Tecent Octop (Opensource): Wait, Tencent just ENDED OpenAI Dots & Muse!?

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Tecent Octop (Opensource): Wait, Tencent just ENDED OpenAI Dots & Muse!? · Channel: AICodeKing · 12:21 · Watch on YouTube

Topics: AI agents, Task and workflow automation, AI developer tools, Local and open-source models, RAG, search and information retrieval

Summary

The video walks through setting up Octop, an open-source Tencent Cloud AI assistant, and building a workflow with specialist agents, document search, team collaboration, coding tools, and scheduled tasks. The demonstration uses fictional Taskboard release notes to check whether answers and reviews stay grounded in the source. The presenter highlights practical considerations, including model costs, data sent to cloud providers, setup requirements, and beta team features. The recommendation is to begin with one expert, one model, and one small document, then add capabilities as needed.

Key points

Octop is open-source under the MIT license, but users provide their own models and may incur cloud API costs.
Self-hosting Octop does not mean every model request is local; information sent to a cloud provider leaves the machine.
The walkthrough covers installation, initial configuration, and creating an expert with a specific role.
A local embedding model can support document search, but it is separate from the chat model.
The Taskboard release-notes example checks for accurate summaries, source references, and unsupported claims.
Expert teams can pair a writer and reviewer, but teams are marked beta and involve additional model calls.
Octop can connect to an external coding agent through ACP; the coding tool needs its own installation and configuration.
Scheduled tasks can send fixed text or ask an agent to perform work, and the Octop process must remain running.
The presenter recommends starting small and adding reviewers or coding tools only when useful.

Chapters

Start	Chapter	Summary
00:02	Introduction and Octop overview	The presenter introduces Octop and outlines its combination of expert agents, documents, scheduled tasks, and coding-tool connections.
01:37	Installation and initial setup	The walkthrough covers installing Octop, creating an admin account, configuring a model provider, and opening the dashboard.
03:41	Creating an expert and testing document search	A release-helper expert is created and tested against fictional Taskboard notes using a knowledge base.
06:44	Teams, reviewers, and coding agents	The presenter creates a writer-reviewer team, tests it against inaccurate claims, and demonstrates delegating file work to an ACP coding agent.
08:15	Scheduled tasks and optional integrations	The video explains recurring reminders, task modes, keeping Octop running, and optional messaging, connector, and browser features.
10:50	Assessment and recommendation	The presenter weighs Octop’s flexible workflow against its setup, costs, and beta features, and recommends building up gradually.

Quotes

It brings together specialist agents, your documents, scheduled tasks, and connections to coding tools like open code and Claude code. — 00:35

Octop itself is free and open-source under the MIT license, but you still supply the models. — 01:06

Agent teams is still marked beta. — 08:15

One expert, one model, and one small document. — 11:51

Concepts and entities

Octop, Tencent Cloud, Ollama, knowledge base, embedding, Agent teams, ACP, OpenCode, Claude Code, SQLite

Full transcript (original language)

00:02 Hi, welcome [music] to another video. Just before we start the video, I have recently posted a video on the membership about how I am using Opus 5.5 and it is almost a tutorial of how to get the best stuff out of it. I am also going to become more active on the members only video now as well. So, if you like what I do and want to support the channel and also get some perks along the way, then I’d highly recommend you to hit the join button and choose a level. The exclusive videos are available on all the tiers from bronze and above. Now, back to the video.

00:35 So, today we’re looking at Octop, an open-source AI assistant from the Tencent Cloud GitHub organization that you can run on your own machine. It brings together specialist agents, your documents, scheduled tasks, and connections to coding tools like open code and Claude code. And the part that interests me is having those things available in one place. You can ask an expert about a document, bring in another expert to review the answer, and hand a coding task to a separate tool without rebuilding the setup every time. So, in this video,

01:06 I’ll walk you through installing it, connecting a model, creating your first expert, and putting together a small release preparation workflow. We’ll give it some sample project notes, check whether it answers accurately, and then try the team and coding agent features. Octop itself is free and open-source under the MIT license, but you still supply the models. So, cloud API usage can cost money. It also supports Ollama for local models, although the hardware requirements and the quality of the results depend on what you run. And self-hosting the app doesn’t automatically make every model request local. If you connect a cloud provider,

01:37 the information you send to that provider still leaves your machine. That’s something to understand before you start uploading your documents. So, let’s get right into it. I’m going to use the terminal installation and the web dashboard for this walk-through. There are desktop downloads for Windows, Mac, and Linux as well, which you’ll find on the project’s releases page. First, open the GitHub repository linked below and scroll down to the installation instructions. On Mac or Linux, copy the shell installer command into your terminal. On Windows, use the PowerShell command instead. I’ve included the exact commands alongside

02:07 the walk-through, so you can copy. The installer handles the Python environment for you, so you don’t need to set up Python manually. Once it finishes, open a new terminal so that the octop command is available. Now, the normal command to start the app is octop run. For this walk-through, I’m using the command on screen that also sets octop’s default time zone to New York. Replace that with your own time zone, especially if you plan to use scheduled tasks. The project’s default is Asia/Shanghai. Keep that terminal running, then open

02:38 the local address shown on screen, which is your own machine on port 8088. For a fresh installation, we’ll complete the setup in the browser. If it asks for a setup password, use the temporary password printed by the server in your terminal. That gets you into the setup process, where you create your regular admin account. Choose SQLite for the database in this walk-through. Then, enter your admin username and a strong password. The password needs at least eight characters, including letters and numbers. Next, configure the model provider. Pick a provider you already have access to, enter its API key, and check the API address. If you’re using a

03:09 custom service, choose the matching API type and enter the base URL supplied by that service. Then, enable the model you want to use and test the connection. I’m keeping this to one chat model initially because I want to confirm that the basic conversation works before adding more moving parts. If the connection fails, check the key, the model name, and whether the account has available quota. Installing octop doesn’t give you a balance with the model provider. You can also connect an Ollama model running on your machine. For this walk-through, whichever model you choose needs to handle tool calls reliably because we’ll be asking it to search documents and delegate work. Finish the setup and we

03:41 can move into the dashboard. Now, octop calls its agents experts. Each expert has its own workspace and configuration, so you can give one a writing role and another a review role. You can also choose different models for different experts. Open the experts page, select my experts, and click new expert. Name it release helper, and pick the model we just configured. Expand prompt files, and edit the agents.md file. I’m asking this expert to write clear release summaries, use the supplied project information, and say

04:12 when something isn’t documented. Save those instructions, and create the expert. There are templates and personality options as well, but for this example, a clear job description is enough. I want this expert to understand what a useful answer looks like. Start the expert, if it isn’t already running. Open its chat, and ask it to explain its role in one sentence. Once you get a normal response, we can give it something more useful to work with. I’ve prepared a small fictional set of release notes for an app called taskboard. This is demo data, and I’ve included the

04:42 file, so you can follow along with the same example. The update adds a CSV export, and a task filter. It also adds a keyboard shortcut for creating a task. But, there’s a known export bug. Device sync isn’t supported, and there isn’t an announced release date. Those details give us something specific to check. If the assistant invents a release date, or says the app has device sync, we’ll know it has gone beyond the document. Now, open knowledge bases, and go into knowledge base settings. Enable the feature, then choose how you

05:12 want to create embeddings. Embeddings are what let the app search for relevant passages in your documents. For this example, choose local on NX, and download an English embedding model from the available list. Save the settings once the model is ready. You can use an online embedding service instead, but that sends document content to that service for processing, and may add another usage cost. This embedding model is separate from the chat model we configured earlier. Running embeddings locally doesn’t change where your chat model runs. Create a knowledge base called taskboard

05:42 demo, and upload the sample release notes file. Wait until the document shows ready, then use add to chat and make sure we’re talking to release helper with that knowledge base selected. So, let’s send it this. Search the task board demo knowledge base and summarize the changes in plain English. Include the known limitations, site the source, and tell me whether device sync is supported and whether a release date has been announced. Give it a moment to search, then open the source reference in the answer and compare it with the file. The answer we’re looking for should

06:13 mention the CSV export, the filter, and the shortcut. It should also explain that the export can include archived tasks. Device sync should be listed as unsupported and the release date should be described as unannounced. If it misses something, ask it to search again and point to the passage it used. This is why I like starting with a tiny document. You can check the whole answer yourself before trusting the same workflow with a much larger collection. Once that works, you could use this for

06:44 internal documentation, project notes, or support material. I think this is a useful starting point because it gives the assistant information that actually belongs to your project. Now, let’s add a second expert called reviewer. Give it the job of checking drafts against the supplied source and calling out unsupported claims. Configure its model, start it, and check that it responds as well. Then open my teams from the experts page and create a team called release team. Add release helper and reviewer as the two members and choose a model for the

07:14 coordinator. The coordinator handles the delegation and brings the results together. The two members do the specialist work. You need at least two members in addition to the coordinator and those members need to be running before you ask the team to work. For this first team demo, I’m pasting the short release notes directly into the team message. Each member has its own workspace, so we’ll give the team the source explicitly. Here’s the request. Have release helper her a short release announcement using these notes. Have reviewer check every product claim against the same notes,

07:45 then return a revised announcement and explain any corrections. Include the source notes in each assignment. Now watch the team conversation as the work comes back. Check who is speaking, what the reviewer flagged, and whether the coordinator’s final answer reflects that feedback. To make the review more interesting, we can supply a deliberately inaccurate draft that says task board now syncs across devices and launches tomorrow. Then ask the reviewer to correct it using the notes. That’s a much better check than asking two agents whether they agree with each other. We already know which claims are wrong, so we can see whether the review catches them. Agent teams is still marked beta.

08:15 And there is a practical limitation here. Stopping the coordinator’s current response doesn’t necessarily stop work already dispatched to the members. Let those jobs finish and check the member activity before restarting the app. I like the idea of having a writer and a reviewer available together, but there are extra model calls involved. So for a simple question, I’d stick with one expert. Add the team when the second role has a useful job to do. Now let’s come to the coding side because Octop can also delegate work to an external coding agent through ACP, which stands for agent client protocol. I’ll use open

08:45 code for this example. It needs to be installed and configured on the same machine that’s running Octop. Make sure the Octop process can find the open code command or set the runner’s command to the full path of the executable. Open the ACP page, enable the open code runner, then select release helper and turn on its ACP runner tool. Create a scratch project folder and put the demo release notes inside it, then ask release helper to use open code through the ACP runner to read that file and create an announcement.md draft in that folder. Give it the full folder path and

09:16 tell it to include the known limitations. We’re keeping this task small so we can check that open code can read and write files in the project folder. If a permission request appears, review it before continuing. Then inspect the file the coding agent created and compare it with the source. Check the actual file even if the chat says the task is finished. The runner connection doesn’t include free access to the coding tools models. Open code still needs its own working configuration and other runners have their own requirements as well. So, this gives you a way to connect a document and chat workflow to a tool that can work on project files. For a larger

09:48 task, you could ask the coding agent to update documentation in a repository and then review the changes yourself. Now, let’s make part of the workflow repeatable. Switch back to the regular release helper expert and ask it to create a weekday reminder at 9:00 in the morning to review the task board release checklist. Ask for a fixed text reminder delivered to this dashboard conversation. Then, open the scheduled tasks page and inspect what it created. Check the schedule, the time zone, and the destination. A message saying it created a reminder is only useful if there’s a corresponding task in the app. Octop has

10:19 two useful task modes here. Text sends a fixed message. Agent asks the model to do work when the task runs. A reminder can use text, while a fresh summary of your project notes would need agent. Use execute now to test the reminder and check the destination conversation for the message. Once you’ve tested it, disable the example schedule if you don’t want to keep receiving it. And remember that the Octop process needs to stay running for these tasks. If you want it in the background, the project provides the Octop service start

10:50 command. Make sure the background service also uses the time zone you intended. There’s more in the dashboard, too. You can connect messaging channels, add connectors, and use the browser tools. Those features have their own setup, so I’d add them after this basic workflow is working. For an optional browser test, install the browser component. If Octop asks for it, open browser AI plus and ask the assistant to open the Octop repository and take a screenshot. Then, compare the screenshot and the visible page with what it reports. That gives you a small task to check before trying

11:21 a longer browser workflow. So, what do I think about Octop? The combination is interesting. You can keep a few experts configured for specific jobs, give them project documents, and bring them together when a task needs more than one role. The coding agent connection makes that especially relevant if your work regularly moves between research, writing, and a repository. There is setup involved, though. You’ll manage model credentials, decide how document search runs, and keep the app available for scheduled work. The team features are still beta, and more agents can also mean a higher bill.

11:51 Personally, I’d start exactly where we started here. One expert, one model, and one small document. Get a useful answer, check it, and then add the reviewer or coding runner when you have a reason to use it. If you want that kind of control over your assistant setup, I think OctoP is worth trying. The source is available. You can run it yourself, and you have room to build up the workflow gradually. Overall, it’s pretty cool. Anyway, let me know your thoughts in the comments. If you like this video, consider donating through the Super Thanks option, or becoming a member by

12:21 clicking [music] the join button. Also, give this video a thumbs up, and subscribe to my channel. I’ll see you in the next one. Until then, bye.