hermitd

Reviewed at commit 91ff2f6 New version, not yet reviewed

Caution: Adds permission rules, including a broad `bash -c` rule, to your project's Claude settings, re-applies them on each `hermitd update`, and downloads and runs prebuilt browser and transcription binaries.

Sends data to Discord (discord.com), Telegram (api.telegram.org), Groq (api.groq.com) or OpenAI (api.openai.com) for voice notes only when you set a key, your telemetry webhook and backup remote (off by default)

Your own local Claude Tag, in the Claude app, Discord, Telegram, or a custom Claude Code channel.

Made by
gtapps
What we checked

Turns a project folder into an always-on Claude Code agent in tmux or Docker that you message from the Claude app, Discord, Telegram or a custom channel. It keeps state in .hermit, edits project Claude settings, saves learned skills and nightly memory notes, transcribes voice notes, and drives its own downloaded Chrome; voice notes reach Groq or OpenAI only when local transcription fails and a key is set.

Review details

Static review of the source at the pinned commit; not run. Version 1.4.31, reviewed again as a whole because the diff since 0c994e6 (1.4.22) is too large for a delta review: hooks.json, mod/register.ts and every shell hook (mcp-auth-gate, helper-report-relay, cache-edit-guard, skill-guard, dream-fence, settings-gate, artifact-backend-guard, channel-responder-invoked, channel-hook, validate-config, skill-lint, generate-summary, usage-track, user-prompt-pipeline, startup-context, stop-pipeline, stop-failure-stamp, subagent-cost, precompact-stamp) were read in full or through their full diff, as were the new browser-install, browser-host, browser-agent, browser, browser-viewer, lib/browser-runtime and browser-pins, transcribe, lib/transcription and transcription-pins, dream, trust-workspace, update-permissions, operator-auto-mode, watch-poll, and the diffs of apply-settings, automode-entries, deny-patterns, hermitd-start, hermitd-update.sh, hermitd-docker.sh, hermitd-cli, channel-send, channel-chats, the evolve-runner agent, the config template and the Docker templates; the rest (doctor-check, tasks, messages, dashboard and other libraries) was covered by targeted search for network calls, process spawning, file writes outside the project, env, settings and permission changes, eval and installs, and the prompt files by grep for allowed-tools, installs, URLs and permission modes. The function-hook module still registers prompt.submit, classic.PreModelSwitch, classic.PostModelSwitch, session.append, session.compact, session.start, classic.SessionStart, turn.step, turn.start and turn.complete, uses $.process.run for its bun bridge (now run from the session root), $.command.run for chat-issued !model, !effort, !compact and similar, auto-allows the matching model switch, drops typed prompts that forge a <channel envelope, and now also confirms /clear and /compact from a session-start.json stamp the SessionStart hook writes. Shell hooks: the PostToolUse tool-counter is gone; new are mcp-auth-gate (PreToolUse on mcp__*__complete_authentication asks, or denies for a non-admin chat sender, when the turn's origin is not a proven typed terminal or admin turn; PostToolUse on mcp__*__authenticate tells the model to send sign-in links only to the admin channel) and dream-fence (limits Edit/Write by the hermitd:dream-* subagents to Claude's auto-memory folder for the project and .hermit/compiled/topic-*.md). settings-gate now also denies task decide/answer/close commands whose claimed chat sender did not write in the turn, asks before shell or Edit/Write changes to scripts named by routines, channels or monitors, asks before Edit/Write of .claude/settings*.json unless the turn was typed at the terminal, and in auto mode lets watch saves (monitors running a shell check on a schedule through watch-poll) through without a prompt for the operator or the owning chat. Like skill-guard, these hooks run in every project, not only hermit projects, but outside a hermit they only add a prompt in edge cases. New features on by default: a managed browser (agent-browser 0.41.0 from registry.npmjs.org, Chrome for Testing from storage.googleapis.com and noVNC, all pinned by SHA hash, installed under ~/.claude/plugins/data/hermitd@<marketplace>/browser in the background at hatch or first use; signed-in profile in .claude.local/browser; cloud metadata URLs refused); voice-note transcription with pinned whisper.cpp, libgomp and ffmpeg from github.com and archive/ports.ubuntu.com and models from huggingface.co, installed under .claude.local/whisper.cpp, with an upload of the audio to api.groq.com or api.openai.com only when local transcription fails and GROQ_API_KEY or OPENAI_API_KEY is set (disclosed in docs/voice-notes.md); and a nightly dream routine (Sonnet, uses tokens) that exports the project's Claude Code transcripts into .hermit/state/dream and has two subagents rewrite Claude's auto-memory files, with before-copies and undo. Opt-in: phone takeover of the browser through Tailscale Serve, set up by the operator-run hermitd run browser-host, which installs Xvnc, openbox and unzip and an AppArmor profile with sudo prompts in the terminal; and trust-workspace apply, which behind an ask rule marks a cloned sub-repository trusted in ~/.claude.json and approves its .mcp.json servers. Permissions: the HERMIT_ALLOW list still contains Bash(bash -c 'AGENT_DIR=".hermit"*) and gains entries for dream, transcribe, brief-digest, task answer/cancel/decide and messages; the evolve-runner agent no longer writes it, instead hermitd update (operator-run) applies the new list without a separate prompt; Artifact is still added to permissions.allow; hatch defaults to permission_mode auto with a generated auto-mode allow policy. Unchanged: Discord and Telegram bot APIs with the channel plugin's token, telemetry webhook and backup git remote off by default, workspace trust seeded in ~/.claude.json at start, --dangerously-skip-permissions only when the operator picks bypassPermissions, optional watchdog service. No eval or remotely loaded code; the downloaded binaries themselves were not reviewed; no npm dependencies (bun required).

What it uses
  • Reads your prompts: yes
  • Sees tool calls: yes
  • Blocks or changes tool calls: yes
  • Reads or writes files: yes
  • Runs programs on your machine: yes
  • Calls the model (uses tokens): yes
  • Sends data over the network: yes
  • Stores data: yes
  • Shows something on screen: no
Hooked events
Screenshot of hermitd
Made this mod? Add a badge to your README

modsforClaude badge for hermitd

Paste this into your README. The badge says reviewed once the current version has been reviewed, and links readers to this page.

[![Listed on modsforClaude](https://modsforclaude.com/badge/hermitd.svg)](https://modsforclaude.com/mods/hermitd/)

← All mods