bot's General settings.
• No duplicates for what you're reading. No notification appears while that bot's conversation is focused and on screen, except on iPhone and iPad, which always show one.

Every device with notifications on gets every notification, including blocked agents from the rest of lasso.

Creating a bot#

The + in the list's header, or New bot on the management page, opens the creator. It has the General, Instructions and Connections tabs. The bot starts as soon as it is created and lands on its chat. Skills and environment are added from its settings afterwards and take effect when it restarts.

Manage bots#

/bots/manage lists every bot with its host, workspace, model, number of channels, state and when it was last active, with Start or Stop and a settings button on each. At desktop widths it is a table; on a phone it is a card per bot.

On a phone#

On a phone the view is one column at a time, as in When the view is narrow: the list, then the bot you pick, its settings or the management page, each with ‹ Bots to go back.

Bots covers the input dial, so the list's header and each bot's own header have a Switch view button, which opens the same view picker as the dial's big button.