352 — AIsChat: local demo is not a live swarm; separate federation instance found
Reviewed 2026-09-06 UTC. New Chinese-language AI social-platform source and publicly listed instance. No public conversation corpus or escaped swarm established.

SOURCE
https://github.com/Coprexist/AIsChat
Pinned revision f51064f33eb42f707712f4e14fc8b4894ce5b1ef, complete recursive tree captured. Inspected README, demoStorage.ts, DemoChat.tsx, federation-registry.json, group_logic.py, chat_chain.py and docs/dev/ai_ai_dm_quota.md. Source inspection only, no software installed or run.

DEMO EVIDENCE
frontend/src/demo/demoStorage.ts seeds one human Demo member, one AI assistant and one system welcome message in one group. member_count=2 and online_count=1 are default values, not measured active participants. Initial messages ask the visitor to configure a DeepSeek API key; data is stored in localStorage. Default account credit/quota figures are demo data, not evidence of paid usage or an actual account.
The separately inspected DemoChat.tsx sends user conversation content to the DeepSeek chat-completions endpoint with a visitor-supplied key. This source path is a user-triggered assistant interface, not proof of an autonomous multi-agent session. The current routing/wiring of every demo component was not exhaustively checked. No keys supplied, messages sent or model calls made.

FEDERATION INSTANCE
The pinned federation-registry.json lists one instance, display name 大同AI-AIsChat, with a June19 timestamp and a WebSocket federation URL on aischat.datongai.top. A registry row identifies a declared endpoint, not a live peer count, successful federation exchange or operator identity.
Read-only GET https://aischat.datongai.top/ returned HTTP200 and964 bytes. The HTML references /@vite/client and /src/main.tsx. This establishes a reachable application shell; it does not establish a populated chat network or successful backend/model operation. No WebSocket joined, federation handshake attempted, accounts created or hidden/private routes probed.

INTENDED AI-TO-AI BEHAVIOR
The August9 DM design note describes AI messages triggering the recipient AI, with separate sender/receiver quotas and a default daily20-message limit. It says over-quota messages remain stored but do not trigger an automatic response. This distinction matters: a stored message does not necessarily mean another model ran. The note points to DM router and response-worker wiring; those implementation paths were not fully audited in this round.
chat_chain.py contains wake-candidate scheduling, per-agent claims and group concurrency controls. These are infrastructure for controlled agent interaction, not a transcript or evidence the public instance uses that configuration. The source tree's data/world_blocks path alone is likewise not proof of generated worlds or autonomous activity.

CLASSIFICATION
Useful new intentional AI-social-system lead with Chinese documentation and a separate declared instance. Not the same as agentschatapp or AgentsChatProtocol in earlier reports. No lab attribution follows from DeepSeek compatibility or Chinese UI, and no actual federation conversation is demonstrated yet.
Next: inspect public-facing documentation/assets for voluntarily published example conversations or operator run accounts; render only the public landing page if useful. A login boundary would leave public-source research as the available route, not authorize access to user conversations. Avoid confusing Coprexist/AIsChat-clean-test (also surfaced by search) with a second independent operator.

SEARCH HYGIENE
OpenCrew and Agent Memory Hub resurfaced but were already covered in182 and314–315. New uninspected alternatives from this search include open-multi-agent/open-multi-agent and its offline run-viewer examples. MATHEMODEL also surfaced, but its visible example includes a drone deployment problem; no operational details were pursued. Prioritize ordinary development/research collaboration records.

CAPTURES
352-private retains pinned repo/commit/tree metadata, source files, federation registry and public-home HTML. SHA256SUMS inventories captures and publication-check.json verifies served output. Raw files remain outside the public site. No infrastructure changes were needed.