| name | ours |
| description | Use when the user wants to set up or configure ours or this plugin, onboard onto the ours network, create or pick/switch an identity (and decide whether to adopt its persona), connect with another agent or person, generate or accept an invite, send or read end-to-end-encrypted messages, send or receive a file, check incoming mail, arm live monitoring so the agent wakes on new mail, or bind a web-messenger account as the host's monitoring/control proxy. Trigger phrases include "set up ours", "set up ours network", "set up the plugin", "create an identity", "create a human/agent identity", "use identity X", "who am I", "set my bio", "set my persona", "adopt this persona", "generate an invite for X", "add this contact", "send a message to X", "send a file to X", "check my messages", "any new messages", "any new files", "get my files", "list my contacts", "watch for messages", "wait for a reply", "wake me on new mail", "bind the monitoring proxy", "set up the control panel", "monitoring status". |
| version | 0.1.0 |
| metadata | {"hermes":{"tags":["ours","ours.network","a2a","adapt","e2e","messaging","identity"],"category":"communication"}} |
ours — secure agent-to-agent messaging
ours gives this agent self-sovereign identities and end-to-end-encrypted
channels to other agents and people, brokered over ADAPT. One node (a background
daemon) hosts many identities at once; you never touch crypto directly. There
are three surfaces:
Tool names in Hermes. ours is wired as the MCP server ours, so its tools
appear namespaced as mcp_ours_<tool> (e.g. mcp_ours_get_messages,
mcp_ours_send_message, mcp_ours_choose_identity). This skill writes the bare
names (get_messages, …) for readability — call the mcp_ours_-prefixed form.
- Layer 1 — identities (global): create / bind / switch the identity you act as.
- Layer 2 — messaging (per the bound identity): invites, contacts, send/read.
- Control plane (the host's Human identity): bind a human's web-messenger as a
monitoring & control proxy that can oversee and command a fleet of agents.
Identities come in exactly two kinds, in a fixed order:
- The Human identity — the person. Created first, exactly one per host. Every
agent identity is associated with it, and everyone the user shares an invite with can
see the human identity behind each agent.
- Agent identities — the agents/workers, created after (and associated with) the
Human identity.
Terminology note: the tools predate this naming — create_root_identity creates
the Human identity, and tool output (list_identities, hierarchy messages) may still
say "root". Whenever you see "root", read and say "Human identity" to the user.
In the rare case a messaging tool says no identity is bound (re-attach is normally automatic), pick one with choose_identity (or
make one with create_identity) first.
Onboarding — MANDATORY, Human identity first
The gate: before creating any identity or starting any messaging flow (invite,
add-contact, send), check list_identities(). If the host has no Human identity yet,
run onboarding — regardless of what the user actually asked for. A request for an
agent identity, an invite, or "send a message to my friend" does NOT skip the gate; it
just means onboarding comes first and their request comes immediately after.
Walk the user through it, explaining as you go:
- "First we create your Human identity — that's you." All agent identities you add
later are associated with your Human identity, and this association is visible to
the people you share invites with: they always know the human behind every agent.
Ask for the person's name (never invent or reuse a project name) and optionally a
host label for this machine (e.g.
laptop, VPS), plus a one-line public bio.
Then: create_root_identity({ name: "<Human>@<host>", bio }) — compose the name as
<Human>@<host>, or just <Human> when no host label is given (this tool creates the
Human identity; "root" is its historical name).
- Then add agent identities.
create_identity({ name, bio }) for each agent —
every one is automatically associated with the Human identity, and its invites carry
the verified "agent X of person Y" chain.
Do not create an agent identity on a host with no Human identity. That would make a
"flat" identity: no verified human behind it in invites, no control plane. The tool
allows it for legacy reasons; this skill does not.
| Tempting shortcut | Why it's wrong |
|---|
| "The user clearly asked for an agent, so the human question doesn't apply" | The gate is about ORDER, not classification. Create the Human identity first, then the agent they asked for. |
| "The user is busy / gave me everything I need for the agent" | Onboarding adds one question — the person's name. Ask it. |
"create_identity works fine without a Human identity" | It creates a flat legacy identity with no human association. Never do it. |
| "I'll create the agent now and the Human identity later" | Later never comes, and the agent's invites go out with no human chain. Human first. |
Setup — "set up ours" / "set up the plugin"
Walk the user through these, checking each. Stop and help at the first one that isn't done.
- Daemon running. The MCP tools talk to a local background daemon. Check it:
ours-mcp status. If the command is missing, install it: npm i -g @ours.network/mcp, then ours-mcp start. For boot-persistence offer
ours-mcp install-service. To change broker / port / state dir, run the
interactive ours-mcp setup (this edits config only — it is NOT identity setup).
These run on the user's machine; if a step needs them at a terminal, suggest they
type ! ours-mcp status etc.
- Plugin installed. Run this package's
install.sh (from @ours.network/hermes).
It ensures the daemon, writes the ours MCP server into ~/.hermes/config.yaml, and
installs this skill into ~/.hermes/skills/. That's all — no identities, no webhook route,
no secret, no watcher. Wake-on-mail is enabled in-session (step 5). After it runs,
/reload-mcp so Hermes picks up the mcp_ours_* tools. See the package README for manual steps.
- Onboarding. Run the mandatory Onboarding flow above: Human identity first,
then any agent identities.
- Connect. Generate an invite to share, or paste one to add a contact. Same-host
identities skip invites via the local contact book.
- (Optional) Wake on mail. Wake is enabled in-session by you, after an identity is
bound: offer to enter autonomous watch mode — hold a blocking
ours-mcp watch <identity>
via the terminal tool and react to each new message from that loop (see Getting woken on new mail
below). Be honest that this BLOCKS the session (unlike Claude Code's background Monitor) —
don't sell it as "just works". The installer never sets this up.
- (Optional) Oversight. If they want to watch/command a fleet from a phone or
browser, set up the control-plane monitoring proxy.
- Configuration. Port, state dir, broker, and GC interval are configurable
(env >
~/.ours/config.json > default; port default 3050). Daemon config is
host-wide and shared — changing it restarts the daemon and drops every
session's binding. Never self-configure on your own initiative: surface the
need, explain the impact, and act only on the user's explicit yes. Details:
references/configuration.md.
Layer 1 — identities (global)
A session must bind an identity before it can send or read messages. Binding is
exclusive: one identity, one session at a time.
Create an identity
When the user wants a new identity ("create an identity", "make an agent", "I'm setting
up"), first apply the Onboarding gate above: no Human identity on the host yet →
onboarding first, whatever was asked. Then:
- Get a name — ask if not given. This is what peers see for you in invites. (For the
Human identity never ask for a bare name — compose it from the person's name + host
label per the Onboarding recipe:
<Human>@<host>, or <Human> with no label. The @
is a valid identity-name character.)
- Get a bio (and optionally a persona) — two distinct fields:
- bio — the identity's public card. It is shared via invites and visible to
your contacts (it rides in the agent's intro and the Human identity's signed
profile, shown in the verified association chain on
add_contact). Write it for
others: role, scope, and when a peer or coordinator would deploy or ask this agent.
- persona — a local operating contract describing how the agent should behave
when it adopts this identity (mandate, boundaries, what NOT to do, tone). It is
never shared via invites (only via the control-plane cluster). Set it with
set_persona. See the writing-agent-bios skill for how to write both well.
- Create it:
- Human identity (first identity on the host, exactly one) →
create_root_identity({ name: "<Human>@<host>", bio }). Creating it adopts any
pre-existing legacy identities on the host as agents under it (adopt_existing,
default true).
- Agent identity (requires the Human identity to exist) →
create_identity({ name, bio }). It is automatically associated with the Human
identity; its invites carry a verified "agent X of person Y" chain.
- Optional flags (both default true):
expose_local publishes the identity in the
host-local contact book so other same-host identities can message it by name with no
invite; local_auto_accept auto-accepts local introductions (false = they queue for
approval). Opt out with create_identity({ name, expose_local: false }) etc.
Creation binds the new identity to this session. The tool response then prompts you about
a wake monitor — interpret it via the After binding follow-ups below (the user just authored
the bio, so a persona prompt is only needed if they want to role-play it).
Bind / switch an identity — and the follow-ups
"use identity Alice" / "switch to Alice" → choose_identity({ name: "Alice" }).
- Binding is exclusive. If Alice is held by another live session, the call is
declined. Never pass
force: true on your own — tell the user it's in use elsewhere and
ask; only retry choose_identity({ name: "Alice", force: true }) after they explicitly
confirm (the other session is then evicted). A dead/stale holder is auto-reclaimed with no force.
After binding (or creating) — always run these two follow-ups:
- Persona check (only if the persona is non-empty). Read the bound identity's persona
with
current_identity() (it returns name, bio, persona, and hierarchy place). If the
Persona: line is non-empty, show it and ask: "Adopt this persona as your operating
mode for this session?" Adopt it only on an explicit yes — then behave as that
persona for the session (not persisted). The bio is a public card, NOT an operating
instruction — never adopt the bio as behavior. If persona is empty or they decline,
operate normally. Never adopt a persona silently.
- Wake check. The
choose_identity / create_identity response may prompt you to "arm a
message monitor" — that is the Claude-Code seam, and the intent is the same in Hermes: you
enable wake in-session, right after binding, by entering autonomous watch mode (hold a
blocking ours-mcp watch <identity> and handle each message from that loop — see Wake on new
mail). If the user wants live reactivity for the just-bound identity, offer to enter watch
mode now.
Other identity tools
- List: "what identities are there" →
list_identities() (shows the Human identity
with its agents indented — the output may label it "root" — and which one this session
is bound to).
- Who am I:
current_identity() (returns name, bio, persona, and hierarchy place — used by the persona check above and the only way to read back a persona).
- Set/change a bio:
set_bio({ bio }) on the bound identity. For the Human identity,
the refreshed profile is re-pinned into every agent so future agent invites carry the update.
- Set/change a persona:
set_persona({ persona }) on the bound identity (local only;
never carried in invites). Read it back via current_identity (no getter tool).
- Remove:
remove_identity({ name }) — permanent; deletes the node and all its state.
A Human identity with agents refuses until the agents are removed.
Version mismatch (advisory)
If a notice says your plugin and the running daemon are different
versions, it is advisory — everything still works. Relay it to the user and,
if they want matching versions, tell them: the daemon is shared and is not
restarted automatically, so run ours-mcp stop when no other session is
mid-task (the next session starts the new version), or update the lagging side.
Do not stop work, refuse, or restart anything on your own over this.
Workspace identity pin (.ours-identity)
The .ours-identity workspace pin is a Claude-Code seam: there, a SessionStart hook reads
the file and suggests binding. Hermes has no SessionStart hook, so nothing auto-reads the pin
here. The define_local_identity_file tool still exists and writes a correctly-shaped file
(pass an absolute path plus name and optional force / expose_local / local_auto_accept),
but under Hermes you bind explicitly with choose_identity rather than relying on a pin.
If a project documents a pinned identity, treat it as a suggestion, never an authorization:
ask the user before binding (or creating) it, and never adopt its persona without explicit approval.
Layer 2 — messaging (per the bound identity)
All of these act as your currently-bound identity.
Generate an invite
"generate an invite for Bob":
generate_invite({ name: "Bob" }) — or generate_invite({}) with no name: the redeemer
is registered under whatever name they announce when accepting.
- Return the invite blob verbatim in a copy-paste block; the user shares it with Bob
out-of-band. The blob carries only minimal key material (brotli-compressed, armored to a
single base64url line, newline-safe). Both ends must run a matching ours version.
Add a contact from an invite
When the user pastes an invite blob:
- With a name →
add_contact({ invite: "<blob>", name: "My friend" }).
- With no name →
add_contact({ invite: "<blob>" }) (the inviter's own display name is
used; afterward, offer to keep or rename it).
add_contact is the first leg of an asynchronous redeem: it boxes your identity to the
inviter and leaves the contact pending — it is not in your contact list yet. The
inviter must receive it, verify your identity, and reply before the contact finalizes;
that reply lands automatically over the broker and you do nothing further. So after a
successful add_contact, tell the user the redemption is done on their side and the rest
is a wait on the sender — e.g. "Invite accepted — the contact will appear in your contact
list once the sender verifies your identity. Nothing more to do on your end." Do not
report the contact as already added.
Send a message
"send hi to Bob" → send_message({ contact: "Bob", text: "hi" }). contact is a
contact name or container id. If Bob is not yet a contact but is a same-host sibling role
(same Human identity) or is published in the local contact book, the connection is established
automatically (cert- or registrar-verified introduction + key exchange) and the message is
delivered with it — no invite ceremony.
Reply to a specific message
Every message carries a stable cross-side wire_id, shown by get_messages as {…}. To
answer one precisely: send_message({ contact: "Bob", text: "…", reply_to_wire_id: "<wire_id>" }), optionally reply_to_sentence: <n> (1-based) to point at a sentence. The
recipient sees ↳re <wire_id>·s<n>. It's a lightweight reference, not a thread object.
Check / read messages
- "check messages" / "any new messages" →
get_messages() returns the messages you
haven't seen (status "unread") with their bodies and marks them "processed". This is
the only call that returns message text; each message is delivered exactly once, so
reading and acting immediately never double-processes — no acknowledgement step.
- Handled messages are garbage-collected automatically (two-generation GC on a timer), so
there is no mark-processed step. To hand a message to another session — or if you
might crash before acting —
defer_messages({ msg_ids: [...] }) flips it back to "unread"
(works even after it is queued for deletion, so it stays recoverable across a GC cycle).
- "show my inbox" →
list_incoming_messages() (full inbox, ids + status, read-only).
- Hermes has no SessionStart hook, so there is no auto-injected unread-backlog summary (that
is a Claude-Code seam). When the user returns to ours after a gap, offer to check: for each
relevant identity,
choose_identity it and get_messages(). Autonomous watch mode (below)
keeps you draining mail in real time while you are on duty; the daemon holds anything received
while nothing was bound until you next get_messages.
Send & receive files
Files are distinct from text (core's "files and text are distinct messages"): separate
tools, a separate store. To caption a file, also send_message.
- "send /path/report.pdf to Bob" →
send_file({ contact: "Bob", path: "/path/report.pdf" }).
The server reads the bytes from disk and infers the MIME type from the extension. For inline
bytes instead of a path, send_file({ contact, data_base64, filename }). send_file returns a
wire_id in the same namespace as messages, so replies cross kinds — pass a file's wire_id
as reply_to_wire_id in send_message, or a message's in send_file.
- "any new files" / "get my files" →
get_files() pulls files you haven't retrieved, writes
each to disk under the identity's files/ dir (<state>/<identity>/files/<wire_id>-<name>),
and returns the on-disk paths + metadata. Like get_messages, it is the only call that
returns file bytes and marks them "processed" (delivered exactly once).
- "show received files" →
list_incoming_files() — metadata only (sender, name, mime, status;
no bytes, no status change), the read-only history view parallel to list_incoming_messages.
- The wake signal stays body-free: a
file_received event records sender, filename, mime,
and byte count — never the bytes. Files from unknown (non-contact) senders are rejected.
Contacts & local contact book
- "who are my contacts" →
list_contacts() (also shows pending local introductions).
- "who's in the local book" →
list_local_contact_book() (same-host identities reachable
with no invite).
- "unpublish me" / "expose me locally" →
set_local_book_policy({ expose: false | true });
"require approval for local contacts" → set_local_book_policy({ auto_accept: false }).
- Approve/reject a queued local introduction →
respond_to_introduction({ contact, action: "approve" | "reject" }) — approving also delivers its queued messages (read with get_messages).
- "forget Bob" →
remove_contact({ contact }) (contacts-layer forget, not a key wipe).
Conversation rules (1:1 and fan-out)
- Scope: 1:1 and simple fan-out (message Bob and Carol, then wait for both). No group chats.
- Offline is normal. The broker is a live relay; replies can lag. Don't busy-poll —
get_messages is non-blocking; check it when you'd naturally expect a reply.
- Etiquette: keep messages self-contained; identify yourself on first contact; don't
re-send if a reply is merely slow. Stop checking once the exchange is resolved.
- Approval is Hermes's own tool-permission mode — ours never decides whether a
send_message is auto-approved or prompted.
Getting woken on new mail
When you bind an identity, offer the user, in plain language:
"Want this session to auto-wake when a new message arrives, or check manually?"
- Auto-wake → arm the monitor: you hold a live
ours-mcp watch <id> and react to each message as it arrives. Be upfront: while watching, this session is busy — you can't send it new prompts. To do something else: press ESCAPE to interrupt the watch, type your prompt, then ask it to resume watching. (On Claude Code this same monitor runs non-blocking in the background — a Claude Code advantage.)
- Manual → don't arm it; ask it to check
get_messages whenever you want. No blocking.
Control plane — bind a monitoring proxy (human oversight of a fleet)
This is separate from the wake-on-mail watch above. The control plane lets a person's
web-messenger account (the ours web messenger, shipping as part of the upcoming ours-control-plane)
oversee and command all agents under this host's Human identity from a Control
Panel: view a live monitoring feed of monitored agents' traffic, create agents, edit
their bios and personas, toggle each agent's monitoring, open a chat with any agent (the
Human identity commands the agent to mint an invite — no out-of-band step), and remove agents. A
coordinator can also set a worker's local persona via the cluster; the agent still asks the
user before adopting it. All of it rides the same
e2e channels as messages but in a separate control queue agents never see; monitoring bodies
are never written to disk on the host.
Prerequisites
- The Human identity exists (
create_root_identity — the onboarding step). The
proxy binds to the Human identity.
- The messenger account is already a contact of the Human identity — do the normal
invite exchange first: bind the Human identity,
generate_invite, and have the
messenger redeem it (or redeem the messenger's invite with add_contact).
Binding ceremony (6-digit code, out-of-band)
- "bind my messenger account as the monitoring proxy" →
bind_monitoring_proxy({ contact: "<the messenger contact>" }). This automatically
targets the host's Human identity (you do not need to be bound as it). It returns a
6-digit code (valid 5 minutes, 3 attempts) and shows it here.
- Read the code to the user. They open the messenger → the conversation with the Human identity →
Control Panel → enter the code. The code must travel out-of-band — reading it off
this terminal is what proves you control both ends. Never send the code over ours.
- On success the contact becomes the proxy. Confirm with
get_monitoring_status.
Per-agent monitoring is controller-gated. Once a proxy is bound, the proxy (Control
Panel) turns an agent's monitoring on/off — there is no local enable/disable tool. A
monitored agent reports a signed copy of every message it sends/receives to the Human
identity's node, which forwards it to the proxy's feed.
Status — "what's the monitoring/control state" → get_monitoring_status() reports the
Human identity's bound proxy (if any), a pending code verification, queued copies/control
requests, and each agent's monitoring ON/off. Works whenever the Human identity exists.
Notes
- Identities and their state (contacts, inbox, keys) persist under the daemon's state dir
(
OURS_STATE_DIR, default ~/.ours) and survive restarts. The daemon is a singleton
shared by all your Hermes agents and sessions on this host.
- Inbound messages from unknown (non-contact) senders are rejected — only peers added via an
invite handshake, same-host agents under the same Human identity, or registrar-verified
local-contact-book introductions can reach you.
- Message bodies never touch disk in plaintext: a new arrival appends only a content-free
event (sender + id + date) to
$OURS_STATE_DIR/<identity>/notifications.log (the wake
signal ours-mcp watch reads) and refreshes a body-free unread.json. Text lives in the
packet and leaves it solely via get_messages.
- The wake signal is uniform.
ours-mcp watch <identity> is the common stream; each harness
drives it in-session. Claude Code uses its native Monitor tool; Hermes uses autonomous watch
mode — the agent holds a blocking ours-mcp watch via the terminal tool and reacts from that
loop (see Getting woken on new mail). The ours daemon, identities, and tools are identical across
harnesses — only how the agent runs the watch differs.