buidlguidl/ethevals-test

ETH Evals

ETH Evals measures how well AI agents do Ethereum work and what bare models know about Ethereum. Tasks cover knowledge questions, building apps and contracts, and acting on a chain.

Each task contains an instruction, an environment, checks, and a working solution. You can run these folders with Harbor, without installing ETH Evals.

What you need

Install uv and Docker with Compose. Start Docker, then install Harbor:

uv tool install --python 3.13 harbor==0.24.0

For a task with a chain, give Docker at least 7 GiB of memory.

Try a task without credentials

The agent registries task asks which ERC defines registries for AI agents. It needs no RPC URL or judge key.

Run its bundled solution:

harbor run --repo https://huggingface.co/datasets/buidlguidl/ethevals-test \
  --include-task-name concepts--agent-registries \
  --agent oracle \
  --n-concurrent 1

oracle applies the solution and runs the checks. It checks that the task works; it does not measure a real agent. A real agent needs a model API key or plan authentication.

Tasks that need a chain or an LLM judge

Some tasks need an RPC URL to fork a chain. Tasks with an LLM judge also need an API key for judging.

For example, the treasury ETH sale task asks an agent to write and run a quarterly sale script. It uses both a mainnet fork at block 26,260,992 and an LLM judge.

This example uses OpenCode through OpenRouter. Set these variables in your shell:

export MAINNET_RPC_URL='your-mainnet-rpc-url'
export ANTHROPIC_API_KEY='your-judge-api-key'
export OPENROUTER_API_KEY='your-agent-api-key'

The RPC provider must serve historical state at the pinned block. The Anthropic key pays for the judge, which checks the agent's work. The OpenRouter key pays for the agent's model calls.

Run the task:

harbor run --repo https://huggingface.co/datasets/buidlguidl/ethevals-test \
  --include-task-name act--treasury-eth-sale \
  --agent opencode \
  --model openrouter/moonshotai/kimi-k3.1-pro \
  --n-concurrent 1

This uses your credentials and can incur API charges. Harbor saves the results under jobs/. A trial passes when its reward is 1.

Choose your agent

Each example below runs the agent registries task, which needs no RPC URL or judge key. Choose one agent and set up its credentials.

A ChatGPT or Claude plan can pay for the agent. For the treasury task, keep the RPC URL and judge key from the earlier example. A plan cannot pay for the judge, which still needs ANTHROPIC_API_KEY.

Codex with a ChatGPT plan or OpenAI API key

With a ChatGPT plan

Install Codex, then sign in and run the task:

codex login
export CODEX_FORCE_AUTH_JSON=YES

harbor run --repo https://huggingface.co/datasets/buidlguidl/ethevals-test \
  --include-task-name concepts--agent-registries \
  --agent codex \
  --model openai/gpt-6-astra \
  --n-concurrent 1

Harbor reads ~/.codex/auth.json and copies it into the task's container. If Codex uses your OS credential store, sign in with codex -c 'cli_auth_credentials_store="file"' login first. If your login file lives elsewhere, run export CODEX_AUTH_JSON_PATH='/path/to/auth.json'. See Codex authentication for login and storage options.

With an OpenAI API key

Clear the plan settings, set your key, and run the task:

unset CODEX_FORCE_AUTH_JSON CODEX_AUTH_JSON_PATH
export OPENAI_API_KEY='your-agent-api-key'

harbor run --repo https://huggingface.co/datasets/buidlguidl/ethevals-test \
  --include-task-name concepts--agent-registries \
  --agent codex \
  --model openai/gpt-6-astra \
  --n-concurrent 1

Claude Code with a Claude plan or Anthropic API key

With a Claude plan

Install Claude Code, then create a token:

claude setup-token

Copy the returned token into CLAUDE_CODE_OAUTH_TOKEN, then run the task:

export CLAUDE_CODE_OAUTH_TOKEN='your-plan-token'
export CLAUDE_FORCE_OAUTH=YES

harbor run --repo https://huggingface.co/datasets/buidlguidl/ethevals-test \
  --include-task-name concepts--agent-registries \
  --agent claude-code \
  --model anthropic/claude-opus-5-5 \
  --n-concurrent 1

CLAUDE_FORCE_OAUTH tells Harbor to use the plan even when the judge's Anthropic key is present.

With an Anthropic API key

Clear the plan settings, set your key, and run the task:

unset CLAUDE_CODE_OAUTH_TOKEN CLAUDE_FORCE_OAUTH
export ANTHROPIC_API_KEY='your-api-key'

harbor run --repo https://huggingface.co/datasets/buidlguidl/ethevals-test \
  --include-task-name concepts--agent-registries \
  --agent claude-code \
  --model anthropic/claude-opus-5-5 \
  --n-concurrent 1

For a task with an LLM judge, the Anthropic key pays for both the agent and the judge.

OpenCode with OpenRouter

Set your OpenRouter key and run the task:

export OPENROUTER_API_KEY='your-agent-api-key'

harbor run --repo https://huggingface.co/datasets/buidlguidl/ethevals-test \
  --include-task-name concepts--agent-registries \
  --agent opencode \
  --model openrouter/moonshotai/kimi-k3.1-pro \
  --n-concurrent 1

The treasury example already uses this setup. Its Anthropic key pays for the judge separately.

Before running another task

Browse the task folders. Use the folder name with --include-task-name, such as concepts--agent-registries.

Read the task's README for its instruction and checks. Then check its configuration:

• environment/docker-compose.yaml names any required RPC variables and the fork block.
• [verifier.env] in task.toml names the judge key, if the task needs one.

For example, the treasury task declares its RPC requirement and judge key separately. Agent credentials are additional to these task requirements.

RPC variables from the exported folders

These names come from each task's Compose file. Compose stops with the message if a value is missing or empty.

• buidlguidl/treasury-eth-sale: MAINNET_RPC_URL
• buidlguidl/vault-v2-concentration-rebalance: MAINNET_RPC_URL
• buidlguidl/vault-v2-market-exit: MAINNET_RPC_URL
• buidlguidl/aave-savings-page: MAINNET_RPC_URL
• buidlguidl/keeper-deposits: MAINNET_RPC_URL
• buidlguidl/signed-votes: MAINNET_RPC_URL
• buidlguidl/swap-page: MAINNET_RPC_URL

See the Harbor guide for more run options. To write a task, read Add a task.

Every task includes its solution, and those solutions can appear in training data. Scores alone do not establish unseen Ethereum knowledge.