Using OrcaRouter with Helix via lsp-ai and aichat

[zhenjunchen-png]

Sharing a setup in case it is useful to others: how to use OrcaRouter with Helix.

OrcaRouter is an OpenAI-compatible LLM gateway: one
API key and endpoint give you access to models from many providers, plus an
adaptive router (orcarouter/auto) that picks the best upstream per request.

Helix has no built-in LLM support by design, and the recommended paths are LLM
language servers and CLI tooling. OrcaRouter plugs into both: an AI language
server (lsp-ai) for in-editor completion, and a CLI in a pipe (aichat) for
transforming selections. Both work with zero changes to Helix itself.

Disclosure: I'm an engineer on the OrcaRouter team.

Get an API key

1. Sign up at https://www.orcarouter.ai

2. Create a key in the console: https://www.orcarouter.ai/console (keys start with sk-orca-)

3. Export it so the tools below can read it:

   export ORCAROUTER_API_KEY=sk-orca-xxxxxxxxxxxx

Full model catalog: https://www.orcarouter.ai/models

Option A: In-editor completion via lsp-ai

lsp-ai is an OpenAI-compatible AI
language server. Point it at OrcaRouter and Helix gets AI completions through the
normal LSP completion menu. Add this to ~/.config/helix/languages.toml (keep
your existing language servers and add lsp-ai alongside them):

[language-server.lsp-ai]
command = "lsp-ai"

[language-server.lsp-ai.config.memory]
file_store = { }

# Pin a NON-reasoning model: lsp-ai always sends `temperature`, which
# reasoning models (and so orcarouter/auto, which may route to one) reject.
[language-server.lsp-ai.config.models.orcarouter]
type = "open_ai"
chat_endpoint = "https://api.orcarouter.ai/v1/chat/completions"
model = "openai/gpt-4.1-mini"
auth_token_env_var_name = "ORCAROUTER_API_KEY"

[language-server.lsp-ai.config.completion]
model = "orcarouter"

[language-server.lsp-ai.config.completion.parameters]
max_tokens = 128
max_context = 2048

[[language-server.lsp-ai.config.completion.parameters.messages]]
role = "system"
content = "Instructions: You are an AI programming assistant. Given a piece of code with the cursor location marked by \"<CURSOR>\", replace \"<CURSOR>\" with the correct code or comment. Only respond with code or comments. Only replace \"<CURSOR>\"; do not repeat existing code. Never include \"<CURSOR>\" in your response. Never wrap your response in markdown code fences."

[[language-server.lsp-ai.config.completion.parameters.messages]]
role = "user"
content = "def greet(name):\n    print(f\"Hello, {<CURSOR>}\")"

[[language-server.lsp-ai.config.completion.parameters.messages]]
role = "assistant"
content = "name"

[[language-server.lsp-ai.config.completion.parameters.messages]]
role = "user"
content = "{CODE}"

[[language]]
name = "python"
language-servers = ["pyright", "lsp-ai"]

Restart Helix and edit a Python file. Completions are served by OrcaRouter.

Notes:

• Completion runs through the chat endpoint with the <CURSOR> technique, so
  no /v1/completions FIM endpoint is needed.
• Pin a non-reasoning model. lsp-ai always sends temperature. Reasoning
  models (anthropic/claude-opus-4.8, the gpt-5 family, deepseek-reasoner,
  ...) reject temperature with a 400, and orcarouter/auto may route to one.
  A fast non-reasoning model such as openai/gpt-4.1-mini accepts temperature
  and suits low-latency inline completion.

Option B: Transform selections via aichat (Unix pipe)

aichat is an LLM CLI. With Helix's |
command (pipe the selection through a shell command and replace it with the
output) you get inline rewrite/explain/fix without any language server.

Save this as ~/.config/aichat/config.yaml. aichat reads the API key from the
environment variable named after the client (ORCAROUTER_API_KEY), so do not put
the key in this file:

model: orcarouter:orcarouter/auto

clients:
  - type: openai-compatible
    name: orcarouter
    api_base: https://api.orcarouter.ai/v1
    models:
      - name: orcarouter/auto
        max_input_tokens: 128000
        supports_function_calling: true

Add a role that returns only code (aichat 0.30+ loads roles from files). Save
this as ~/.config/aichat/roles/code.md:

You are a coding assistant embedded in a text editor. The user pipes a
selection of source code to you. Apply the requested change and reply with
ONLY the resulting code: no explanations, no commentary, no markdown fences.

In Helix: select code, press |, and run:

aichat -m orcarouter:orcarouter/auto -r code

The selection is sent to OrcaRouter and replaced with the result. You can also
keep an interactive aichat session open in a split terminal.

Verify your key and endpoint

curl https://api.orcarouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $ORCAROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"orcarouter/auto","messages":[{"role":"user","content":"Reply with the single word: pong"}]}'

Limitations

• These third-party tools talk plain OpenAI Chat Completions. OrcaRouter
  extensions such as extra_body routing preferences and attribution headers
  are not exposed by lsp-ai or aichat today; orcarouter/auto routing still
  works because it is selected by model name.
• A native, branded integration would require Helix's Steel plugin system, which
  is still under development and not part of a stable release.