Skip to content

BYOK Language Models

Early access

DEVI is in beta and free for every workspace. You bring your own provider key, and your provider bills the requests. Answers are grounded in supplied scan evidence and the shipped documentation and may still be incomplete. DEVI never executes or approves infrastructure changes.

Bring your own API key. CloudSlash routes DEVI through whatever model you configure: GPT-4, Claude, Gemini, or Ollama running locally.

Interactive Setup

The fastest way to configure your AI provider:

cs config

The wizard walks you through:

  1. Provider: Ollama, OpenAI, Anthropic, Gemini, Groq, or custom endpoint
  2. API key: provider-specific input with placeholder hints
  3. Model: selected from the live list fetched from your provider

If your provider is unreachable (wrong key, network issue), model selection is skipped. Fix the key and re-run cs config.

Cloud Providers

Gemini

cs config set ai.provider gemini
cs config set ai.gemini_key "AIza..."

Or via environment:

export CLOUDSLASH_AI_MODEL=gemini-2.5-pro
export CLOUDSLASH_AI_KEY="AIza..."

OpenAI

cs config set ai.provider openai
cs config set ai.api_key "sk-..."

Anthropic

Claude

cs config set ai.provider anthropic
cs config set ai.anthropic_key "sk-ant-api..."

Groq

cs config set ai.provider groq
cs config set ai.api_key "gsk_..."
cs config set ai.base_url "https://api.groq.com/openai/v1"

Local Inference

Ollama

Run inference completely air-gapped with no external data egress.

  1. Install and start Ollama:
ollama serve
  1. Pull a model:
ollama pull llama3.2
  1. Configure CloudSlash:
cs config set ai.provider ollama

CloudSlash auto-detects Ollama at http://localhost:11434. To use a custom endpoint:

cs config set ai.ollama_endpoint "http://192.168.1.50:11434"

Manual Configuration

All settings live in ~/.cloudslash/config.yaml:

ai:
  provider: gemini
  model: gemini-2.5-pro
  gemini_key: "AIza..."

  # For Ollama:
  # provider: ollama
  # model: llama3.2
  # ollama_endpoint: "http://localhost:11434"

  # For custom OpenAI-compatible endpoints:
  # provider: other
  # model: my-model
  # api_key: "..."
  # base_url: "https://my-endpoint.com/v1"

Verification

Check that your AI backend is reachable:

cs doctor

The doctor command validates the configured provider, checks API key presence, and tests endpoint connectivity.

Sessions and sign-out

The browser workspace ends an idle session after 48 hours; active use refreshes the clock. When Horizon rejects an expired CLI session, cs sync signs the local CLI out instead of retrying with a dead token — run cs login to reconnect. The TUI is local-first: it does not need a Horizon session to scan, and it never displays or stores provider keys.

Backup keys and failover

The main model is the primary. Below it, Backup models lets you add any number of named keys, including multiple keys for the same provider. Each block has a provider, key name, model and API key.

If the primary provider fails, DEVI tries the backups in order until one answers. The reply shows which model was used, fallback: <key name>, and an estimated cost for the reply. Estimates use list prices and are not a bill; your provider charges the actual amount. A reaction (Love, Thanks, Noted) stores the question and answer only when you choose it.