How to use the Aider Delegate agent skill (aider-delegate)

What the aider-delegate skill makes the agent do

The skill turns the agent into an orchestrator. The agent writes a task brief, sends it to Aider as a separate implementer, and Aider edits files in your git repository. The agent then reads the result and decides whether to commit it.

The skill only activates when you explicitly ask for delegation to Aider. Aider's own automatic commits are switched off by design, so its edits arrive as an uncommitted diff for you or the agent to check.

• The agent writes a brief covering the goal, current state, what to change, what to leave alone, and the project's real test commands.
• Aider makes the edits inside its own run.
• The agent verifies the diff and keeps commit authority.

Install the aider-delegate skill

Download the ZIP for this skill from this page and unzip it. Put the folder in ~/.claude/skills/ to make it available everywhere, or in a project's .claude/skills/ folder to limit it to that project.

You can also upload the skill in the Claude apps under Settings → Capabilities → Skills.

Before you rely on it, note that this bundle contains documentation only. The helper script scripts/relay.mjs is not included, so the documented workflow cannot run as written until a reviewed runtime is installed separately.

• Option 1: unzip into ~/.claude/skills/ for all projects.
• Option 2: unzip into a project's .claude/skills/ folder.
• Option 3: upload through Settings → Capabilities → Skills in the Claude apps.

Install and check Aider and its requirements

The skill needs the Aider CLI, Node 18 or newer, git, and a model with credentials. Aider is installed with pip, and you should confirm it runs before dispatching any task.

Aider reads provider keys from environment variables such as OPENAI_API_KEY or ANTHROPIC_API_KEY, or from its own config. Work inside the target git repository, or point the tool at it.

• Install Aider with the pip command shown here.
• Confirm the install with aider --version.
• Set a provider key or configure Aider's own model settings.
• Make sure the target folder is a git repository.

How to trigger the skill with a natural request

Ask for delegation explicitly, by naming Aider. The skill is written to act only when you request it, so a general request to fix something will not hand the work to Aider.

A good request says what the bounded task is and which repository it concerns. Keep one task per request so the diff stays easy to review.

• Delegate this change to Aider: add input validation to the signup form in this repo.
• Have Aider implement the rename of the config parser in /path/to/repo, then show me the diff before committing.
• Send this bounded refactor to Aider and review what it changes.

Use a local or self-hosted model with Aider

Aider works with any OpenAI-compatible endpoint, so the skill can drive a model on your own hardware. The source names llama.cpp's server, Ollama, vLLM, and LM Studio as examples. Pass the model with --model and the server address with --api-base.

Three details differ from a hosted provider. The model name must begin with the openai/ prefix. You still need to export a non-empty OPENAI_API_KEY even if the server ignores it. Local models often fail Aider's default diff edit format, so the source suggests whole with --edit-format whole and a tight file scope with --file.

A local endpoint that is not running can look like a hang rather than an error, so confirm the server is up before sending a long brief.

• Use the openai/ prefix before your served model name.
• Export any non-empty OPENAI_API_KEY for local endpoints.
• Try --edit-format whole if the default edit format fails.
• Limit scope with --file to keep rewrites small.

Safety notes on scripts and network use

This skill runs scripts and uses the network. Delegation sends the brief, the selected code, and repository context to the configured model provider. Confirm the endpoint and its data policy before you send private or secret code, and never send a real provider key to an arbitrary api-base address.

The listed flags reduce some automatic network requests, but they do not prove the run is offline. Model requests, repository lint or hook commands, and dependency installs can still reach the network. If you need offline behaviour, enforce it outside the process. Logs and history can keep private code and keys, so protect and review retention.

• Model requests send your code and brief to the configured provider.
• Repository lint, test, and hook commands run during a task; review them first.
• A dry run can still write history or cache files and may incur model costs.
• Set a cost limit where your provider allows it.