Coding agents

Coding agents send a very particular workload: streamed tool calls, tool results fed back as tool messages, files written through tools, and a large context on every turn. OSMEXO speaks the OpenAI wire format, so every agent that takes an OpenAI-compatible base URL works as is. Pick a model that carries the Coding ready badge in the dashboard (filter the list by Coding ready), then set the tool up as below.

Set-up per tool

All of them need the same three things: the base URL __GATEWAY_URL__/v1, your OSMEXO key (sk-llm-…, from API Keys) and a model id such as anthropic/claude-sonnet-5. Key names below are the tools' as of late 2026; check the tool's own docs if a field has moved.

OpenCode

opencode.json in the project (or ~/.config/opencode/opencode.json):

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "osmexo": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "OSMEXO",
      "options": { "baseURL": "https://api.osmexo.com/v1", "apiKey": "{env:OSMEXO_API_KEY}" },
      "models": { "anthropic/claude-sonnet-5": { "name": "Claude Sonnet 5 via OSMEXO" } }
    }
  },
  "model": "osmexo/anthropic/claude-sonnet-5"
}

Cline and Roo Code

In the extension settings choose the OpenAI Compatible provider, then:

Leave "supports streaming" on. For reasoning models (Kimi K3, DeepSeek V4, Gemini Pro) raise the max output tokens if the tool exposes it; see the note on max_tokens below.

Continue

~/.continue/config.yaml:

models:
  - name: Claude Sonnet 5 via OSMEXO
    provider: openai
    model: anthropic/claude-sonnet-5
    apiBase: https://api.osmexo.com/v1
    apiKey: ${{ secrets.OSMEXO_API_KEY }}
    roles: [chat, edit, apply]

For codebase indexing add an embedding model the same way (roles: [embed], e.g. venice/venice-embed-1).

Aider

Aider reaches OpenAI-compatible endpoints through the openai/ prefix:

export OPENAI_API_BASE=https://api.osmexo.com/v1
export OPENAI_API_KEY=sk-llm-...
aider --model openai/anthropic/claude-sonnet-5

Codex CLI

~/.codex/config.toml, using the chat wire format (OSMEXO's Responses API is a subset; see Responses):

model_provider = "osmexo"
model = "anthropic/claude-sonnet-5"

[model_providers.osmexo]
name = "OSMEXO"
base_url = "https://api.osmexo.com/v1"
env_key = "OSMEXO_API_KEY"
wire_api = "chat"

What "Coding ready" means

Before a model gets the badge, OSMEXO's operators send the agents' workload through the route and judge each step deterministically. The badge is shown only when every required test passed; the model page lists the tests with what happened.

The check judges behaviour on the route, not identity; Verify a Model is the identity check, and you can run that one yourself.

Things to know

• Caching. Agents resend a large context every turn. Some routes report cached prompt tokens and some do not (the "Cached tokens reported" line on the model page says which). Today cached tokens are billed at the model's input price on every route, so for very long sessions compare the per-token price with the vendor's cached rate before you commit.
• max_tokens on reasoning models. Kimi K3, DeepSeek V4 and the Gemini Pro models spend part of the output budget thinking before the visible answer or tool call. Give them at least 1,200 output tokens for a tool call and a few thousand for a file; a small budget returns an empty answer. Most agents set this high by default.
• Streaming. Keep it on. Non-streaming calls must return headers within 60 seconds, which a long file generation can exceed; streamed calls only need to keep sending.
• One key per tool is enough: budgets, model allow-lists and spend per key are on API Keys, so a runaway agent can only spend what you gave that key.