Manus is presented as a general action engine that can execute workflows through web, Slack, browser, Microsoft 365, and API interfaces.
2
The Manus API runs asynchronous tasks, supports follow-up messages, accepts files and connectors, and can notify applications through webhooks.
3
A practical research workflow combines private Notion data, uploaded documents, OCR, browser access, and generated reports or updates.
Summary
Ivan Leo presents Manus as an agent that can execute tasks rather than only answer questions. He demonstrates applications built through prompting, including a French-learning app, an event discovery site, email automation, browser-based coffee search, and dashboards generated from uploaded data. The workshop then focuses on the Manus API. Developers create asynchronous tasks, poll their status, continue the same task with a task ID, attach files or public URLs, use connectors such as Notion, and receive completion notifications through webhooks. A longer example connects Slack, uploaded receipt images, OCR, and a company policy in Notion to produce an expense response. Leo also shows how to keep Slack conversations tied to the same Manus task. He is candid about unfinished areas: browser permissions need more work, memory is not yet available, and some live coding fails. Future plans include exporting generated content to PPTX and PDF.
Manus is designed to execute tasks across the interfaces people already use
Ivan Leo describes Manus as an "action engine that goes beyond answers to execute task automate workflows extend human reach." The goal is a general agent that users can access in different settings. He lists the web application, Slack app, API, browser operator, Microsoft 365 integration, Mail Manus, and the iOS app. The Microsoft 365 integration is intended to help edit PowerPoint files and repair Excel templates. Leo says the team built these entry points so users can work with Manus through a mailbox, Slack, a custom programmatic workflow, or a mobile app.
Prompting Manus can produce complete web applications with data and integrations
Leo shows a French-learning application he created by prompting Manus. The app corrects writing inline, generates a corrected version, explains words, and uses a language model to build a profile of the learner. He then shows an event site created by asking Manus to scrape conference events, place them in a JSON file, and build a searchable interface with calendar links, starred events, and a personal timeline. He supplied Chroma and OpenAI keys for the project. Because the web apps use a full Docker image, he says developers can install packages, Redis, Stripe, webhooks, and other custom components.
The API exposes the same agent capabilities through asynchronous tasks
Ivan introduces the Manus API as a way to use the agent inside another application or workflow. A request returns a task ID, task title, and task URL. The task ID lets an application send clarification or follow-up input to the same session. Tasks can be running, pending, completed, or in an error state. When no chat mode is specified, a backend router chooses between simple chat and the full Manus agent. Developers can also select Manus 1.5 or Manus 1.5 light, with 1.5 recommended for complex work and 1.5 light for simpler, faster queries.
Files, URLs, images, and connectors give research tasks their working context
The API accepts uploaded files, public URLs, and base64-encoded images. Uploaded files receive a file ID and an S3 link, and Leo says the API deletes uploaded files automatically after 48 hours unless they are deleted sooner. Manus can process PDFs and other multimodal content. He demonstrates a Rick and Morty JSON dataset that is turned into a website, and a Berkshire Hathaway investor letter that is summarized and analyzed. Existing connectors, including Gmail and Notion, can be added by copying a connector UID into the task payload.
Webhooks are the practical way to run many longer tasks
Polling is the simplest way to prototype an API integration, but Leo recommends webhooks as the number of tasks grows. The application registers an endpoint, and Manus sends notifications when a task starts and when it stops. He says a typical task takes roughly three to five minutes, so repeatedly polling every task would create unnecessary workers and cost. The webhook payload includes the task ID, task URL, status, and output. The receiving application can then request the full conversation history or process the completed result.
A Slack integration needs fast acknowledgement and task mapping
The workshop builds a Slack bot with a public endpoint deployed through Modal. Slack sends events to the endpoint, which must respond in roughly three seconds or Slack will retry the request. The integration creates a Manus task from the Slack message and posts a link back to the live task. To support multi-turn conversations, the application stores a mapping from the Slack thread timestamp to the Manus task ID. Later messages in the same thread are sent to the existing task. Leo also converts Manus Markdown into Slack-compatible blocks and uploads files to the correct channel thread.
Private research workflows can combine receipts, OCR, and company policies
Leo demonstrates a more involved workflow using a receipt image, Slack, and a Notion connector containing a fictitious company expense policy. Manus receives the file, extracts its details with OCR, reads the policy in Notion, and produces a response about the expense. The agent identifies the merchant and amount, then updates the relevant Notion page with a Markdown table. Leo presents this pattern as a way to build internal research tools that search company information and return answers based on editable internal policies.
Memory and browser permissions remain unfinished parts of the product
During the Q&A, Leo discusses using Python and six Selenium instances to search Singapore's government pickleball booking site. He says this worked because the agent had its own sandbox and could write and run code. A participant asks whether the API can use their personal browser. Leo says browser access requires choosing a browser and approving actions because users should not have random tabs opened for them. He says a better permission system is on the roadmap. He also says persistent memory is being explored but is not currently available, so users must provide context explicitly.
