Harnesses: how log works in each one

A harness is the agent app someone talks to log through: Claude, ChatGPT, Cursor, Codex, Hermes, OpenCode, Antigravity or Grok. log is the same MCP server in all of them. What changes from one to the next is four things:

  1. Whether it draws MCP Apps. That is the MCP extension io.modelcontextprotocol/ui. Two of log's tools draw through it: report (the generated view) and browserView (the live browser for signing an account in). A host that draws MCP Apps shows them inline in the chat; any other host gets the same answer as text.
  2. How log's guidance reaches the model. log has three ways in, used in this order:
    • The prompt router. It matches the owner's request against every use case's "How the owner asks for it" phrases and returns the use case to follow (use cases). Where log has a prompt hook (Claude Code, Codex, OpenCode and Antigravity in the repo, and Hermes through the native logHooks plugin's pre_llm_call), it runs on every prompt by itself; when the hook's context service is down, the hook falls back to the same word router, so the play is still named. Cursor's prompt hook can't add context, so Cursor gets a sessionStart and a postToolUse hook and routes with ask. Everywhere else the model runs it: getGuide({ ask: "<the owner's words>" }) calls the same route on the server, local or hosted. On the word score alone both give the same answer for the same prompt, and a golden set of prompts checks it (src/useCases/router/fixtures/golden.json). The hook's service adds a meaning score from a local embedding model, which the ask path does without. On a clear match ask loads that play for the caller; on a tie or a near tie it returns a short menu and loads nothing.
    • The server's instructions. Sent at connect, they tell the model to load getGuide({ topic: "logMcpInstructions" }) first.
    • The tool descriptions. Each one names its guide and the use cases that call it.
  3. How it connects. Either to the hosted server, https://mcp.listeningkit.com/mcp over OAuth, or to logMcp on the same machine over stdio.
  4. Where an account is signed in. In the chat through browserView, in the window on the owner's screen when log runs on their machine, or in the log console. All three drive the same live browser, src/browser/viewer.
HarnessPrompt hookSession-start hookTool hooksDraws MCP AppsServer instructionsSkillsURL elicitationPrompt routerHook text limitConnects toHooks from
Claude Codeyesyesyesnoyesyesyesthe prompt hook10,000 charactershosted or local.claude/settings.json
Codexyesyesyesunknownyesunknownunknownthe prompt hook7,500 charactershosted or local.codex/hooks.json
CursornoyesyesyesunknownunknownunknowngetGuide({ ask })7,500 charactershosted or local.cursor/hooks.json
Hermesyesnoyesunknownunknownyesnothe pre_llm_call hook (native plugin)7,500 charactershosted or localhermesPlugin/plugin.yaml
OpenCodeyesyesyesunknownunknownunknownunknownthe prompt hook20,000 characterslocal.opencode/plugins/logHooks/index.ts
Antigravityyesyesyesunknownunknownunknownunknownthe prompt hook7,500 characterslocal.agents/hooks.json
Claude (web, desktop, mobile)nononoyesunknownunknownunknowngetGuide({ ask })no hookshostednone
ChatGPTnononoyesyesunknownunknowngetGuide({ ask })no hookshostednone
Grok (xAI API)nononounknownunknownunknownunknowngetGuide({ ask })no hookshostednone

The table is HARNESSES in src/agentHooks/harness/types.ts; every hook file and this table are written from it and from AGENT_HOOKS. "unknown" means no source in sources.md states it. Each hook's text is cut to the limit shown. "Draws MCP Apps" comes from the MCP extension support matrix, read on 2026-11-13. It is community-maintained, so test before relying on it.

What this means for log

  • Hosts that draw MCP Apps:
    • Views: report shows the view inline, with pages per person and buttons that ask the next question.
    • Sign-in: browserView shows the account's live browser inline, whether log runs on the owner's machine (a loopback relay, or frame pulls) or remotely (hosted or a paired workspace). Locally, its window is on the owner's screen too.
  • Hermes Desktop: doesn't draw MCP Apps, but log's Desktop plugin draws the same two inline: the agent writes the ::logview line report names and the ::logbrowser{account="..."} line browserView names, each on its own line.
  • Hosts that draw neither:
    • Views: they get the view as text.
    • Sign-in: happens in the window or the log console.
  • Where there is no prompt hook, the model routes each request with getGuide({ ask }). Both servers' instructions say to, and so does getGuide's own description, so it works whether or not a host passes the instructions on.