Chat apps & coding tools

Every application here has a setting for an OpenAI-compatible endpoint. Point it at the marketplace with a base URL and a Heimdallm API key (sk-hd-v1-…, see Authentication), and name models by their author/slug catalogue id, copied verbatim from the catalogue.

The base URL for all of them is

https://api-staging.heimdallm.com/api/v1

None of these applications has a row in the compatibility matrix, which drives client SDKs (see SDKs & frameworks). The settings follow each project's documented OpenAI-compatible configuration; the first request is the check.

Tool use and agents depend on the model. A coding tool sends tools with its requests, and the router sends a request that uses tools only to endpoints that declare them. Choose a model whose catalogue entry lists tools in supported_parameters (Choosing a model), and see Tool calling for how a call is returned.

Open WebUI#

In Settings → Admin → Connections, add a connection under Manage OpenAI API Connections:

• URL: the base URL above, without a trailing slash.
• API Key: the Heimdallm key.
• Model IDs: optional, needed only for a server that does not list its models. The marketplace answers GET /models.

Verify Connection calls GET /models and saving does not, so verify before the first chat. The environment variables OPENAI_API_BASE_URL and OPENAI_API_KEY seed the same connection on the first start; afterwards the admin panel holds the value.

LibreChat#

LibreChat takes a custom endpoint in librechat.yaml. Add this entry to the file's endpoints.custom list, replacing <the base URL above> with the base URL at the top of this page. The key is read from the environment by reference, so the file holds no secret:

endpoints:
  custom:
    - name: 'Heimdallm'
      apiKey: '${HEIMDALL_API_KEY}'
      baseURL: '<the base URL above>'
      models:
        default: ['deepseek/deepseek-v4-flash-0731-abliterated']
        fetch: true
      titleConvo: true
      modelDisplayLabel: 'Heimdallm'

Add HEIMDALL_API_KEY=… to LibreChat's .env, and restart after changing either file. An endpoint missing name, baseURL, apiKey or models is dropped without an error, and models needs either fetch: true or at least one default entry.

Cline#

Open Cline's settings and choose the OpenAI Compatible provider:

• Base URL: the base URL above. It must not be OpenAI's own.
• API Key: the Heimdallm key.
• Model ID: a catalogue id.

Under Model Configuration, set Context Window to the model's context_length and Max Output Tokens to a ceiling the model allows, and enable Image Support only for a model whose input_modalities list image. The price fields are the extension's own estimate; the marketplace's charge for a request is the usage.cost in its reply.

Roo Code#

Roo Code's settings take the same three values under the OpenAI Compatible provider: Base URL, API Key and Model, with the same optional Model Configuration (context window, maximum output tokens, image support). Roo Code expects native OpenAI-compatible tool calling, so choose a model that declares tools.

Continue#

Continue's config.yaml declares a model with the openai provider and an apiBase:

name: Heimdallm
version: 0.0.1
schema: v1

models:
- name: Heimdallm chat
  provider: openai
  model: deepseek/deepseek-v4-flash-0731-abliterated
  apiBase: https://api-staging.heimdallm.com/api/v1
  apiKey: <your Heimdallm API key>

model is the catalogue id and provider: openai selects the OpenAI wire, which the marketplace serves under that base URL.

aider#

aider reaches an OpenAI-compatible endpoint through two environment variables and an openai/ model prefix:

export OPENAI_API_BASE=https://api-staging.heimdallm.com/api/v1
export OPENAI_API_KEY=$HEIMDALL_API_KEY

aider --model openai/deepseek/deepseek-v4-flash-0731-abliterated

The openai/ prefix tells aider which wire to speak; the rest is the catalogue id. aider may warn that it does not recognize the model, because the id is a marketplace catalogue id; its documentation on model warnings covers declaring the context window.

Coding harnesses#

Claude Code, Codex and the other terminal agents are configured for the dialect they send rather than for an OpenAI endpoint. Switching to Heimdallm lists the base URL and key variable each uses, and the playground runs seven harnesses on a machine that is already connected.