A personal AI crew you can just text on Signal

For people who don't write code. Everything below actually runs today; anything not built yet is marked “planned” or “designed”.

implemented built and available (not necessarily production-stable)　partial / designed works with gaps, or is specified but not fully built　planned not yet

No new app is required: message the assistant (“Sophon”) in a chat app you already use — here, Signal. A replaceable crew handles research, building and checking; anything that would speak for you in public or spend money stays a draft until you approve it.

1. Entry & communication

Problem for the spine: if every conversation and delivery path belongs to one app, switching the app means losing the relationship, archive, or both.

Signal can be the everyday door through a linked device; self-hosted web chat keeps per-contact histories as JSONL files on your disk; email adds a durable local spool, attachment intake, and mail-to-agent wakeup. If SMTP delivery fails, the mail body stays on the machine rather than falling back to a cloud service. A token-gated web page can drive the agent on your own machine from your phone, while fleet-drop moves files from another machine to Signal on your phone. The communication spec adds user-owned inboxes, addresses, trust tiers, lifecycle receipts, and a provider-neutral identity/delegated-authority design.

implemented signal-agent-pipeline, selfhost-chat, agent-mail-stack · designed / partial starshard-communication

2. Dispatch & replaceable roles

Problem for the spine: a workflow tied to one model's personality or tool set is still lock-in, even when its data is portable.

The crew uses seats, not permanent model names: the chat door you text hands the request to a dispatcher, which assigns a work packet with acceptance checks; builders do the work; a fresh verifier re-runs those checks; an independent reviewer looks for a different class of failure. Any seat can move to another model, machine, or platform. Cooperative work claims reduce duplicate work across sessions; they are leases, not authorization or exactly-once execution.

implemented, v0 experimental cross-session facts, work claims, receipts, and conflict quarantine in fleet-coordination-protocol · runnable subset agent-continuity-demo

3. One shared, user-owned memory

Problem for the spine: switching agents is painless only when decisions, preferences, task state, and receipts outlive every session.

All seats can use one searchable memory hub over HTTP or MCP. The reference server stores it in SQLite on infrastructure you control. Its nightly Mirror proposes consolidations; its assumption checker surfaces logical conflicts; its weekly digest emails what the system learned so improvement stays visible; and the continuity demo quarantines a competing belief instead of silently overwriting a held fact. Chat transcripts and mail remain separate open archives rather than being trapped inside model memory.

implemented reference-impl · architecture and safety contract: architecture-v1

4. Hard rules & approvals

Problem for the spine: replaceable models need stable limits outside their prompts; otherwise every swap changes what the system may do.

• Draft before consequence. External messages, publishing, and spending wait for human approval; a timeout is never approval.
• Receipts before irreversible actions. The communication design specifies intent and execution receipts, so the authorization can be checked before the effect and the result recorded after it. This stronger two-phase protocol is designed, not a guarantee of the FCP v0 library.
• Content-gated permission boundary. Public material must pass a mechanical clean-content scan that fails closed; failing content does not ship merely because the actor otherwise has a publish tool.
• Trust and authority stay narrow. Discovering a tool does not expose it to the system; secrets stay out of shared memory, and preemption, inheritance, and attenuation-only authority envelopes never widen authority. The identity/authority specification is designed.
• Spend has two walls. The self-hosted public chat checks a local daily budget before model calls and recommends a separate upstream account limit.

5. Verification & observability

Problem for the spine: portable data is not enough if a replacement cannot tell what really happened, whether a channel works, or where a result came from.

A completion receipt can name the artifact, its SHA-256 fingerprint, the rule profile, and evidence from a clean-context or different-machine re-check. In the running trial described by the layered-architecture note, an evidence-only Gate has four verdicts; only verified-mechanical, issued after a fresh check on a non-builder host, may be reported as “verified.” The trial blind-tests the verifier with planted faults; one missed critical fault revokes its pass authority, and a kill switch can disable the seat.

Channel health uses an end-to-end probe: a running process is only a heartbeat; successful receipt is health. For Signal, proof of send is the server's SUCCESS response, not the sending process's exit code. SPEC-0002 adds false-negative monitoring, filtered-message digests, uncertainty escalation, resend detection, and a runnable demo to catch what was silently filtered. The reviewer skills add a drift mirror, constraint-based adversarial panel, reuse check, frontend smoke gate, bounded delegation, and repository triage. Skill provenance records maker, source, version, license, and support path without marketing claims.

skills available reviewer-wheels (six wheels) · skill-provenance · not every receipt is independently verified; each repo states its proof boundary.

6. Open components & repos

Problem for the spine: a stack you cannot rebuild from its parts is not truly yours.

Take one branch or assemble several; none requires surrendering the whole data flow to one vendor. Licenses are mixed MIT / Apache-2.0; the communication whitepaper is CC-BY-SA-4.0.

These are experimental components extracted from real systems. They will have bugs; read each repository's limits before trusting it with important data or actions.