Skip to content
tomo-labs

Tools

One research page per wired agent: what it is, its command surface, how the lab drives it, how it works inside, the system prompt it actually sent, and a Hi! run traced end to end.

Eleven agents run through the harness today. Each one gets a page here that treats it as the object of study, not just a row in a table.

A tool page covers the same six things every time so two agents are easy to compare:

  • What it is, and who builds it.
  • Its command surface, down to the subcommands and flags the lab relies on.
  • How the lab drives it, which is one adapter script and one Dockerfile.
  • Its architecture, as far as the trace and the tool's own code show it.
  • The system prompt it actually sent, recovered from the trace and checked against the tool's public source.
  • A Hi! run traced end to end, from the request the tool built to the reply that reached the user.

The system prompt sections quote the real text the proxy captured. The full verbatim prompts live under prompts, one page per tool, regenerated with lab prompts <tool> so any drift shows up in the diff.

tomo tomo is the harness author's own personal AI agent, one Go binary from github.com/tamnd/tomo, driven headless with tomo -p and speaking native chat-completions. codex OpenAI's Codex CLI, installed from npm and driven headless through codex exec over the Responses wire, which the proxy shims to chat-completions for the shared free model. opencode The sst/opencode terminal coding agent, driven headless through opencode run against the lab's trace proxy. claude-code Anthropic's Claude Code CLI, run headless through the trace proxy as if it were the Anthropic API, carrying the largest baseline prompt of any wired tool. openclaw A self-hosted personal-assistant CLI installed from npm, wired into the lab as its own tool and measured through its native OpenAI chat-completions provider. hermes hermes is Nous Research's Hermes Agent, driven headless with hermes chat -q and speaking native chat-completions to the trace proxy. gemini-cli gemini-cli is Google's Gemini CLI, a Node command driven headless with gemini -p and speaking the Gemini generateContent wire that the proxy shims to chat-completions. pi pi is the earendil-works coding agent CLI, driven headless with pi -p and pointed at the trace proxy through a custom lab provider. kilocode Kilo Code, an opencode fork, driven headless through kilo run against the lab's trace proxy on the same fixed model as every other agent. aider aider, the whole-file-editing pair programmer, driven headless through aider --message against the lab's trace proxy on the same fixed model as every other agent. copilot The GitHub Copilot CLI, driven headless through copilot -p in BYOK mode against the lab's trace proxy on the same fixed model as every other agent.