AI-assisted coding with evroc Think

evroc think code launches popular AI coding agents — Claude Code, Codex, OpenCode, Aider, and more — with evroc Think as the model backend. The CLI handles API keys, proxy translation, and environment configuration automatically.

You get AI-assisted coding on European infrastructure without manually configuring each tool's API endpoint, base URL, or auth headers.

Supported agents

Note: Claude Code and Codex use protocol-specific APIs that Think doesn't expose directly. The CLI runs a local translation proxy for these agents — you don't need to configure anything extra.

Prerequisites

Before you begin, you'll need:

• The evroc CLI installed and logged in, or a Think API key
• At least one agent installed (e.g., npm install -g @anthropic-ai/claude-code)

Authentication

The CLI resolves your Think API key in the following order:

1. --api-key flag
2. EVROC_API_KEY environment variable
3. EVROC_THINK_API_KEY environment variable
4. Auto-created from your evroc CLI session (OIDC)

For CI and scripting, set EVROC_THINK_API_KEY to skip the OIDC flow:

export EVROC_THINK_API_KEY="<your-api-key>"
evroc think code claude -- -p "Explain this function"

Create an API key with:

evroc think apikey create my-ci-key

Launch an agent

evroc think code claude

This starts Claude Code with Think configured as the backend. The CLI sets the appropriate environment variables for the agent (e.g., ANTHROPIC_API_KEY and ANTHROPIC_BASE_URL for Claude Code, OPENAI_API_KEY and OPENAI_BASE_URL for most other agents), starts a translation proxy if needed, and spawns the agent.

Choose a model

By default, the CLI uses moonshotai/Kimi-K2.5 from the shared models catalogue. Specify a different model with --model:

evroc think code claude --model zai-org/GLM-5.2
evroc think code aider --model moonshotai/Kimi-K2.5

See Supported Models for the full catalogue.

Pass flags to the agent

Arguments after -- are passed through to the underlying agent:

evroc think code claude -- -p "Say hello world, nothing else"
evroc think code codex -- exec --skip-git-repo-check "Refactor this function"
evroc think code aider -- --yes --message "Add a README"

Permanent configuration with evroc think code init

evroc think code runs an agent for a single session — no config files are written. Use evroc think code init to write permanent config files that connect an agent to Think:

evroc think code init claude
evroc think code init codex
evroc think code init aider

This writes agent-specific config files in your project directory or home directory:

The config files use your Think API key and endpoint. You can commit them to share with your team.

Important: The init command writes config files, but doesn't run a proxy. For Claude Code and Codex, use evroc think code claude / evroc think code codex to start the translation proxy. The init config is for reference and for agents that talk directly to OpenAI-compatible endpoints.

How it works

evroc think code claude
        │
        ├─ 1. Resolve API key (flag → env → OIDC session)
        ├─ 2. Start local Anthropic→OpenAI proxy (Claude only)
        ├─ 3. Set ANTHROPIC_API_KEY, ANTHROPIC_BASE_URL env vars
        └─ 4. Spawn: claude-code
                │
                └─ Agent talks to local proxy → Think Models API

For agents that are already OpenAI-compatible (Aider, OpenCode, Hermes, etc.), the CLI sets the environment variables and spawns the agent directly — no proxy needed.

The local proxy is torn down when the agent exits. No ports are left listening.

Environment variables

Next steps

• Supported models — Browse available models
• OpenAI compatibility — Use Think with any OpenAI-compatible client
• CLI reference — Full evroc think code command reference