DeepSeek not working using `Anthropic` mode

[SkYNewZ]

Hello! It's me again.

EDIT : 0346dd3 should fix the original issue using regarding OpenAI OpenAI Chat Completions provider. I am still on v0.5.2

I am currently configuring the app again (npx fix okay!).
I'm having trouble using DeepSeek. The settings are simple : https://api-docs.deepseek.com/
Also, their "Anthropic" endpoint does not work neither because of the /v1. Testing connection (get models) simply returns 404. Not tester if messages work although.

PARAM	VALUE
base_url (OpenAI)	https://api.deepseek.com
base_url (Anthropic)	https://api.deepseek.com/anthropic
api_key	apply for an API key
model*	deepseek-v4-flashdeepseek-v4-prodeepseek-chat (to be deprecated on 2026/07/24)deepseek-reasoner (to be deprecated on 2026/07/24)

On Agent tab, All good, models recognized, all green!

On Chat, I got a 404 error message.

Note

I wonder if a path has been added to the API’s base URL, which would result in the routes being concatenated incorrectly and therefore make it impossible to use.

[SkYNewZ]

Root cause — Anthropic provider builds the wrong path (/messages instead of /v1/messages)

I traced the 404 by following how the URL is assembled:

• AnthropicMessagesProvider builds its endpoint as baseURL.appending(path: "messages") (same for models / embeddings).
• The default Anthropic base URL is hard-coded as https://api.anthropic.com/v1 (LLMAPIKind.defaultBaseURL), so the /v1 segment is expected to live inside the user's base URL.

DeepSeek (like every Anthropic-compatible gateway) documents its base URL as https://api.deepseek.com/anthropic without /v1, because the official Anthropic SDKs append /v1/messages themselves. So when a user enters the documented value, the app sends:

POST https://api.deepseek.com/anthropic/messages      → 404   (the /v1 is missing)

instead of the correct:

POST https://api.deepseek.com/anthropic/v1/messages   ✓

This is exactly the "a path gets concatenated incorrectly" suspicion in the issue.

Fix — align with the Anthropic SDK convention: base URL = host root, provider supplies /v1

Add the /v1 in the provider and tolerate a trailing /v1 already present in saved configs (so existing https://api.anthropic.com/v1 providers keep working). Resolves identically for all three forms:

Configured base URL	Resulting endpoint
https://api.anthropic.com	…/v1/messages
https://api.anthropic.com/v1 (legacy state)	…/v1/messages
https://api.deepseek.com/anthropic	…/anthropic/v1/messages

Patch (Sources/FyxLocalProviders/Anthropic/AnthropicMessagesProvider.swift):

/// Builds an Anthropic API endpoint URL from the configured base URL.
/// The Anthropic wire API lives under `/v1`. The official SDKs treat the base
/// URL as the host root and add `/v1` themselves, which is how compatible
/// gateways document it (e.g. DeepSeek: https://api.deepseek.com/anthropic).
/// We append `v1/<path>`, stripping a trailing `/v1` first so both conventions
/// resolve identically.
func endpoint(_ path: String) -> URL {
    var base = baseURL
    if base.lastPathComponent == "v1" {
        base = base.deletingLastPathComponent()
    }
    return base.appending(path: "v1/\(path)")
}

Then replace the three call sites:

• baseURL.appending(path: "models") → endpoint("models")
• baseURL.appending(path: "embeddings") → endpoint("embeddings")
• baseURL.appending(path: "messages") → endpoint("messages")

And update the default placeholder (Sources/FyxLocalCore/Models/ProviderRecord.swift, LLMAPIKind.defaultBaseURL): https://api.anthropic.com/v1 → https://api.anthropic.com.

Tests added in Tests/FyxLocalProvidersTests/AnthropicMessagesTests.swift cover host-root, legacy /v1, gateway sub-path, and trailing-slash cases (plus the updated defaultBaseURL assertion).

One caveat for DeepSeek specifically (separate from this fix): the connection check / model picker calls listModels() (GET …/v1/models). DeepSeek's Anthropic-compatible gateway only implements /v1/messages, not model listing, so the green check there is still misleading. The fully working path to use DeepSeek today is the OpenAI (Chat Completions) provider kind with base https://api.deepseek.com — note that OpenAI (Responses) will 404 because DeepSeek doesn't implement /responses.

Prepared by Claude (Claude Code) for @SkYNewZ.

[FyrbyAdditive]

Thanks for this report! I'm currently deep into enhancing the RAG and hope to look at this directly after!

[FyrbyAdditive]

@SkYNewZ I do not have a DeepSeek account, I wanted to check the models url (that lists the models available) exists? The DeepSeek documentation doesn't seem to mention it (or perhaps this is me being sleepy on a Friday and I missed it!)

[SkYNewZ]

Yep here it is!

curl -s https://api.deepseek.com/models \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \
| jq
{
  "object": "list",
  "data": [
    {
      "id": "deepseek-v4-flash",
      "object": "model",
      "owned_by": "deepseek"
    },
    {
      "id": "deepseek-v4-pro",
      "object": "model",
      "owned_by": "deepseek"
    }
  ]
}

[FyrbyAdditive]

Great thank you! I will have a look - an interim workaround for DeepSeek can be to use the OpenAI Completions API in 0.6.0, which seems to be supported according to docs.

[FyrbyAdditive]

After a look today you are correct, DeepSeek does this in a different way to the Anthropic API. DeepSeek uses the base URL of https://api.deepseek.com/anthropic however models live at https://api.deepseek.com/models, below the base directory (uhhh).

I suppose the nice fix is to remove the /v1 assumption and add a hack to fail back to https://api.deepseek.com/models when it inevitably fails because they didn't copy the API properly. I'll try and get an API key from somewhere as I much prefer to test things!

[FyrbyAdditive]

I have dropped DeepSeek API support an email about this this morning, explaining this and the case for fixing it as I believe this should be resolved their end to be more compatible with the Anthropic API, as this will benefit them. I will otherwise fix this anyway to be clear if they do not!