This aims to serve as an alternative, but not a replacement, for Spermack. If Spermack works fine for you or you like its features there is no reason to use this beyond curiosity.
Credit to Barbiariskaa/Spermack for the original idea to interface with Claude in Slack in this fashion.
The main difference to Spermack is in the way the context is sent to Slack. Instead of sending messages directly to Claude as DMs, Slaude will instead create threads in the configured Slack channel. The entire workflow is:
- Split the context into multiple parts that each fit into a single Slack message without cutoff
- Send the first part to the configured channel
- Send all other parts as replies to the original message
- Ping Claude with a final message, triggering his response
When working with threads Claude receives the entire message history of a thread as context. However, Claude will not actually respond to a thread until specifically asked to by pinging him. Claude has no knowledge of anything going on outside of the thread he's been called to with a ping.
Editing your chat history in SillyTavern is still possible with this solution as each prompt creates a new thread. The amount of message spam won't really be any different than with Spermack, it'll just happen in a different channel and collapsed into threads.
It does feel like sending the context this way has given Claude a bit more coherency but this might be entirely imagined. Feel free to try for yourself.
First you will have to create a Slack workspace. You can do this for free at https://slack.com. Using an existing account and workspace is possible but very much not recommended. You'll also want to use open your workspace in a browser, not the Slack desktop app.
Once your workspace is set up, add the Claude app to your workspace. You will also need a channel to send the prompt message to but any of the ones that a new Slack workspace comes with by default will do.
Clone this repository or download it as a zip and extract it somewhere. You will also need Node and NPM installed but if you're using SillyTavern there is a good chance this is already the case so I'm not going to explain that here.
Open the config.js file with a text editor of your choice. You will have to fill in these values with information from your Slack workspace. The following will be described assuming you are using Chrome as a browser, but this should be possible with all of them with some minor differences.
With your Slack workspace open in a browser, press F12 to open the developer tools and switch to the Network tab. With the developer tools still open, send a Slack message in any channel. You can use this opportunity to send any message to Claude which will prompt you to accept its ToS as that is required before using it. With that done, look for an entry starting with chat.postMessage in the Network tab, open it and click "Payload". In the Form Data you should find the token, starting with xoxc-.
Still in the developer tools, switch from the Network tab to the Application tab. On the left look for Cookies -> https://app.slack.com then copy the value starting with xoxd-. After this we are done with the developer tools.
This is the ID of your Slack workspace. You can find this by clicking on the name of Slack workspace in the top left. It should show your workspace URL. The part before .slack.com will be your TEAM_ID. For example, if the URL is slaude-workspace.slack.com then slaude-workspace is the TEAM_ID.
The ID of the channel we want to start the prompt threads in. This can be any channel, but a good fit is the default #random channel that comes with your workspace by default. Whatever channel you go with, open it and click the little arrow at the top next to the channel name. You can find the channel ID at the bottom of the resulting popup.
Once you've picked a channel and copied its ID, send a message anywhere in the channel that pings Claude with "@Claude". This will open a dialog asking for confirmation that Claude should be invited to the channel. Click on "Invite them" and Claude will now be able answer to your prompts in that channel.
Open your Claude DMs and similarly to the above step click on the little arrow next to Claude at the top. The Member ID is what we're looking for here, not the Channel ID.
This lets you configure what gets sent in the message that we use to ping Claude. This seems to have a pretty huge influence on his response, similar to a jailbreak. The default message is "Assistant:" and will prompt Claude to continue the context we sent him to the best of his abilites. This has the usual side effects of responses getting longer and longer as the context gets larger. Feel free to experiment with different messages here but be aware that this can lead to unpredictable behavior. You also can't use SillyTavern replacements like {{char}} or {{user}}. If you need to refer to either, your best bet will be to stick to calling the character "Assistant" and the user "Human".
One thing that I personally tried was "Continue the story as Assistant. Keep your response brief and stay in character." and it resulted in consistently shorter responses, even in longer conversations, but it also had the side effect of all messages being a single paragraph, even if a longer response would have been more appropriate for the situation. Probably good if you want to stick to more of an actual chat conversation format.
This only needs to be changed if you have anything else running on the same port already.
It is recommended but not required to use the latest dev version of SillyTavern.
Run start.bat or start.sh depending on your system of choice or just do npm install and node app.js manually. It's not rocket science. Once the server is running you should see Slaude is running at http://localhost:5004 if you're using the default port.
In SillyTavern, click the API connections button and switch the API to OpenAI. Enter whatever you want in the API key field. The selected model doesn't matter either.
In the AI Response Configuration, paste the above URL into the Reverse Proxy field. I recommend creating a new preset for Claude. If you already have one for Spermack that should work fine. Set the context size to something sensible. Something between 4000-5600 should be safe. Max Response Length, Temperature, Frequency Penalty and Presence Penalty are all irrelevant and will be ignored, as will most other OpenAI specific settings. Streaming should work but I personally don't use it so I didn't test it that much.
What to use for your prompts is up to personal preference. I personally don't believe Claude needs any jailbreaks and never used any. If you do decide to send the jailbreak do not use the default SillyTavern jailbreak.
Once all that is set up press "Connect" back in API Connections. If everything was set up correctly the dot under the button should go green and say "Valid".
If you configured everything in Slaude correctly as well and made sure to accept Claude's ToS you should now be good to start prompting.
Note that you do not need to have Slack open in your browser for this to work. Once this is set up you can in theory just never open your workspace again.
Not tested if this works myself, but this guide was provided by a nice anon. This guide assumes you've already set up SillyTavern with termux.
Copy your config.js file from your desktop to your phone if you already have it set up there. Make sure you're using the same fork. If you don't have it setup then: Download the https://github.com/PandarusAnon/slaude zip off of GitHub using your phone's browser. (Use desktop mode code download zip). Unzip the contents, open config.js file in a text editor. Edit the contents of the file following the instructions on the git page and save. Use kiwi browser for access to developer tools. It's a hassle on mobile, just use desktop this one time copying the cookies and shit. If you don't have a text editor (wow) use Acode or FX explorer.
termux-storage-get requires the Termux:API addon. https://wiki.termux.com/wiki/Termux-storage-get Get it from here: https://www.f-droid.org/en/packages/com.termux.api/ and then run pkg install termux-api -y in termux. If you don't want to install an extra app, you can learn to use a terminal text editor (look up how to use nano or vim) or follow the guide here: https://wiki.termux.com/wiki/Internal_and_external_storage to connect your termux storage to FX Explorer and edit files from there.
Now open termux and run the following:
termux-setup-storage click allow storage access.
git clone https://github.com/PandarusAnon/slaude slaude use the git link for whatever fork you want to use.
cd slaude
npm install
termux-storage-get config.js Browse to the directory where you have your edited config.js file, and select the file to import it.
Now you have installed and setup all the files necessary. To run Slaude do: node app.js. This step may be called something else depending on which git you selected so refer to the git page for the command to run the server.
Now to run Silly, you need to run a separate Termux session. You can do that by pulling the sidebar from the top left. Tap and pull from the top left corner. Look up how to pull sidebars with gesture navigation on youtube if you're having trouble.
Click on new session. You're loaded into a new session now, in the slaude directory though so you'll have to change to the silly directory. Do: cd ~/silly_directory_name. Now run your silly server with the regular node server.js. Follow the API instructions on the git page, and voila!
To run Slaude + Tavern next time, open termux like usual. In the first session do: cd slaude && node app.js to run slaude. Switch to a new session and cd ~/silly_directory && node server.js.
>I can't open the sidebar!
Do: nano ~/.termux/termux.properties
Type the line: shortcut.create-session=ctrl + t anywhere in the file, just make sure you're writing on a fresh line and press ctrl+o, enter, ctrl+x.
Now to run a new session, simply press ctrl+t.
slaude now requests multiple times simultaneously so you have bigger chances of getting pass the filter (shorthand multi reply).
this only works if you're editting in the ping however. i.e. edit_msg_with_ping: true,
Before you set multi_response in the config to 1000, know that you'll fuck yourself over because:
* it seems like there's a limited amount of Claude responses at the same time for each workspace, so you'll have to wait for every request to finish to get your next ones
* requests might poison one another, if one takes too long to start, maybe
I also fixed the automatic retry, both on streaming/non-streaming and pinging/editting.
You can limit the number of retries. retry_count: 5,
- I think it didn't work at all if you were ping messaging (old way) and not editing the ping in (newest update way) (i.e. edit_msg_with_ping: false).
- added more timeouts, there's 3 total
- timeout if reply is taking too long to start being received
- timeout if reply is message is taking too long to update
- timeout if waiting just for the last few multi reply
Anons had reported, and I also had a bit of evidence that this new way of editing the ping in might somehow be filtering more than adding a ping at the end.
So to be clear, in the config:
when
edit_msg_with_ping: false
the prompt is sent the exact same way as old slaude
when
edit_msg_with_ping: true
instead of sending another message after your prompt, with the Claude ping and your PING_MESSAGE_PREFIX and PING_MESSAGE
The ping is edited in to the top of your prompt, before it. So it could technically go out of context even if you're lucky or put lots of tokens in PING_SUFFIX. (I actually will try doing that right now, dunno why I didn't yet)
Now never ignores responses, from possibly previous threads, from the same request. Something is weird with how events and timeouts race.
Thanks PandarusAnon for making this.