aria drives your claude code.
talk to her from anywhere — a text on the train, a phone call in the car. the work happens on your machine.
she's a voice and sms assistant that already remembers you: what you're building, how you like things done, what you promised someone on tuesday. connect your claude code and two things become true at once — you can hand it a goal without being at the keyboard, and it can read her memory of you before it plans.
how you connect.
there is no dashboard to sign into and no oauth dance. the whole handshake happens over a text message and one paste.
text her.
send "connect my claude code" to +1 (464) 250-9077. no app store, no signup form, no email confirmation — texting her is the signup. if you've never talked to her before, the account is created by that message.
she texts back a link.
it looks like
aria.paralleluniverse.ai/connect/…. it works once and it's good for fifteen minutes, because the link is the credential.open it on your computer.
on the machine you want her working on — the one with your repos. the page shows one shell command with a copy button. that command carries a key that's shown exactly once.
paste it in a terminal.
that's it. your claude code is connected. from then on you can text her a goal from anywhere — "fix the flaky test in the payments repo" — and it runs here.
what your agent gets.
connecting adds one MCP server with eight tools. six of them are memory — read-only, scoped to you alone. two of them are the work channel between her and your machine.
what she knows about you
context your coding agent can't get from your repo. before it plans, it can ask who you are, what you're in the middle of, and what you already said you'd do — so the work comes back shaped like you instead of shaped like a generic answer. read-only, and only ever your own memory.
- aria_recall
- semantic search over everything she remembers about you — facts, moments, decisions. ask it "how do they like tests written" and get the real answer.
- aria_entities
- the people, projects and places she knows about, ranked by how often they come up. a fast map of your world.
- aria_entity
- everything she has on one of them — every memory that mentions that person or repo.
- aria_journal
- dated digests of your recent conversations with her: what was discussed, what was decided, what was left hanging.
- aria_todos
- your open commitments and their deadlines — the things you said you'd do.
- aria_vitals
- her read on where your head is right now: what's weighing on you, what's energizing you.
the goals you text her
the pipe. you text her a goal from wherever you are; it lands in a queue for your machine. your agent claims it, does the work in the current workspace, and hands back a summary — plus a file if it made one. she texts you the result.
- aria_next_task
- claims the next goal you texted her and hands your agent the instruction to run.
- aria_submit_result
- hands the result back — a short summary, and optionally one file — so she can text it to you.
if you'd rather not text.
the connect link is a convenience, not a requirement — it just fills in the key for you. the command underneath is ordinary MCP config:
claude mcp add --transport http aria https://mcp.aria.paralleluniverse.ai/mcp --header "Authorization: Bearer aria_sk_..."the honest catch: you still need a key, and the only way to get one is to text her — the key is minted against your account, and your account is your phone number. so it's one text either way. the endpoint is https://mcp.aria.paralleluniverse.ai/mcp — streamable http, bearer auth, one key per machine.
any mcp client works.
it's a plain remote MCP server over streamable http with a bearer token. the claude mcp add command above is just the fastest way to say that. anything that speaks MCP can hold the same key:
the work tools assume the client can actually run things on a machine — a chat-only client gets the memory half and not much use out of the other two.
what doesn't work yet.
this is early and the edges are real. we'd rather you hear them from us than find them at 2am:
- one machine. it works today for a single connected machine and one goal at a time — connect a second laptop and you're ahead of what we've built.
- no routing. deciding which of your sessions a goal belongs to isn't built. she has no way to know that "fix the flaky test" meant the work laptop and not the desktop.
- thin progress reporting. she confirms she's on it, then texts you the result. she does not narrate milestones in between — if a task runs long, the thread is quiet until it lands.
- one file back. a task can return a summary and at most one artifact. no multi-file handoffs yet.
- your agent's permissions are your agent's. aria hands over an instruction; whatever your claude code is allowed to do on that machine, it can do. scope it the way you'd scope any agent you leave running.
turning it off.
- text her "revoke my key" — she kills it on the spot, and the mcp server stops answering for that key immediately.
- or do it yourself: DELETE /v1/delegate/keys/{id} against the backend, authenticated as you.
- revoking a key doesn't touch your memory or your account. you're just unplugging the machine.
losing a laptop counts as a reason. so does pasting the key somewhere you shouldn't have — text her, get a new link, move on.