How to connect SillyTavern to KoboldCpp

KoboldCpp SillyTavern setup in short: start KoboldCpp with a model, then in SillyTavern open API Connections (the plug icon), choose Text Completion, set the API type to KoboldCpp, enter http://localhost:5001 and click Connect. Leave the API key empty unless you started KoboldCpp with --password: SillyTavern shows a red Missing key marker anyway, and the connection works. For Chat Completion use the custom OpenAI-compatible source with http://localhost:5001/v1. Menu names checked in SillyTavern 1.19.0.

Checked against KoboldCpp 1.124.2 · updated 13 Nov 2026
  • lab-lnx: Linux test VM · 4 vCPU · 15 GB · no GPU
  • gha-u24: Ubuntu 24.04 test machine · 4 vCPU · 16 GB · no GPU
How we test
On this page
  1. Before you connect: two checks
  2. KoboldCpp SillyTavern: Text Completion or Chat Completion?
  3. Connect with Text Completion (recommended)
  4. Connect with Chat Completion
  5. API key and password
  6. Match context size and response length
  7. Pick the right instruct template
  8. SillyTavern on your phone, KoboldCpp on your PC
  9. When it does not work
  10. Images, voice and more from KoboldCpp
  11. Check it worked
  12. Undo

Before you connect: two checks

Before you connect 0 of 3 done

If the page does not open, fix that first: Will not start or Troubleshooting. You can also check from a terminal: curl http://localhost:5001/api/v1/model should answer with the model name (API page). SillyTavern itself is installed from its own documentation.

KoboldCpp SillyTavern: Text Completion or Chat Completion?

The two ways SillyTavern can talk to KoboldCpp
CriterionText Completion → KoboldCppChat Completion → Custom
Who builds the promptSillyTavern, with its instruct templateKoboldCpp, with the model's chat template
Addresshttp://localhost:5001http://localhost:5001/v1
Connected on 1.19.0Yes, model name shown works F48Yes, "Valid", model listed works F48
A chat reply through itReceived works F48Not exercised untested

Our recommendation: use Text Completion with the API type KoboldCpp. It is the one we got a reply through, and SillyTavern then knows it is talking to KoboldCpp. Choose Chat Completion if you want KoboldCpp to apply the model’s own template (OpenAI API page).

Connect with Text Completion (recommended)

  1. Open API Connections

    Click the plug icon in SillyTavern’s top bar. On a first start SillyTavern asks for a persona name before anything else works.

    First start of SillyTavern: you must choose a persona name before connecting anything.gha-st-firstrun-01 · gha-u24 · KoboldCpp 1.124.2 · 13 Nov 2026
    SillyTavern welcome dialog asking for a persona name
    The plug icon opens API Connections.gha-st-connect-01 · gha-u24 · KoboldCpp 1.124.2 · 13 Nov 2026
    SillyTavern API Connections panel
  2. Choose Text Completion and the API type KoboldCpp

    Set API to Text Completion, then API Type to KoboldCpp. The URL field shows the example http://127.0.0.1:5001, and the API key is marked optional F43.

    Text Completion with API Type set to KoboldCpp: the API key is optional and the example URL is http://127.0.0.1:5001.gha-st-connect-02 · gha-u24 · KoboldCpp 1.124.2 · 13 Nov 2026
    SillyTavern Text Completion settings with KoboldCpp selected
  3. Enter the address

    Type http://localhost:5001 (or http://127.0.0.1:5001: both worked on the same PC F48).

  4. Click Connect

    A green dot and the model name appear, such as koboldcpp/qwen2.5-0.5b-instruct-q4_k_m F43. If a list of earlier addresses covers the Connect button, press Escape first F47.

    Connected: the green dot shows the model KoboldCpp has loaded.gha-st-textcomp-correct-01 · gha-u24 · KoboldCpp 1.124.2 · 13 Nov 2026
    SillyTavern connected to KoboldCpp showing the loaded model name

Connect with Chat Completion

Choose Chat Completion, source Custom (OpenAI-compatible), and use the base address http://localhost:5001/v1. SillyTavern reported “Valid” and listed the loaded model F48:

Chat Completion with the documented /v1 base URL.gha-st-chatcomp-correct-01 · gha-u24 · KoboldCpp 1.124.2 · 13 Nov 2026
SillyTavern Chat Completion custom endpoint with /v1

SillyTavern’s own hint says to add /v1 if the address does not work; in our test the address without /v1 was also reported valid F48:

Chat Completion, Custom (OpenAI-compatible): the base URL validated and listed the model even without /v1.gha-st-chatcomp-missing-v1-01 · gha-u24 · KoboldCpp 1.124.2 · 13 Nov 2026
SillyTavern Chat Completion custom endpoint validated against KoboldCpp

The model list shows koboldcpp/<model file>. We did not send a message through this mode, and we have not checked whether SillyTavern applies its own instruct template in it. See the OpenAI API page for what KoboldCpp does with chat requests.

API key and password

There is no key unless you started KoboldCpp with . SillyTavern marks the key field with a red “Missing key” whenever it is empty, even though the connection works and no key is needed F43:

A URL ending in /v1 also connects in KoboldCpp mode; the red "Missing key" marker appears although no key is needed.gha-st-textcomp-v1-suffix-01 · gha-u24 · KoboldCpp 1.124.2 · 13 Nov 2026
SillyTavern connected with a /v1 URL and a Missing key marker

If you did set a password, paste it into the key field of the connection you use: KoboldCpp answers generation requests with 401 without it F25. A 401 after connecting means the key is missing or wrong. See the password table.

Match context size and response length

KoboldCpp’s context is set when it starts: 16384 by default, with 256 added F05 F06. SillyTavern has its own Context and Response sliders. Keep SillyTavern’s context at or below KoboldCpp’s and the response length well below it. We have not tested how SillyTavern reads the server’s value or what happens when its number is larger. For reprocessing and ContextShift in long chats, see context size.

Pick the right instruct template

In Text Completion SillyTavern formats the prompt, so its instruct and context templates must match the model’s family. A mismatch shows as the model speaking for you, role tags in replies, endless output or visible thinking text (users report all of these). We have not tested SillyTavern’s template selection. See chat templates when it is published.

SillyTavern on your phone, KoboldCpp on your PC

localhost means the device you type it on, so on a phone it is the phone. Use your PC’s address on your network instead, such as http://192.168.1.20:5001, and test it in the phone’s browser first: the KoboldAI Lite page should open. Windows may ask to allow KoboldCpp through the firewall. We have not tested this on a phone. The steps for finding the address and the risks are in remote access; for KoboldCpp running on the phone itself see Android.

When it does not work

SillyTavern connection problems
"No connection..." and a red plug iconKoboldCpp is not running, the port is wrong, or you typed https:// for a normal server. A wrong port (5002) and https both gave this.tested, F48
"Not connected to API!" in the chat boxNot connected yet, or the connection dropped. Click Connect again.tested, F47
A red "Missing key" next to the API key fieldHarmless without --password: the connection works.tested, F43
A list of addresses covers the Connect buttonSillyTavern keeps a history of addresses you tried. Press Escape or click elsewhere first.tested, F47
Nothing works on the first startSillyTavern asks for a persona name before anything else: enter one.tested, F43
401 or Unauthorized when sending a messageKoboldCpp has a --password and the key field is empty or wrong: password table.tested, F25
"Not Found" or "endpoint does not exist"The address or API type does not match: URL mistakes.tested, F23 (the KoboldCpp message)
The old model name stays after you swap models in KoboldCppClick Connect again to refresh it.users report
Replies are cut shortSillyTavern's response length is too low.users report
Every reply reprocesses the whole chatThe start of the prompt changes (lorebook entries, a changing character note) or the context is mismatched: ContextShift.tested for the mechanism, F66

A second tab or a second chat waits its turn, because KoboldCpp answers one request at a time by default F26.

Images, voice and more from KoboldCpp

KoboldCpp can also serve image generation (A1111-style routes) and speech to apps (API). We have not tested SillyTavern’s image or voice features against it, so we do not say they work. See images and voice when those pages are published.

Check it worked

Send a message in a chat. A reply arrives, like this one through the connected backend:

A first reply through SillyTavern from KoboldCpp.gha-st-chat-01 · gha-u24 · KoboldCpp 1.124.2 · 13 Nov 2026
SillyTavern chat with a reply generated by KoboldCpp

The KoboldCpp console prints the request and its CtxLimit status line (how to read it).

Undo

Click the plug icon and disconnect, or clear the address. Nothing changes in KoboldCpp. If you opened the firewall or used for a phone, close it again: remote access.

Frequently asked questions

Does SillyTavern need an API key for KoboldCpp?

No. KoboldCpp has no key unless you start it with --password. SillyTavern's KoboldCpp page shows the field as optional but marks it with a red Missing key even when the connection works, which is a common source of confusion. Only if you set --password do you paste that password into the field.

Text Completion or Chat Completion for KoboldCpp?

Text Completion with the API type KoboldCpp is the simplest: SillyTavern then knows it is talking to KoboldCpp, and it formats the prompt with its own instruct template. Chat Completion lets KoboldCpp apply the model's chat template instead. We connected both and sent a reply through Text Completion; we did not exercise generation through Chat Completion.

What URL do I enter in SillyTavern?

http://localhost:5001 (or http://127.0.0.1:5001) for Text Completion, and http://localhost:5001/v1 for Chat Completion. In our test SillyTavern 1.19.0 also accepted /v1 in Text Completion, but the plain address is the documented one.

Why does SillyTavern say no connection?

Usually KoboldCpp is not running, the port differs from 5001, or the address uses https:// against a normal server. In our test a wrong port and https:// against an http server both gave 'No connection...' and turned the plug icon red. If SillyTavern runs on another device, localhost does not mean the PC.

How do I use SillyTavern on my phone with KoboldCpp on my PC?

Enter the PC's address on your network instead of localhost, for example http://192.168.1.20:5001, and allow KoboldCpp through the firewall. Test the address in the phone's browser first. We have not tested this on a phone yet.

Which context size should I set in SillyTavern?

No larger than KoboldCpp's, which is 16384 by default plus 256 reserved, and leave response length well below it. A number above the server's gets the oldest text trimmed. We have not tested how SillyTavern reads the server's value.

Can SillyTavern use KoboldCpp for images and voice?

KoboldCpp serves image and speech routes that apps can call, but we have not tested SillyTavern's image or voice features against it.