Hermes for Organisations

Run one Hermes Agent per employee, and manage what they share from one place. A bundle playbook carries the organisation's baseline: LLM providers, MCP servers, skills, the bots to install and the shared knowledge to read. apb hermes sync applies it to a machine. Every bot is its own playbook.

This guide covers the delivery mechanism. For memory itself see Hermes Memory Provider.

The pieces

Bundle and bot playbooks are read from the agent endpoint (GET /api/playbooks/?format=json). It answers anonymously for public and unlisted playbooks, so on an instance that only the intranet can reach, no employee needs a credential to receive the baseline. Do not put anything secret in a bundle: its config is readable by everyone who can reach the instance.

config.hermes

Set it in the bundle playbook's config field. Every key is optional.

{
  "hermes": {
    "managed": {
      "providers": {
        "local": { "name": "Local Ollama", "base_url": "http://llm.intranet:11434/v1",
                   "api_key": "no-auth", "models": ["qwen3.8:27b"], "discover_models": false }
      },
      "mcp_servers": { "sql": { "url": "http://sql-mcp.intranet:8006/mcp", "timeout": 120 } }
    },
    "defaults": { "model": { "default": "qwen3.8:27b", "provider": "local" } },
    "env": { "MSGRAPH_CLIENT_ID": "00000000-0000-0000-0000-000000000000" },
    "bots": ["<bot-playbook-guid>"],
    "shared_memory": ["<shared-knowledge-guid>"]
  }
}

A bot playbook may set config.hermes.profile_name. Otherwise the profile name comes from the playbook name, lowercased and with accents removed. It may also set config.hermes.version.

Sync a machine

apb hermes sync <bundle-guid> --url=https://agents.example.org \
  --managed-dir="%LOCALAPPDATA%\example\hermes-managed" \
  --memory=<personal-memory-guid> --apply

Without --apply the command only shows its plan. With it:

• It writes config.yaml and .env into the managed directory. The bundle's

skills go to skills/, which is pinned as skills.external_dirs.

• It writes each bot as a profile distribution under bots//. Hermes

installs it with hermes profile install, or hermes profile update once it exists. A same-named profile that was not installed from a distribution is left alone.

• It merges defaults, plus a bot's own MCP servers, into each profile's

config.yaml where unset.

• With --memory, it configures the memory provider in every one of those

profiles: the personal playbook, plus shared_memory. It reads the personal key from AGENTPLAYBOOKS_MEMORY_API_KEY once and keeps it in the managed .env.

• It records state.json: bundle, time, skills, bots and any failures.

Credentials never come from the bundle. Pass them with --env-file=: a KEY=VALUE file that only the intended people can read, for example on a protected share. Its values go into the managed .env and stay there for later syncs, even when the file cannot be reached. A key that the bundle's env used to set, and no longer does, is removed.

Hermes reads the managed layer only when HERMES_MANAGED_DIR points at the managed directory. Set it once, as a user environment variable, in your installer. On an intranet instance without TLS, add --allow-insecure-http; see Self-Hosting.

Run the same command from a scheduled task, for example at logon and daily. That keeps every machine on the current bundle. A bot update never touches the bot's sessions or memories.

Learning without leaking cases

Hermes keeps learning locally. Its own MEMORY.md and the employee's private playbook may hold case details. Shared knowledge is kept separate:

• Every profile reads the shared knowledge playbook. It cannot write to it.
• Lessons reach it only through review. A person proposes a generalised

principle or method, with no names, case numbers or document content. A curator checks it and promotes it.

• Nothing is promoted automatically.

Limits of this version

• A user-level managed directory is advisory: the employee can edit it. Real

enforcement needs a machine-level HERMES_MANAGED_DIR that users can only read, set by your device management.

• Bundle and bot playbooks must be public or unlisted. Organisation-only

visibility is planned.

• Each employee creates their private memory playbook and its key once, in the

web app.