Skip to content

Hermes Agent Configuration Inventory

盘点 Nous Research Hermes Agent(v0.16.0)所有配置项、存储位置、读写方式

日期:2026-06-28 状态:verified(实验 + 官方文档交叉验证) 关联:Profile Tab 补全、Library Extension(mcp / skill / hooks / personality)


0. 方法论

每个结论标注来源:

标记含义
🧪实验验证(agent-box 创建 exp-hermes profile + 现有 hermes-main profile 观察)
📖官方文档(hermes-agent.nousresearch.com + GitHub NousResearch/hermes-agent)
🧪📖实验 + 文档双重验证

环境:Hermes Agent v0.16.0(Python 3.11,pip 安装),~/.hermes/ host 目录,agent-box 模板在 src/agent_box/templates/hermes/


1. 配置版图总览

1.1 用户可编辑配置(agent-box 应管理)

#配置项Hermes 读取路径Profile 内对应路径写入方式验证
1config.yaml~/.hermes/config.yaml (YAML, ~139 lines)dot-hermes/config.yamlYAML 合并写入🧪📖
2.env~/.hermes/.env (dotenv)dot-hermes/.env文本覆盖(保留 # 注释)🧪📖
3SOUL.md~/.hermes/SOUL.mddot-hermes/SOUL.mdMarkdown 覆盖🧪📖
4MCP Serversconfig.yamlmcp_servers:同 config.yamlYAML 内嵌,无独立文件🧪📖
5Skills~/.hermes/skills/ + external_dirsdot-hermes/skills/目录拷贝🧪📖
6Shell Hooksconfig.yamlhooks:同 config.yamlYAML 内嵌 + allowlist 文件🧪📖
7Memories configconfig.yamlmemory:同 config.yamlYAML 写入📖
8Compressionconfig.yamlcompression:同 config.yamlYAML 写入🧪
9Custom Personalitiesconfig.yamlagent.personalities同 config.yamlYAML 写入🧪
10Auxiliary modelsconfig.yamlauxiliary.*同 config.yamlYAML 写入📖
11Skills external_dirsconfig.yamlskills.external_dirs同 config.yamlYAML 写入📖

1.2 Hermes 自动生成(agent-box 不管理,但需保留)

#路径内容验证
12~/.hermes/.install_method安装方式标记(git / pip)🧪
13~/.hermes/.update_check更新检查状态(JSON)🧪
14~/.hermes/.env.bak.<timestamp>.env 的自动备份(每次编辑前生成)🧪
15~/.hermes/config.yaml.bak.<timestamp>config.yaml 的自动备份🧪
16~/.hermes/.hermes_historyCLI 交互历史🧪
17~/.hermes/auth.jsonOAuth 凭据池(credential_pool)🧪📖
18~/.hermes/auth.lockauth.json 写锁(OAuth 认证期间)🧪
19~/.hermes/audio_cache/TTS 音频缓存🧪
20~/.hermes/bin/ (tirith, uv, uvx)Hermes 自带的二进制依赖(预编译)🧪
21~/.hermes/cache/model_catalog.json模型目录缓存🧪
22~/.hermes/lsp/ (含 node_modules, package.json)LSP 服务依赖(Node.js 包)🧪
23~/.hermes/sessions/request_dump_<ts>_<hash>.json完整请求/响应 dump(按时间戳)🧪
24~/.hermes/memories/持久化记忆文件目录📖
25~/.hermes/skills/.bundled_manifest内置 skill 版本追踪(name:hash 行格式)🧪📖
26~/.hermes/skills/.curator_backups/skill curator 自动备份🧪
27~/.hermes/skills/.curator_statecurator 内部状态🧪
28~/.hermes/skills/.usage.json (+ .lock)skill 使用统计🧪
29~/.hermes/skills/.skills_prompt_snapshot.json当前 prompt 中注入的 skills 快照🧪
30~/.hermes/shell-hooks-allowlist.jsonshell hooks 一次性确认 allowlist📖
31~/.hermes/mcp-tokens/<server>.jsonOAuth MCP 凭据缓存📖

1.3 项目级 / 扩展配置(不在 Profile 内,非 agent-box 管辖)

#配置项路径备注验证
32Skill bundles~/.hermes/skill-bundles/<slug>.yaml优先级高于 individual skills📖
33Gateway hooks~/.hermes/hooks/<hook_name>/{HOOK.yaml,handler.py}Python-based,独立机制📖
34External skillsconfig.yaml::skills.external_dirs 列出的目录支持 ~${VAR} 展开📖
35Plugin hooks通过 plugin 加载(ctx.register_hook()与 shell hooks 不同的注册机制📖
36System-managed layerrequirements.toml 类机制(不存在,Hermes 无)n/a📖

2. 关键配置项详解

2.1 config.yaml 主要 sections

路径~/.hermes/config.yaml → Profile dot-hermes/config.yaml

优先级(从高到低):

CLI args (e.g., /model, --accept-hooks, --ignore-rules)
  > ~/.hermes/config.yaml
  > ~/.hermes/.env (仅用于 secrets)
  > built-in defaults

Schema 顶层 sections(📖 来自官方 configuration 文档 + 🧪 实际观察):

yaml
# ── 模型 ──
model:
  default: mimo-v2.5-pro           # 实际部署的模型
  provider: custom
  base_url: https://token-plan-cn.xiaomimimo.com/v1
  api_key: tp-...                  # ⚠️ 也常在 .env 中

# ── Terminal ──
terminal:
  backend: local                   # local|docker|ssh
  cwd: .
  timeout: 180
  docker_mount_cwd_to_workspace: false
  lifetime_seconds: 300
  container_cpu: 1
  container_memory: 5120
  container_disk: 51200
  container_persistent: true

# ── Browser ──
browser:
  inactivity_timeout: 120

# ── Tool loop guardrails ──
tool_loop_guardrails:
  warnings_enabled: true
  hard_stop_enabled: false
  warn_after:
    exact_failure: 2
    same_tool_failure: 3
    idempotent_no_progress: 2
  hard_stop_after:
    exact_failure: 5
    same_tool_failure: 8
    idempotent_no_progress: 5

# ── Compression ──
compression:
  enabled: true
  threshold: 0.5
  target_ratio: 0.2
  protect_last_n: 20
  protect_first_n: 3

# ── Prompt caching ──
prompt_caching:
  cache_ttl: 5m

# ── Memory ──
memory:
  memory_enabled: true
  user_profile_enabled: true
  memory_char_limit: 2200
  user_char_limit: 1375
  nudge_interval: 10
  flush_min_turns: 6
  write_approval: ...

# ── Session reset ──
session_reset:
  mode: both                       # both|idle|scheduled|none
  idle_minutes: 1440
  at_hour: 4

# ── Concurrency ──
max_concurrent_sessions: null
group_sessions_per_user: true

# ── Streaming ──
streaming:
  enabled: false

# ── Skills ──
skills:
  config: {}                       # 细粒度 skill 配置
  guard_agent_created: ...
  write_approval: ...
  external_dirs:                   # ⭐ 跨 agent 共享 skills
    - ~/.agents/skills

# ── Agent(行为 + personalities)──
agent:
  max_turns: 60
  api_max_retries: 3
  reasoning_effort: medium
  tool_use_enforcement: auto
  disabled_toolsets: []
  personalities:                   # 自定义 personalities
    helpful: You are a helpful...
    concise: ...
    # (14 个 built-in + 用户扩展)

# ── Auxiliary models(多模型)──
auxiliary:
  vision:        { model, provider, base_url, api_key }
  web_extract:   { ... }
  approval:      { ... }
  tts_audio_tags:{ ... }
  compression:   { ... }
  title_generation: { ... }
  skills_hub:    { ... }
  mcp:           { ... }
  triage_specifier: { ... }
  kanban_decomposer: { ... }
  profile_describer: { ... }
  delegation:    { ... }

# ── TTS / STT ──
tts:
  provider: openai
  voice: alloy
stt:
  provider: groq

# ── Display ──
display:
  compact: false
  streaming: true
  personality: null                # 当前激活的 personality

# ── Updates ──
updates:
  pre_update_backup: true
  backup_keep: 5
  non_interactive_local_changes: stash

# ── MCP Servers(嵌入 config.yaml)──
mcp_servers:
  my-server:
    command: npx
    args: ["-y", "@some/mcp-server"]
    env: { API_KEY: "..." }
  my-http-server:
    url: https://example.com/mcp
    auth: oauth                    # 或省略为匿名 HTTP
    headers: { X-Custom: "value" }

# ── Shell Hooks(嵌入)──
hooks:
  pre_tool_call:
    - matcher: "<regex>"
      command: "/path/to/hook.sh"
      timeout: 60
  post_tool_call:
    - matcher: ""
      command: "..."
hooks_auto_accept: false

# ── Code execution ──
code_execution: ...

# ── Worktree ──
worktree: ...

# ── Privacy ──
privacy: ...

# ── Other ──
quick_commands: ...
human_delay: ...
context_file_max_chars: ...
file_read_max_chars: ...
tool_output: ...
credential_pool_strategies: ...
providers.<id>: ...
unauthorized_dm_behavior: ...

2.2 .env 结构

路径~/.hermes/.env (dotenv 格式)

支持 keys(部分确认,部分 📖 推测):

bash
# 主要 LLM providers
OPENROUTER_API_KEY=
OPENAI_API_KEY=
OPENAI_BASE_URL=

# Auxiliary models
VOICE_TOOLS_OPENAI_KEY=
GROQ_API_KEY=
GROQ_BASE_URL=
STT_GROQ_MODEL=
STT_OPENAI_MODEL=
STT_OPENAI_BASE_URL=

# Search / web
EXA_API_KEY=
PARALLEL_API_KEY=
FIRECRAWL_API_KEY=
TAVILY_API_KEY=

# Browser
BROWSERBASE_API_KEY=
BROWSER_USE_API_KEY=

# Image / Video
FAL_API_KEY=

# LLM
ANTHROPIC_API_KEY=
GMI_API_KEY=
GOOGLE_API_KEY=

# Compute
MODAL_TOKEN_ID=
MODAL_TOKEN_SECRET=
DAYTONA_API_KEY=

# Terminal (SSH)
TERMINAL_SSH_HOST=
TERMINAL_SSH_PORT=
TERMINAL_SSH_USER=
TERMINAL_SSH_KEY=
TERMINAL_SSH_PERSISTENT=
TERMINAL_LOCAL_PERSISTENT=
TERMINAL_SCRATCH_DIR=
TERMINAL_SANDBOX_DIR=

# Internal
HERMES_API_TIMEOUT=
HERMES_API_CALL_STALE_TIMEOUT=
HERMES_STREAM_READ_TIMEOUT=
HERMES_STREAM_STALE_TIMEOUT=
HERMES_DOCKER_BINARY=
HERMES_HOME=
HERMES_REAL_HOME=
HERMES_LANGUAGE=
HERMES_FILE_MUTATION_VERIFIER=0
HERMES_ACCEPT_HOOKS=1

Hermes 区分 secrets 与 config:📖 docs 明确 "Secrets → .env. Everything else → config.yaml"。但 🧪 实际观察到 config.yaml 中也常包含 api_key 字段(如 host 中 model.api_key),这是一个潜在的重复配置点

2.3 SOUL.md — Personality 文件

路径~/.hermes/SOUL.md

📖 官方文档:"Primary agent identity",是 system prompt 的 slot #1。Hermes 只从 $HERMES_HOME 加载,不读 cwd。

机制

  • 文件不存在 → 自动创建默认模板(永不覆盖已有
  • 文件为空/不可读 → fallback 到 built-in 默认 identity
  • 内容(去 comment 后)原样注入 prompt slot #1
  • 14 个 built-in personalities:helpful, concise, technical, creative, teacher, kawaii, catgirl, pirate, shakespeare, surfer, noir, uwu, philosopher, hype
  • 自定义 personalities 在 agent.personalities.<name>: "..." 中定义
  • AGENTS.md(项目约定)和 /personality(会话级 overlay)区分

实验验证(🧪):host ~/.hermes/SOUL.md 是 537 字节的注释模板(未定制),由 user 后续编辑启用。

2.4 MCP Servers — 嵌入 config.yaml

📖 Hermes 没有独立的 MCP 文件。所有 MCP 配置都在 config.yamlmcp_servers: 块下。Hermes 根据字段存在性推断 transport:

  • command → stdio
  • url → HTTP
  • auth: oauth → OAuth-authenticated HTTP
  • client_cert/client_key → mTLS

Schema

yaml
mcp_servers:
  my-stdio: # stdio
    command: npx
    args: ["-y", "@some/mcp"]
    env: { API_KEY: "..." }
    timeout: 30
    connect_timeout: 10
    enabled: true
    supports_parallel_tool_calls: true
    tools:
      include: ["tool1", "tool2"]
      exclude: ["tool3"]

  my-http: # anonymous HTTP
    url: https://example.com/mcp
    headers: { X-Custom: "value" }

  my-oauth: # OAuth
    url: https://api.example.com/mcp
    auth: oauth
    oauth:
      client_id: ...
      client_secret: ...

  my-mtls: # mTLS
    url: https://internal.example.com/mcp
    client_cert: |
      -----BEGIN CERTIFICATE-----
      ...
    client_key: |
      -----BEGIN PRIVATE KEY-----
      ...

# 高级:per-server sampling config
mcp_servers.my-server:
  sampling:
    enabled: true
    model: anthropic/claude-opus-4.6
    max_tokens_cap: 4096
    timeout: 30
    log_level: info

OAuth token 缓存~/.hermes/mcp-tokens/<server>.json

Catalog:仓库 optional-mcps/<name>/manifest.yaml

Per-server 命名规则:工具以 mcp_<server>_<tool> 暴露给模型

冲突规则:include 与 exclude 同时存在时,include wins

Hermes 作为 MCP serverhermes mcp serve(stdio,~200ms 事件轮询,暴露 10 个工具如 conversations_list, messages_send, permissions_respond

CLI

  • hermes mcp add — discovery-first install
  • hermes mcp list / ls
  • hermes mcp test <server> — 测试连接
  • hermes mcp configure / config — 切换工具选择
  • hermes mcp login — OAuth 强制重认证
  • hermes mcp picker / catalog / install

2.5 Skills — ~/.hermes/skills/

Single source of truth~/.hermes/skills/<name>/SKILL.md(📖 官方明确)

External dirs(在 config.yaml::skills.external_dirs):

yaml
skills:
  external_dirs:
    - ~/.agents/skills # ⭐ 与 Claude Code / OpenCode 共享
    - /home/shared/team-skills
    - ${SKILLS_REPO}/skills # ${VAR} 展开

优先级:local (~/.hermes/skills/) > external(按 external_dirs 顺序)

Bundled skills:仓库 skills/ → 首次安装时拷贝到 ~/.hermes/skills/,版本追踪在 .bundled_manifest(格式:name:hash 行)

Hub sources(📖):official, skills-sh, well-known, github(默认 taps: openai, anthropics, huggingface, NVIDIA, gstack), clawhub, claude-marketplace, lobehub, browse-sh, url

Trust levelsbuiltin > official > trusted > community

格式:agentskills.io 兼容

发现方式:三层 progressive disclosure:skills_list()skill_view(name)skill_view(name, path)

Skill bundles~/.hermes/skill-bundles/<slug>.yaml):一组 skills 的集合定义,优先级高于 individual skills

实验验证(🧪):host ~/.hermes/skills/ 包含:

  • 多个真实目录:apple/, autonomous-ai-agents/, creative/, data-science/, devops/, dogfood/, dw-handoff/, email/, github/, media/, mlops/, note-taking/, productivity/, red-teaming/, research/, smart-home/, social-media/, software-development/, yuanbao/
  • 多个符号链接~/.agents/skills/frontend-design, frontend-design-review, mmx-cli, pdf, skill-creator, slack-gif-creator, theme-factory, vercel-*, web-*, writing-guidelines, xlsx
  • 自动生成文件:.bundled_manifest(记录 50+ bundled skill hash), .curator_backups/, .curator_state, .usage.json (+ lock), .skills_prompt_snapshot.json

Cross-agent 关系:Hermes 通过 skills.external_dirs 读取 ~/.agents/skills/,与 Claude Code 的 cross-tool 标准路径一致。agent-box 应确保 hermes profile 的 config.yaml::skills.external_dirs 包含 ~/.agents/skills(如果 host 中存在)。

2.6 Hooks — 三种机制并存

📖 官方文档明确:Hermes 有三种 hook 系统,不是一种:

系统注册位置作用范围实现
Gateway hooks~/.hermes/hooks/<name>/{HOOK.yaml,handler.py}仅 GatewayPython
Plugin hooksctx.register_hook() in pluginCLI + GatewayPython
Shell hooksconfig.yaml::hooks: blockCLI + GatewayShell

Shell hooks(最常用,可被 agent-box 管理):

yaml
hooks:
  pre_tool_call:
    - matcher: "<regex>" # 工具名 regex 匹配
      command: "/path/to/hook.sh"
      timeout: 60
  post_tool_call:
    - matcher: ""
      command: "..."
  pre_llm_call:
  post_llm_call:
  on_session_start:
  on_session_end:
  on_session_finalize:
  on_session_reset:
  subagent_start:
  subagent_stop:
  pre_gateway_dispatch:
  pre_approval_request:
  post_approval_response:
  transform_tool_result:
  transform_terminal_output:
  transform_llm_output:

hooks_auto_accept: false # true 则跳过 allowlist 提示

Wire protocol:JSON via stdin/stdout

Allowlist~/.hermes/shell-hooks-allowlist.json

json
{
  "approvals": [
    {
      "event": "post_llm_call",
      "command": "/home/.../hook.py"
    }
  ]
}

每个唯一 (event, command)首次运行时提示用户确认,之后持久化在 allowlist。

绕过机制

  • CLI flag: --accept-hooks
  • Env var: HERMES_ACCEPT_HOOKS=1
  • Config: hooks_auto_accept: true

CLI

  • hermes hooks list | ls — 列出所有 hooks
  • hermes hooks test <event> — 用合成 payload 测试
  • hermes hooks revoke / remove / rm — 撤销 allowlist
  • hermes hooks doctor — 检查 exec bit、allowlist、mtime drift、JSON validity、synthetic run timing

Ordering:Python plugin hooks 先 → shell hooks 后;首个 valid {"action": "block", ...} wins

Gateway 事件gateway:startup, session:start/end/reset, agent:start/step/end, command:*

2.7 覆盖机制

CLI 覆盖

bash
hermes -m anthropic/claude-opus-4.6          # 短选模型
hermes --provider custom --base-url https://...
hermes --accept-hooks                        # 跳过 hooks allowlist
hermes --ignore-rules                        # 忽略项目 AGENTS.md
hermes --ignore-user-config                  # 完全跳过 ~/.hermes
hermes --resume <session-id>
hermes --continue                            # 继续上次
hermes --worktree
hermes --yolo                                # 跳过所有权限
hermes --pass-session-id
hermes --tui | --cli | --dev

Slash commands/new, /reset, /retry, /undo, /compress, /usage, /insights [--days N], /stop, /platforms, /status, /sethome, /personality, /model, /skills, /reload-mcp, /learn, /skin

Env var override

  • HERMES_HOME / HERMES_REAL_HOME — 覆盖 home dir
  • TERMINAL_<KEY_UPPERCASE> — 覆盖 terminal.* 配置
  • AUXILIARY_VISION_* / AUXILIARY_WEB_EXTRACT_* — 覆盖 auxiliary 模型
  • ${VAR_NAME} 在 config.yaml 中替换(undefined 时保留原样)

Subcommands(50+,agent-box 应通过 CLI 包装而非 GUI 覆盖):

chat, model, fallback, secrets, migrate, gateway, proxy, lsp,
setup, postinstall, whatsapp, slack, send, login, logout, auth,
status, cron, webhook, portal, kanban, hooks, doctor, security,
dump, debug, backup, checkpoints, import, config, pairing, skills,
bundles, plugins, curator, memory, tools, computer-use, mcp,
sessions, insights, claw, version, update, uninstall, acp, profile,
completion, dashboard, desktop, gui, logs, prompt-size

2.8 备份机制(Hermes 自动)

📖 updates.pre_update_backup: true + backup_keep: 5 + non_interactive_local_changes: stash|discard

.env.bak.20260610_145143
.env.bak.20260610_150307
.env.bak.20260612_085832
...
config.yaml.bak.20260610_145143
...

⚠️ agent-box 复制模板到 profile 时,应避免携带这些 .bak.* 文件,或显式列出 ignore 规则。

2.9 Schema 版本化

📖 hermes config migrate 会自动添加缺失的配置项。hermes doctor 会报告 "Config version outdated (v0 → v28)" 提示用户升级。


3. 关键差异 vs Claude Code

维度Claude CodeHermes影响
配置格式JSONYAML + dotenv (.env)agent-box config editor 需要 YAML reader/writer
MCP 文件~/.claude.json::mcpServersconfig.yaml::mcp_servers:(嵌入)agent-box mcp.apply_hermes 直接修改 config.yaml
HooksJSON hooks.PreToolUse[]YAML hooks.pre_tool_call[] + 3 种 hook 系统agent-box hooks.apply_hermes 需要支持三种系统的不同入口
SOUL/personality无对应SOUL.md + agent.personalities + 14 built-inagent-box 可加 Personality tab
Skills 路径~/.claude/skills/ + ~/.agents/skills/~/.hermes/skills/ + external_dirsagent-box skills.apply_hermes 复制到 dot-hermes/skills/<id>/
Auxiliary models11+ 独立模型(vision, web_extract, approval...)agent-box 可选择性暴露
CLI 标志/model, /clear/model, /personality, /skills, /learn 等 18+ slash commandsagent-box 可命令包装
备份机制.env.bak.* + config.yaml.bak.* 自动备份agent-box profile copy 需 ignore .bak.*
OAuth无对应auth.json credential_pool + 多 provider 支持agent-box 可统一管理
Gateway / desktop / gui / dashboard / acp无对应内置 5 种 server modeagent-box 不应触碰(用户独立选择)

4. agent-box 当前实现状态

4.1 ✅ 已实现

功能实现
MCP 应用mcp.py::_apply_hermes → 写入 config.yaml::mcp_servers.<id>,剥离 type 字段(Hermes 通过 command/url 推断 transport)
Skills 应用skills.py → 复制到 dot-hermes/skills/<id>/
Profile 模板templates/hermes/ 包含 config.yaml + .env

4.2 ❌ 缺失 / 待实现

缺失项影响
Shell hooks 应用用户无法通过 GUI 配置 hooks.pre_tool_call
Shell hooks allowlist 管理用户首次运行新 hook 时卡在 stdin 提示
SOUL.md 编辑用户无法通过 GUI 编辑 personality
Personalities 管理14 个 built-in + 自定义 personalities 无 GUI 入口
Auxiliary models 配置11+ 独立模型的 GUI 缺失
.env 多 key 管理30+ API keys 无统一管理
Skills external_dirs 注入用户无法通过 GUI 启用 ~/.agents/skills 共享
OAuth credential pool 视图auth.json 中的 credential_pool 不可见
Backup ignoreprofile copy 携带 .bak.* 文件
Bundled manifest 保护dot-hermes/skills/.bundled_manifest 可能被覆盖
Config version migrationhermes config migrate 不在 GUI 触发
50+ subcommand 集成用户必须跳出 agent-box 才能用 cron, webhook, kanban
LSP / Bundles / Plugins / Memory用户无法通过 GUI 配置

5. Template 改进建议

src/agent_box/templates/hermes/ 当前内容:

config.yaml  — 17 行(model, terminal, memory, compression, display 等)
.env         — 1 行(HERMES_API_KEY=)

建议补全

yaml
# config.yaml additions
skills:
  external_dirs:
    - ~/.agents/skills # 启用 cross-tool skills 共享

hooks_auto_accept: false # 强制 allowlist 流程

agent:
  max_turns: 60
  reasoning_effort: medium

streaming:
  enabled: false

memory:
  memory_enabled: true
  user_profile_enabled: true

session_reset:
  mode: both
  idle_minutes: 1440

updates:
  pre_update_backup: true
  backup_keep: 5
  non_interactive_local_changes: stash

新增文件:

  • SOUL.md — 默认 personality 模板(537 字节,Hermes 自动创建默认模板)
  • skills/.gitkeep — 让 skills 目录在 profile 落地时存在
  • memories/.gitkeep — 让 memories 目录存在

Profile copy 时 ignore

  • .bak.*(Hermes 自动备份)
  • .usage.json.usage.json.lock.curator_backups/.curator_state(runtime 状态)
  • .skills_prompt_snapshot.json(运行时快照)
  • auth.json 中的 credential_pool(运行时会刷新)
  • bin/(tirith, uv, uvx,Hermes 自带二进制)
  • lsp/node_modules/(LSP 包依赖,体积大)
  • sessions/request_dump_*(请求 dump,体积大且可能含敏感信息)

6. Profile Tab 设计建议

Hermes profile 应包含以下 tab(基于本 inventory 的用户可编辑配置项):

Tab包含项编辑方式
Generaldisplay_name, description, agent_type表单
Providermodel.default, model.provider, model.base_url, api_key表单 + 文本
API Keys30+ .env keys(OpenRouter, OpenAI, Exa, Parallel...)Key-value editor
Auxiliaryauxiliary.vision, web_extract, approval, tts, stt 等 11+ 模型表单
Terminalterminal.backend, timeout, cwd, docker_* 配置表单
Memorymemory.* 全字段, compression.*表单
MCP Servers启用/禁用 mcp_servers 条目Library 引用
Skills启用/禁用 skills(来自 Library),external_dirs 配置Library 引用
Hooksshell hooks(按 event 分组)YAML editor + 测试按钮
PersonalitySOUL.md 编辑, personalities 列表, 当前激活 personalityMarkdown + 表单
Sessionsession_reset, max_concurrent_sessions, group_sessions_per_user表单
Updatesupdates.pre_update_backup, backup_keep, non_interactive_local_changes表单
OAuthcredential_pool 中各 provider 状态只读 + 重新认证
Advancedraw config.yaml + raw .envYAML / dotenv editor

7. 待验证 / 已知缺口

状态备注
~/.hermes/memory/ 是否独立目录📖 推测但 🧪 未观察到docs 说有 memories/ 但 host 中无
14 built-in personalities 完整列表🧪 部分(kawaii, catgirl, pirate, shakespeare, surfer, noir, uwu, philosopher, hype, helpful, concise, technical, creative, teacher — 共 14)🧪 host config.yaml 中找到全部
所有 .env keys 完整性📖 部分docs 列举一些,host 中 .env 还含 OPENROUTER 等
LSP / Bundles / Plugins / Curator 子系统📖 未深入子命令存在但详细 schema 未文档化
gateway / proxy / portal 配置📖 未深入子命令存在但与 agent-box 关系不明确
dashboards / desktop / gui 启动📖 未深入推测为 Hermes 自己的 server 模式
acp (Agent Client Protocol) server📖 未深入与 opencode acp 类似

8. Sources

类型URL
📖 官方https://hermes-agent.nousresearch.com/docs/user-guide/configuration
📖 官方https://hermes-agent.nousresearch.com/docs/user-guide/features/skills
📖 官方https://hermes-agent.nousresearch.com/docs/user-guide/features/hooks
📖 官方https://hermes-agent.nousresearch.com/docs/user-guide/features/mcp
📖 官方https://hermes-agent.nousresearch.com/docs/user-guide/features/personality
📖 GitHubhttps://github.com/NousResearch/hermes-agent
🧪 实验~/.hermes/ host 目录结构(2026-06-28)
🧪 实验~/.agent-box/profiles/hermes-main/dot-hermes/ 已运行 profile
🧪 实验agent-box create exp-hermes --type hermes(本次创建的 test profile)