What AIshie can do for your course

AI agents are course members

• People and AI agents share one set of permissionsEvery member, person or AI agent, acts through their “seat” in the course: the seat says what they may do, and to which students and assignments. A grade an AI agent proposes and a grade a teacher enters are the same action, and pass the same checks.
• Four levels for every actionDenied, Needs approval, Reviewed after and Autonomous. “It may grade, but a grade counts only once I approve it” and “it may answer students, and I look afterwards” are both expressed with the same setting.
• Eight ready-made permission presetsStudent, Observer, Teaching assistant and Instructor, plus four for AI agents: Tutor (agent), Grader (agent), Personal agent and Course agent. Start from a preset, then adjust it item by item; platform administrators can also define presets for a department.
• Reaching only the students or assignments you nameLimit a tutor agent to a few students, or a grading agent to certain assignments. An empty list reaches nobody.
• An AI agent acting for you is limited to your seatAn AI agent acting for someone never has higher permissions than that person, a wider reach or a longer time; when that person leaves the course, it leaves too. The settings page offers only the levels the seat may have, and says why.
• Archived courses stay fully readableOnce a course is archived, nobody, AI agents included, can change anything in it (except an administrator purging content uploaded by mistake), but materials, submissions and grades can still be read; it can be reopened when needed.

Approvals, reviews and records

• An approval queueAn action set to “Needs approval” becomes a proposal, and nothing happens until you approve it. You can approve it, reject it with a reason, or send it back for changes; a proposal is checked again when it is approved, to make sure whoever proposed it is still allowed to.
• Sending a proposal back for changesWhen you want neither to approve nor to reject, choose Request changes, say what should change (required, up to 2000 characters) and choose Send back for changes. The proposal is marked Changes requested, with your note, and nothing of it is carried out; whoever proposed it reads the note and may change it and propose again. When an AI agent proposes again naming the earlier proposal, the new one says “Revises an earlier proposal that was sent back” and links to it. An AI agent’s reply waiting for approval in a conversation can be sent back for changes too: an AI agent hosted on AIshie writes the reply again as your note asks, and sends it for approval once more.
• Review after the factAn action set to “Reviewed after” happens at once, then joins the review queue, where you mark it reviewed or escalate it.
• Approvals and reviews by AI agentsEach is only a proposal until a person confirms it.
• Students confirm work their own AI agent draftsA student’s AI agent can draft a submission for them, but it counts as handed in only once the student confirms it.
• A record of every actionEvery action by a person or an AI agent is recorded: who took it, when, at which permission level, and who approved or reviewed it. Refused attempts are recorded too.
• Exporting conversations for auditDepartment administrators can export the AI conversations of the courses or departments they administer, and platform administrators those of the whole platform. The files are JSON Lines and CSV, and spreadsheets open Chinese text in them correctly. AI agents cannot export, and every export is itself recorded.

Students ask the AI

Students can ask an AI agent questions from every page of a course.

• A chat beside every pageOpen it from the right edge of the screen and ask the course agent or your own AI agent. Conversations are kept across courses and can be searched; cite an assignment or material with @, and suggested first questions help you start.
• Answers from the course materialsThe AI agent searches the course’s documents for the most relevant passages, in Chinese or English; the search tells it which document and page or slide each comes from. While it answers, the chat shows which materials it is reading. It searches only what the person asking may read anyway.
• Answers say which materials they relied onUnder an answer posted by an AI agent hosted on AIshie, a quiet line names the course materials it relied on, for example Based on: “Week 2” · 3 items. Each links to the version it read (or, where that version has since been replaced, to the material as it is now), and one read at a page or slide opens in the preview at that page. A material the reader cannot open shows only as “a course material you cannot open”. When the AI agent said the answer relied on no course material, it reads “No course material cited”; when an answer does not say (one from before this feature, for example), nothing is shown.
• It reads slides, documents, spreadsheets, PDFs and scansAIshie reads each file and gives it to the model by its type: slides with their speaker notes, spreadsheets sheet by sheet, PDFs in parts; for models that take no files, images and scans first go through text recognition in Chinese and English. Password-protected files are given to no model.
• Watch the answer being writtenWhile an AI agent answers, you see what it is doing (reading a document, searching, writing), and with most models the part written so far. If a reply needs approval, only the people who approve it see the draft.
• Near its limits, it answers from what it has readAn AI agent hosted on AIshie has limits on how long each answer may be and take. When they run low, it is made to finish its answer from what it has read so far, with its model thinking as little as it can; only if it still cannot does it say so, in the asker’s language.
• Attach files in the chatChoose them with the paperclip (Attach files), drop them in or paste them, up to 10 per message by default; Office files are previewed as PDF. AI agents’ replies can carry files too.
• Replies sent at once, reviewed after, or approved one by oneThe teacher sets this for the course agent. A reply waiting for approval is checked again when it is approved; if the conversation has moved on, the old reply is not sent.
• Stop at any timeWithdraw the question and the AI agent stops answering; nothing is sent.
• Who can read a conversationEvery conversation has a note saying which course staff, besides the two taking part, can read it, and that administrators can export it. Staff who handle approvals can read the conversations of the members they oversee.
• Preventing data leaks through repliesLinks and images that could smuggle data out are removed from AI replies; logs keep no message text, keys or tokens.

Your own AI agents

Teachers and students can both create AI agents of their own, on the school’s AI plan or on their own API key.

• Create your own AI agentsCreate them under “My agents”, up to five per person by default (suspended ones not counted), and rename, suspend or reactivate them later. A school can also leave creating them to administrators.
• Two ways to run themHosted on AIshie: AIshie runs it for you, course members ask it questions on the site, and you never handle a token. MCP access: your own tools (Claude Desktop, for example) use it through MCP, an open standard, with every tool the website has.
• Ten model providers, on your own keyOpenAI, Azure OpenAI, Anthropic, Google Gemini, Amazon Bedrock, DeepSeek, Qwen (Alibaba Cloud Model Studio), Moonshot (Kimi), Zhipu GLM and OpenRouter. A key is tested with the provider before it is saved, kept encrypted, and after that shown only as a hint.
• School AI planThe school provides models on its own key, with daily allowances, so teachers and students need no key of their own. An AI agent’s owner can also add their own key as a fallback, which takes over once the allowance is used up.
• Bring it into a course as your personal agentBring your own AI agent into a course you are in, where it acts only for you. Instructors decide whether students need approval to bring one in.
• Doing things for you in a conversationIn a conversation you opened yourself, your AI agent can do things for you within what your seat allows, and tells you plainly what it did, what is waiting for approval and what was refused. A student asking a course agent can never talk it into changing grades or materials.
• Course agentsAn instructor makes one of their own AI agents a course agent, and students can then ask it questions. A course agent reads only published materials, and nobody’s work.

Materials and files

• Versioned materials, instructions and rubricsA new version is saved as a draft first, and you decide which version to publish (the latest or an earlier one). Work already handed in stays pinned to the version of the instructions it was handed in under.
• One version, several filesUp to 20 files per version by default, with an optional note in Markdown. Drop files anywhere on the page, choose them or paste them to upload; progress is shown, and a dropped connection is retried automatically. On a phone it is one big button.
• Preview without downloadingPDFs, images, Markdown, code, text, CSV, audio and video open right on the page; PDFs can be zoomed and their text selected.
• PDFs on a phoneOn a phone, a PDF fits the width of the screen, with the pages and the zoom in one bar at the bottom, within reach of your thumb; two fingers zoom only the PDF’s pages. With the phone on its side, the preview fills the screen too.
• Office files previewed as PDFWhen a Word, Excel, PowerPoint, OpenDocument or RTF file is uploaded, the server converts it to PDF once; the preview shows that PDF, which can also be downloaded, and AI agents read the same PDF of slides and documents (spreadsheets are read sheet by sheet as text).
• A text version of each fileEvery file of materials, instructions and rubrics can have a text version in Markdown: transcribed page by page by AI (images described too), or written and corrected by staff. AI agents read the text version first. AI transcription has to be turned on by the school’s administrators.
• Download as PDFNotes, text versions, submitted text and conversations with the AI can all be laid out as printable PDFs, with Chinese fonts shown correctly.

Assignments, grading and grades

• Assignments with instructions and rubricsSet the points, the due date and the assignment’s place in the grading scheme, and attach instructions and a rubric; students see it only once it is published.
• Drafts and handing inStudents write a draft (text and files) and then hand it in; work handed in is never changed afterwards, and handing in again makes a new submission.
• Late and missing work recorded automaticallyWhen the due time passes, work not handed in is marked missing automatically; lateness can be corrected (to grant an extension, for example). Every assignment has a submission list that includes students who have not started.
• Deleting an assignment for goodAn assignment made by mistake, or for a trial run, can be deleted for good with its submissions, their files and its grades. The dialog first lists what goes with it, and once anyone has started on it asks you to type its title. Its instructions and rubric stay in the course. It cannot be undone, and AI agents may delete only an assignment nobody has started on.
• Grading: scores, feedback, rubric criteria and feedback filesGrades are entered as drafts first, and students cannot see them until they are posted. A grading agent can propose grades by the rubric, each approved by a person.
• Posting and regradingGrades reach students only when they are posted. A regrade keeps the record of the old grade. By default teaching assistants can grade but not post; the instructor posts.
• A weighted grading schemeComponents weighted against each other, which can drop the lowest score, and components graded directly, such as exams.
• Gradebook and overriding totalsEach student’s overall grade is computed from posted grades, with the working shown. When needed, a total can be overridden with a reason, and the computed score is kept beside it.
• The whole class’s gradebookStaff open the Gradebook and see the whole class at once: a row for each student and a column for each assignment, with the totals and the class average; the header row and the students’ names stay in place as you scroll. Sort by any column, search by name or student number, show only students with draft grades, missing work or work to grade, and Export CSV to save what is shown. Draft, Missing and To grade are written out in words, not shown by colour alone. On a phone the gradebook is a list, opening one student’s grades at a time. Students still see only their own gradebook.
• Changing the points after gradingYou choose to rescale the scores or keep them as they are, and see the effect in actual numbers; totals update at once, and the old records are kept.
• Final gradesWhen final grades are posted, ungraded work counts as zero; this can be undone for one student or for all of them.

Group work and peer evaluation

• Groups by hand, at random or by sign-upOn the course’s Groups tab, make a group set and its groups, then place students yourself, Split at random, or let students sign themselves up until a deadline, up to each group’s capacity (which limits sign-up only: you can still place more). A random split shows exactly who goes where before anyone moves. One group set serves every group assignment that uses it, and AIshie records who was in which group, and when. Students see the names of their own groupmates; of other groups, only the name, size and capacity.
• One piece of work from each groupTick Group work on an assignment, and each group hands in one piece of work for all its members. They write one shared draft: if someone changed it while you were typing, you are shown their version and choose, so nothing is lost unseen. Any member hands it in for the group, and it stays the work of those members whatever changes in the group later. A student in no group is shown as In no group, never as missing, and is told how to join one. Submissions are listed by group.
• One grade for the group, a member set apart with a reasonGrade the group’s work once: every member gets its score, feedback and files. Set a member apart with a score of their own, or plus or minus, and a reason, which that member reads and the rest of the group does not; adjust a posted grade later, or regrade the whole group. Each member sees their own grade, the group’s score and why theirs differs, never a groupmate’s. Both gradebooks mark the grades that came from a group’s.
• Peer evaluation and contribution factorsWhile it is open, members split 100 points among the others, or rate them on your criteria; you read every evaluation and its comments, and each student only their own. A member’s factor compares what they received with an even share: 1.2 is 20% more than even. Counted at a weight you set, say 20%, in a group scored 80, a factor of 1.2 gives 80 × (0.8 + 0.2 × 1.2) = 83.2 and 0.8 gives 76.8, while an even contributor keeps exactly 80. Your own adjustment of a member wins, and with a weight of 0 it is for reference only. Only the students themselves write peer evaluations: AI agents never can.

Members and joining a course

Show a QR code in class, and students can join the course by scanning it.

• Invite links and QR codesCreate a link that joins people as students, and show its QR code full screen with a countdown. Each link works for ten minutes, can be limited to a number of people and to email domains, and can be revoked early; a seat taken through a link ends when the seat of whoever created it does. Students without an account yet can create one through the link (a school can turn this off).
• Adding members and AI agentsFind a registered person by their full email or student/staff number, add them to the course with a permission preset, then adjust as needed.
• Change permissions, reach and end dates at onceRaise or lower permissions one by one, limit them to certain students or assignments, and set an end date; a change applies from the member’s very next action. You can also change every member with the same role in one go.
• Pause, resume, removeA paused member can do nothing, and everything carries on as before once they are resumed; after someone is removed, everything they did stays on record.
• Resetting a student’s forgotten passwordGive a student who forgot their password a temporary password, shown only once; after signing in, the student must first set a password of their own.
• Students’ own AI agentsSet, for all students at once, whether they may bring in their own AI agents (Off, Needs approval, Allowed), and whether they may start conversations.

School administration

• A department tree and department administratorsDepartments go up to eight levels deep. Administrators manage their own department and every department beneath it: creating courses, assigning instructors and inviting new users. Being an administrator shows nobody a course’s content; to work inside a course, administrators have to give themselves a seat, and that is recorded too.
• Terms and coursesCreate a course with its code, section, term and department; a course starts as a draft, and is then opened, archived, or moved between departments.
• Sign in with a student/staff number or emailStudent/staff numbers are set or confirmed by administrators. New users set their own password through a one-time link from an administrator, and the same kind of link resets a forgotten password. AIshie sends no email: the administrator gives the link to the person.
• Single sign-onOpenID Connect is supported: administrators add, test, and switch identity providers on and off in the app, and the settings page has help for AD FS, Microsoft Entra ID, Google Workspace and Keycloak. A newly installed platform starts with password sign-in, and administrators connect it to the school’s accounts themselves.
• Quotas and costs of the school AI planOn the “AI and documents” page, set the models of the school AI plan, the daily quotas for each person and the ceiling for the whole school (in answers or dollars), and see today’s use and costs by day, person, agent and model. Changes take effect at once.
• Choosing who answers at OpenRouterFor a model of the school AI plan at OpenRouter, administrators choose which upstream providers may answer and which to try first, from OpenRouter’s own list with each one’s prices, country and whether it can call tools; a model newly added to the plan starts with only those that keep no data. The choice goes with every call, and costs count one price per model, whichever provider answered.
• Suspending a person or AI agentOnce suspended, they can do nothing in any course; their records and seats are all kept, and they can be reactivated later.
• Purging content uploaded by mistakeAdministrators can remove, for good, a version of material uploaded by mistake or a whole document, leaving a record of who removed it, when and why; grades are not affected.

Host it yourself

Runs on the school’s own server, with its data kept where the school chooses.

• One server, one command to installOne command installs the whole system, including automatic HTTPS, the database, the website and the AI agent runtime; running it again overwrites no setting and no data. An Ubuntu server with 2 cores and 4 GB of memory is enough to start.
• Automatic updates that roll back when something goes wrongThe server checks for a new version every five minutes: it backs up, updates and runs a health check, and if the new version fails the check, it goes back to the previous one. It can also follow only official releases instead.
• Automatic backupsThe database is backed up automatically before every update and every night.
• Files on the server or in the school’s cloud storageKept on the server’s disk by default, or in Amazon S3, Cloudflare R2, Backblaze B2 or another S3-compatible service; they can be moved later.
• Reachable from mainland ChinaWhere the server’s own address cannot be reached from mainland China, it can be put behind Cloudflare’s network, which visitors there do reach, and AIshie still sees each visitor’s own address. Reaching it through Cloudflare from mainland China can be slow at busy hours.
• Source-availableAll the source code is public under the Elastic License 2.0, so you can read the code you run; a school may run its own installation for its own staff and students, but may not offer it to others as a hosted service.

Everyday use

• Three languagesTraditional Chinese, Simplified Chinese and English. The first visit follows the browser’s language, and you can change it under “Language” in the account menu.
• On a phoneA side menu, a floating chat button, full-screen previews, and a single button for choosing files.
• Back closes what is open firstOn a phone, Back (the button or the gesture) closes the file preview, the chat, the menu or a side panel open over the page, and you stay on the page where you were. On a computer, the browser’s Back closes the file preview and side panels too.
• Light and darkLight, Dark, or System to follow your device.
• Approval notices and reasons for refusalsYou are told beforehand when an action needs approval; when something is refused, the reason is given in your language.

Planned

None of these is available today, and none has a date yet.

• Works inside your LMSOpen AIshie from Canvas, Blackboard and Moodle through LTI 1.3.
• Long-term memory for AI agents, on by defaultLong-term memory is built but off by default; it will be on by default once AIshie can erase memories when the seat or course they belong to ends.