Glossary
The words we use for how a Koshra team works, each in a sentence or two.
- AG-UI
- An open protocol for streaming agent runs to a user interface. Koshra's client console is built on it and is in testing.
- AI employee
- A bot in Koshra with a role name, a written job, a limit, its own model and its own memory.
- Allow, Ask, Block
- The three rule settings for a tool. Allow runs it, Ask stops for your approval and Block refuses it.
- App connection
- Access to an app such as Gmail or HubSpot through a connection layer, limited per bot to what its role needs.
- Approval card
- The card a bot raises when a rule says Ask. It shows what the bot wants to do, and you answer Allow or Deny.
- Approvals inbox
- The one place every waiting approval card is collected. Only the space owner and admins see it.
- Backup chain
- The models behind a bot's main model. If a provider is down or out of credit, the next model in the chain answers.
- Brain
- Your team's shared store of facts, edited as plain text in the app, with every edit saved so you can undo it.
- Brain sync
- A routine that reads recent work and proposes Brain updates. Proposals wait for your review.
- Branch
- A new thread started from any message in a chat, leaving the original thread untouched.
- chief_of_staff
- The bot at the top of a Koshra team. It prepares the day's list and names the officer, desk and next step for each item.
- Cost per reply
- The model, cost in USD and time shown on every reply.
- CRM guard rails
- Rules that block deleting or merging CRM records and hold bulk changes and large writes for approval.
- Desk
- An AI employee that does one focused job inside a lead's area, such as outbound_desk or pipeline_desk.
- Dispatcher
- The part of Koshra that turns a request into tasks, picks the bot that owns each one and shows a plan card before work starts.
- Handoff
- Moving a task to the bot that owns that kind of work. Handoffs are capped per task.
- Lead
- An AI employee that owns an area of work, such as research_lead or support_lead, and has a team chat with its desks.
- llms.txt
- A plain-text index of a website for AI assistants. Koshra's lists every page, and llms-full.txt holds the whole site as Markdown.
- Loop guard
- A check that stops two bots passing work back and forth, and stops repeat visits to the same bot.
- MCP
- The Model Context Protocol, an open standard for giving AI models tools. Koshra adds MCP servers per bot, under the same rules as everything else.
- Memory bank
- A bot's private memory. It keeps what the bot learns and recalls it in later chats.
- Model picker
- The control in the chat composer that switches a bot's model. The next reply uses the new model.
- Owner bot
- The one bot that owns a task. You can tell it to continue, retry the task or reassign it.
- Permission mode
- A per-bot setting on top of the rules. It can make a bot stricter, never looser.
- Plan card
- The card the dispatcher shows before work starts, with its reasons, the bots it picked and a countdown you can stop.
- Role
- The job an AI employee is named for, in lowercase snake_case, such as research_desk or support_triage. The role name is what you see on chats, tasks and approval cards.
- Routine
- Work a bot repeats on a schedule, run on the server. Creating or changing a schedule needs your approval.
- Scoped API key
- A key a Koshra space admin makes so another agent can use the REST API or MCP server. It acts as that admin in one space, only within the scopes picked, and the secret is shown once.
- Self-hosted
- Running Koshra on a server you control, as a small set of Docker containers, next to your own tools and data.
- Send limit
- How many messages a bot may send on its own. For drafting roles it is zero: they draft and you send.
- Space
- A workspace in Koshra with its own bots, rules, tasks and Brain. The dispatcher and tasks board are switched on per space.
- Spend cap
- The most a single dispatch may spend on model calls before it stops. Dispatches also have token, handoff and run caps.
- Stuck run
- A run that has stopped moving for longer than the threshold you set. It shows in the stuck-run bar with a cancel button.
- Subagent
- A short-lived helper a bot runs for a side job, which reports back to the bot that started it.
- Tasks board
- The board where tasks move through their stages, each showing its owner bot and status.
- Team chat
- A chat shared by a lead and its desks, with the lead listed first. You can read along and step in.
- Team computer
- A shared container on the server where an allowed bot can run tools and work with files.
- Web Push
- Browser alerts on your phone or laptop when an approval card is waiting.
Also as Markdown for AI agents.
Last updated