AI coding agents

DirectInference is a drop-in endpoint, so coding assistants and CLIs work the same way your app does — point them at the base URL and key, keep your model ids.

Running a standalone agent platform?

This page covers IDE and CLI coding assistants. For self-hosted agent gateways and autonomous agent runtimes — OpenClaw and the like — see Agent frameworks, which has worked, per-platform setups.

Point your agent at DirectInference

Cursor, Claude Code, and most agent frameworks let you override the API base URL and key — usually through environment variables or a settings field (full provider forms are covered in Custom providers). Set them to your DI surface and the agent sends requests through DirectInference unchanged.

curl

# Most OpenAI-compatible agents and CLIs read these:
export OPENAI_BASE_URL="https://api.directinference.com/di/v1"
export OPENAI_API_KEY="llm_live_..."

# Anthropic-compatible agents (e.g. Claude Code):
export ANTHROPIC_BASE_URL="https://api.directinference.com/di"
export ANTHROPIC_AUTH_TOKEN="llm_live_..."

Keep the model id the agent already uses

Whatever model the tool defaults to is treated as intent and echoed back. You do not need to configure a DirectInference model name — there is only one model, and request handling is automatic. If the agent has fast/smart model slots (a background model and a main model), use di-saver and di-max — the same model with effort pinned low and high.

Using a tool with a structured provider config?

OpenCode, Cline, Roo Code, Continue, Goose, Kilo Code, Zed, Aider, and Crush don’t take a base-URL env var — they need a provider block with the model’s capabilities (tools, reasoning, vision) and context window declared by hand, because di-fusion is new to their model catalog. Copy-pasteable setup for each is in Coding agents & orchestrators.

Machine-readable docs

For assistants that read documentation directly, DirectInference publishes condensed, plain-text indexes of the API at the root of the public site. Point a coding agent at them to teach it the endpoint, request types, and effort contract in one fetch.

llms.txt

https://directinference.com/llms.txt — a compact overview.

https://directinference.com/llms-full.txt — the expanded reference.

A migration an agent can do for you

Because the switch is purely a base-URL-and-key change with the model ids left intact, an agent can convert an existing OpenAI, Anthropic, or Gemini integration mechanically — find the client construction, change two values, done. The full manual walkthrough is in Migrate to DirectInference.

curl

# Install the packaged skill once (shown: Claude Code's personal skills directory)
git clone https://github.com/Direct-Inference/skills.git
mkdir -p ~/.claude/skills
cp -R skills/skills/use-directinference ~/.claude/skills/

# Then, in any project, just ask: "switch this app to DirectInference"

The packaged skill

https://github.com/Direct-Inference/skills — this migration as an installable Agent Skill: it inventories your call sites, applies the base-URL-and-key swap, and verifies the result with a bundled script. Agents without a skills directory can follow the raw SKILL.md in the repo directly.