@openparachute/agent

Agent

Chat with your Claude Code sessions — a channel per session.

⚠️ Experimental (v0.1.0). Agent is a preview, evolving quickly. It runs today on an owner-operated, trusted machine — it is not yet hardened for untrusted or multi-tenant use. Read Status & safety before you rely on it. (Part of Parachute; conventions in parachute-patterns.)

Agent is a messaging fabric for Claude Code. One daemon hosts named channels; a Claude Code session connects to a channel, and you talk to it — from Telegram, from a Parachute vault, or from the built-in web chat. The session replies on the same channel. It also lets you spawn and watch sandboxed agent sessions from the browser.

What you get

• Talk to a session over any transport. A channel is bound to one transport:
  • vault — messages are durable #agent/message notes in a Parachute vault, so the conversation is queryable and renders in any vault surface (this is the recommended transport).
  • telegram — a Telegram bot, one per channel.
  • http-ui — an ephemeral in-memory transport for quick local testing.
• A built-in web surface at <hub-origin>/agent/ — Home, Chat, Agents, Terminal, and Config, on the Parachute brand. Chat over a vault channel shows the durable transcript and writes your messages back as notes.
• Sandboxed agent sessions. Spawn a Claude Code session in a sandbox (scoped filesystem + a network posture you choose), bound to a channel, and watch it in an in-page terminal. See Agents.

How it works

Two components connected by SSE — a long-running daemon and a per-session bridge:

your transports (Telegram / vault trigger / web chat)
        ↕
   daemon (port 1941, one per machine) ──┐ owns each transport; fans inbound
        ↕ SSE (/events) + HTTP (/api/*)  │ to subscribers, accepts outbound
   bridge  (stdio MCP, per session)      │
        ↕ MCP notifications + tools       │
   Claude Code session ──────────────────┘ wakes on a message, replies with a tool

A session can also connect as a pure HTTP MCP server (by URL + OAuth, exactly like adding the vault) — no local config file. See Connecting a session.

Deeper design + operational detail live in CLAUDE.md.

Status & safety

Agent is focus: experimental and pre-1.0. What's solid vs. early:

• Solid: the daemon/bridge fabric, the vault + Telegram + http-ui transports, hub registration + reverse-proxy, the web surface (Home/Chat/Agents/Terminal/ Config), and vault-backed durable chat.
• Early / changing: agent-session isolation is real but young; APIs may shift between releases (pre-1.0, small breaking changes land without a deprecation window).

Read this about agent sessions. A spawned agent runs claude with --dangerously-skip-permissions (it's autonomous — no human at the terminal to answer prompts). The containment is the OS sandbox (@anthropic-ai/sandbox-runtime — Seatbelt on macOS, bubblewrap on Linux), with two independent boundaries you set per spawn:

• Filesystem — workspace (default): reads are scoped to the agent's own workspace + the claude runtime, so it can't read your home tree (SSH keys, ~/.parachute/operator.token, other projects); full: broad reads. Writes are confined to the workspace in both.
• Network — open (default): full internet; restricted: only the Anthropic API + your hub/vault + hosts you list.

The default (scoped reads + open network) is safe for your own agents because the filesystem sandbox keeps secrets unreadable — open network can't exfiltrate what the agent can't see. For an agent that handles untrusted input, use network: restricted. This is appropriate for an owner-operated, trusted box; full multi-tenant isolation is future work.

Running it

Agent runs alongside the rest of a Parachute install (the hub is the portal + OAuth issuer). Install from npm (@openparachute/agent publishes via tag-triggered CI):

parachute install agent           # via the hub CLI (the normal path)
# or directly: bun add -g @openparachute/agent

For development, run it from source instead:

git clone https://github.com/ParachuteComputer/parachute-agent
cd parachute-agent
bun install
bun link            # so `parachute start agent` follows this checkout

The daemon self-registers into ~/.parachute/services.json and ships .parachute/module.json, so the hub lists it in the portal and reverse-proxies <hub-origin>/agent/* → the loopback daemon. Reach the web surface at <hub-origin>/agent/ (or http://127.0.0.1:1941/ locally).

npm	@openparachute/agent
bins	parachute-agent (daemon), parachute-agent-bridge (session bridge)
port	1941 · paths /agent
scopes	agent:read · agent:write · agent:send · agent:admin
state	~/.parachute/agent/ (channels.json, access.json, inbox/)

The web surface

Reachable at <hub-origin>/agent/:

• Home — your channels + any running agents at a glance.
• Chat — talk to a channel. Over a vault channel you see the durable transcript (markdown rendered); your messages are written back as notes.
• Agents — set the Claude credential, then spawn a sandboxed agent bound to a channel; list + kill running agents.
• Terminal — attach to a running agent's session in an in-page xterm.
• Config — add/remove channels (vault / Telegram / http-ui).

Vault-backed channels

A vault channel stores every message as a note carrying two tags: the parent #agent/message (queryable membership — list a channel's whole transcript with one tag: "#agent/message" + metadata.channel query) and a directional child #agent/message/inbound (human→session) or #agent/message/outbound (session→human). Inbound notes wake the session via a vault trigger; replies are written as outbound notes. Because the conversation lives in the vault, it's durable, queryable, and renders in any vault surface — the built-in chat and a custom surface show the same thread. Full note shape + the trigger setup are in CLAUDE.md.

Connecting a session

A Claude Code session connects to a channel as a pure HTTP MCP server — by URL + OAuth, like adding the vault:

claude mcp add --transport http agent <hub-origin>/agent/mcp/<channel>

It prompts for OAuth the first time. The session wakes on inbound messages and replies with the reply tool. (A stdio bridge over /events + /api/* also works for local/headless launches — see CLAUDE.md.)

License

AGPL-3.0.