Short answer: Semibot is a desktop AI agent installed on your computer—not a web chat box. Conversations, files, knowledge, and run history are stored in a local SQLite database. The same workbench handles research and writing, and can open a real code repository to search code, edit files, review changes, and run tests.
What Semibot is
- A desktop app (Electron + TypeScript local core) with an AI runtime adapter. It runs on macOS (Apple Silicon) and Windows x64.
- Office-capable: built-in browser research, document and slide writing, image generation, and a knowledge library that organizes materials into source-linked wiki articles.
- Coding-capable: opens real projects, scans the codebase in the background, plans changes, edits files, runs commands and tests, and delivers a ChangeSet you can review.
- Local-first: core data stays on your machine. Secrets go to the OS keychain, not plaintext config.
What Semibot is not
- It is not a web-only chat interface.
- It is not an IDE plugin—coding happens in the same workbench as office work.
- It is not “AI that only suggests.” Intermediate steps (browsing, writing, editing, running commands) execute within the access you grant.
Quick comparison
| Dimension | ChatGPT Desktop | Claude Desktop | Cursor | Semibot |
|---|---|---|---|---|
| Primary shape | Chat | Chat | IDE | Desktop workbench |
| Office writing / research | Strong | Strong | Limited | Built-in |
| Repo-level coding | Limited | Limited | Strong | Built-in |
| Local data default | Cloud | Cloud | Mixed | Local SQLite first |
| Continuous follow-up | Reminders | Limited | — | AI secretary |
| Folder grants + checkpoints | — | — | Partial | Yes |
Positioning summary: chat-first tools answer well; IDE-first tools code well. Semibot is for people who must produce work and run projects in one place. The trade-off is breadth vs depth—a dedicated IDE will always have deeper coding features, and a dedicated chat app will always have a more polished conversation experience. The workbench model prioritizes reducing context-switching across different kinds of work.
How the workbench works
- You describe a goal in the composer.
- Semibot plans and executes intermediate steps—web research, file writing, code edits, commands—within granted folders and approval gates.
- Results land as artifacts (documents, reports, slides) or as a coding ChangeSet (files, diffs, commands, tests).
- Knowledge and memory persist locally so the next session picks up context without re-explaining.
How connectors work
Semibot connects to external services through connectors—each connector is a bridge between Semibot and a specific platform. Current connectors include Feishu/Lark, DingTalk, Discord, Telegram, Slack, Gmail, calendar, local file watchers, BlueBubbles/iMessage, and MCP tool servers. Connectors are managed in a single connectors page and can be enabled or disabled individually.
Important: MCP tool connectors do not automatically inherit your workspace data or system credentials. That boundary is designed to prevent an MCP server from reading your entire workspace without explicit permission.
Connectors are individually toggleable—you enable only the ones you need and can disable any connector at any time. Each connector operates within its own permission scope, and the secretary can use connectors to deliver results to the platforms where your team already communicates.
The knowledge library
Semibot’s knowledge library has three modes in one workspace: Browse (real folder tree with file preview), Search (hybrid semantic search across one or all workspaces), and Wiki (continuously organized topic articles with traceable sources). The Wiki mode is not a one-shot report—it grows over time as materials are added and organized.
The knowledge library is local-first: its contents live in the local SQLite database. If you later switch to another tool, the raw files are still on your machine.
Workflow scenarios
To make the description concrete, here are three real workflows that show how the workbench operates in practice.
Scenario 1: You inherit an unfamiliar codebase and need a bug fix in 2 hours
- Open the project folder in Semibot. The background scanner indexes the code graph while you keep working.
- Describe the bug in the composer. Semibot uses the code graph to locate the relevant modules and trace the call chain.
- It proposes a plan: which files to edit, which tests to run. You review the plan before anything is written.
- After approval, Semibot edits the files and presents a ChangeSet—every touched file with a diff, the commands it ran, and test output with exit codes.
- You accept or roll back individual files from the checkpoint. No git history required.
Scenario 2: Weekly client brief from scattered sources
- Set up a knowledge workspace and add folders containing client documents, meeting notes, and email exports.
- Tell the secretary: "Every Friday at 17:00, summarize client A's weekly progress—what moved, what risks, what to plan next."
- The secretary checks the workspace for changes at the scheduled rhythm, compiles a digest, and places it on the secretary page.
- If it needs clarification—a missing document, an ambiguous status—it pings you. Otherwise the digest appears without interruption.
- Over time, the wiki mode organizes these digests into an evolving source-linked article about client A.
Scenario 3: Research a topic and produce a slide deck
- Ask Semibot to research a topic using the built-in browser. It reads pages, extracts key claims, and saves source-linked notes to the knowledge library.
- Request a slide deck from those materials. Semibot drafts slides with the content, pulling quotes and data points that link back to their sources.
- Review and refine in the same workbench. The knowledge library retains the research trail so you can verify any claim.
Decision tree: which tool fits?
There is no universal answer. Use this decision tree to match a tool to your actual workflow:
- If your work is mostly conversation and writing with no code projects, ChatGPT Desktop or Claude Desktop will serve you well. Both are polished, mainstream, and have large communities.
- If your work is mostly coding and you rarely leave the editor, Cursor is built for that tight editor loop. It is mature and widely used by developers.
- If you need both office work and code projects in one place, and you care about keeping data local, Semibot is designed around that combination. Be aware of its younger ecosystem and platform caveats.
- If you need the absolute deepest IDE integration (custom extensions, language servers, debugger integrations), a dedicated IDE with AI features is a better fit than any workbench.
- If you need full offline operation, no current cloud-model-based tool will work. You would need a local model setup, which is a separate infrastructure decision.
Real connector examples
Connectors turn Semibot from a single-user tool into something that touches your existing workflows. Each connector is a bridge to a specific platform, managed in a single connectors page and individually toggleable. Here are concrete examples:
- Feishu/Lark: Pull messages and documents from team channels. Useful for a secretary task that monitors a project channel and summarizes weekly activity. Write-back lets the secretary post digests into the channel.
- Gmail: Draft follow-up emails after meetings. The secretary can watch for calendar events and prepare reply drafts based on the meeting notes in your workspace.
- Calendar: Meeting preparation. Before a meeting, the secretary reviews related documents in the workspace and produces a one-page brief with action items from the last meeting and any new materials.
- Slack / Discord / Telegram: Team communication monitoring. Set a secretary watch on specific channels for mentions of a project or keyword, and collect relevant threads as notes in the knowledge library.
- DingTalk: Enterprise messaging workflows. Pull task assignments and status updates into the workspace so the secretary can track project progress across tools.
- BlueBubbles/iMessage: For personal workflow integration, monitor messages related to specific contacts or topics and surface relevant context.
- MCP tool servers: Connect external tool APIs. MCP connectors do not inherit workspace data or system credentials by default—you explicitly control the boundary.
Boundary conditions: when does this not work?
- Very large codebases (10,000+ files): The code graph indexes in the background, but extremely large monorepos will take longer to scan and may exceed practical context limits for any single task. The graph narrows what the agent reads, but cannot eliminate context constraints entirely.
- Network-dependent model calls: If you are on a flight or in a low-connectivity environment, cloud model inference is unavailable. Semibot does not currently ship a built-in local model—you would need to configure a local provider yourself.
- Real-time collaboration: Semibot is a single-user desktop app. There is no multi-user live editing or shared cursor. If your workflow requires real-time co-editing, you need a cloud-native tool.
- Mobile access: There is no mobile client. The desktop workbench is where all work happens. Connectors let you receive results on mobile platforms (e.g., via messaging), but you cannot run tasks from mobile.
Limitations to be aware of
- Windows builds are currently unsigned. The installer may trigger OS security warnings. This is a trust and convenience problem worth acknowledging.
- Linux desktop is still on the roadmap. Not yet shipped.
- Cloud model calls need network. Local-first is about where data is stored, not about offline operation.
- Younger ecosystem. Fewer third-party reviews, plugins, and community patterns than established tools. Evaluate it on its own claims.
- Single-app trade-off. Combining office and coding in one workbench means neither gets the full depth of a specialized tool. If you only code, Cursor is deeper; if you only write, Claude Desktop may be more refined.
Who it is for
- Solo developers who need to read unfamiliar repos and ship fixes.
- Business and research roles that need scheduled briefs and material tracking.
- Teams that want one shared assistant with visible, confirmable, auditable steps.
- People who care about local-first data and are willing to use a less mainstream tool that trades ecosystem size for data control and workflow breadth.
Who should skip it: if you only want casual small talk, a chat app is a better fit. If you only code in an IDE, Cursor is deeper. If you want the most polished chat experience with a massive plugin ecosystem, ChatGPT Desktop is more mature. Semibot is built for people who need to finish work across both documents and code.
FAQ
What is Semibot?
A local-first desktop AI agent for office work and real code projects. See the definition above.
How is it different from ChatGPT Desktop or Cursor?
Chat-first vs IDE-first vs workbench-first. Semibot adds an AI secretary that keeps following up, a knowledge library that grows into wiki articles, and coding behind folder grants, checkpoints, and approvals.
Does my data stay on my machine?
Yes by default: local SQLite for conversations and knowledge; secrets in the OS keychain. Data only leaves when you sign in, call a cloud model, or use a connector.
Is it free?
The client is free to download. First launch creates a free account with trial quota; you can also bring your own model provider.
How long does setup take?
Download and first launch take a few minutes. The free trial quota is available immediately—no API key required to start. Configuring your own model provider or connectors adds a few more minutes per integration.
Can I use Semibot offline?
Not for model inference with cloud providers. Local-first is about where data is stored (local SQLite, OS keychain), not about offline model execution. You would need to configure a local model provider for any offline capability, which is a separate infrastructure decision.
What model providers does it support?
Semibot supports OpenAI-compatible and Anthropic-compatible API endpoints, so you can bring your own key from common providers. The built-in trial uses a hosted model so you can try the product before configuring anything.
Can I export my data?
Conversations, knowledge, and run history live in a local SQLite database on your machine. You can access the database directly using standard SQLite tooling. There is no proprietary lock-in on the storage layer.
How does it handle large projects?
Semibot scans codebases in the background using a native TypeScript code graph engine. The graph narrows what the agent needs to read, but very large monorepos will take longer and may hit practical context limits for any single task. The workbench does not freeze during scans.
Does it force Git?
No. Git is optional source control. Semibot detects Git if present and uses it for version control features, but file checkpoints work independently. Approving a git commit never implicitly grants folder access.
Can teams share a Semibot instance?
Semibot is a single-user desktop app. There is no shared server mode. Multiple team members can each run their own instance, and connectors (like Feishu/Lark or Slack) let them share results through existing communication channels.
What about migration from other tools?
There is no one-click import from ChatGPT or Claude. You can bring context by adding documents to a workspace, importing files into the knowledge library, or pasting previous conversations. The knowledge library can index local files directly—so any exported data from other tools can be added as source material.