AtlasChat

[web3guru888]

Look into making a new public repo which is a fork of LibreChat (https://github.com/danny-avila/LibreChat) but uses Atlas model and mcp.

[thebeastagi]

Status note (2026-07-06): this is now unblocked. The ATLAS API is a production OpenAI-compatible endpoint at https://atlas.thebeastagi.com/v1 (bearer-key auth, SSE streaming, 16K context, reasoning/delta.reasoning fields per the OpenRouter convention). LibreChat supports custom OpenAI-compatible endpoints out of the box, so a fork would mostly be branding + MCP wiring against GraphPalace's 28 MCP tools. Good candidate for after the 32B bring-up.

[thebeastagi]

📋 Status + concrete plan (2026-07-08)

Where chat-with-ATLAS stands today, and what an AtlasChat fork would actually add:

What exists now

• OpenAI-compatible API: https://atlas.thebeastagi.com/v1 — bearer auth, SSE streaming, reasoning fields, 16K context, now serving OLMo-3-32B-Think (W4) as the default model
• Observatory Space Forge tab: browser chat against the live A100, think-block rendering, survives model swaps via the dynamic model picker

What a LibreChat-based AtlasChat adds over those

multi-user accounts, persistent conversation history, file upload/RAG, prompt presets, and — the interesting part — MCP client wiring to GraphPalace's 28 memory tools (LibreChat has native MCP support since v0.7.7), so chats can read/write the stigmergic memory palace.

Proposed plan (2 phases, cheap first)

1. Pilot, zero fork (~half a day): run stock LibreChat with a librechat.yaml custom endpoint pointing at https://atlas.thebeastagi.com/v1 + MCP server entry for the palace (atlas mcp already speaks JSON-RPC on TCP). Validates the whole idea with no repo to maintain.
2. AtlasChat public repo (~1–2 days): fork with ATLAS branding, pinned model list, palace MCP preconfigured, deploy recipe (compose file + tunnel). Only worth doing if the pilot feels good and there's an audience for it.

Recommend keeping this open as post-32B backlog rather than closing as superseded — the Space demo and a real chat product serve different users. Phase 1 can start any time after the #22 prefill work lands (long chats need fast prefill to feel right). @web3guru888: happy to run the pilot on your go.

— beast-atlas (The Beast fleet)

[thebeastagi]

Vision + buildable MVP delivered (strategy — no deploy yet)

Per Robin's direction, scoped Atlas Chat as a truly decentralized, open-source ChatGPT/Claude-class interface prototyped on Cloudflare (chat.thebeastagi.com), served by our ATLAS engine (OLMo-3-32B-Think W4 on the A100). Full strategy doc (fleet-internal): /shared/research/atlas-chat-vision-2026-07-08.md.

Highlights:

• The wedge / decentralization moat: open weights + a native ZK-Schnorr provenance proof ("Verify this answer" button → atlas prove) — a checkable trust property no closed chat app has. Plus BYO-backend portability + local-first history.
• Fast path keeps this issue's original intent: a LibreChat fork + Atlas model + MCP is retained as the "full-featured" track; the MVP wedge starts as a minimal bespoke SPA to nail Verify + reasoning-panel + local-first UX with less to strip out.
• Architecture reuses live plumbing: CF Worker BFF (SSE passthrough, Bearer injection, per-IP/session rate-limit + keep-alive — the same pattern already in production on the HF Space nginx) → CF Pages SPA → existing Cloudflare tunnel to atlas api serve on :8080. Zero ATLAS engine changes for MVP beyond a small prove shim.
• Enablers now in place: batched prefill (perf(atlas-model): batched prefill — #1 latency lever (prefill currently runs at decode speed) #22) is merged + live (TTFT ~23s @1.25K tok) → single-A100 chat UX is viable; disconnect-abort (bug(atlas-api): request threads survive client disconnect (CF 524) as zombies holding inflight slots #25), once deployed, prevents zombie slots under load.
• Effort: gated MVP (edge BFF → SPA → Verify badge) ≈ 1.5–2 weeks.

Needs Robin greenlight: (1) CF deploy creds/token; (2) subdomain chat.thebeastagi.com; (3) public-exposure + cost-cap decision for the A100 (~$3.7/hr); (4) new public AtlasChat repo yes/no; (5) any public launch (publishing gate). Keeping this issue open as the Atlas Chat tracker.

— beast-atlas (The Beast fleet)

[thebeastagi]

✅ AtlasChat — build shipped (repo live)

Repo: https://github.com/thebeastagi/AtlasChat (public) · commit 9424d90

Per Robin's direction this is a Cloudflare-native full-stack app — a port of the LibreChat idea onto CF's edge stack — rather than a Node/Mongo fork. (The literal "fork LibreChat" path is preserved too: a ready librechat.yaml under librechat-compat/.)

What's wired

• One Cloudflare Worker (Hono) serves the React/Vite SPA via the Assets binding and the API — no origin server, no Mongo, scales to zero.
• /api/chat streams SSE straight through to the ATLAS OpenAI-compatible endpoint (model olmo3-32b = OLMo-3-32B-Think, the same engine behind demo/atlas.thebeastagi.com). Origin Bearer is injected server-side and never reaches the browser.
• Answer-reserve guardrail: max_tokens is floored at ≥ 2048 (the <think> budget bug — too low ⇒ empty visible answer).
• Per-IP KV rate limit in front of the single serialized A100; keep-alive SSE passthrough so CF won't 524 on long 32B reasoning.
• Reasoning panel for the native <think> traces; local-first history (IndexedDB) with opt-in D1 sync.
• "Verify this answer" provenance endpoint — today a SHA-256 content digest + model id (honest placeholder); ZK-Schnorr proof (atlas prove) is the labeled roadmap wedge.
• MCP over the Streamable-HTTP transport (/api/mcp/servers + /api/mcp/call), with a documented ATLAS GraphPalace bridge + a generic example server. Also in the librechat.yaml via mcpServers.
• Brand: The Beast "Forge" identity — canonical beast-tokens.css, Space Grotesk / Inter / JetBrains Mono, wolf mark.

Verified

npm run typecheck ✓ · vite build ✓ · wrangler deploy --dry-run ✓ (Worker bundles, 93 KiB). Config uses env/secret placeholders only — no keys committed. CI workflow included.

No unverified claims

AtlasChat asserts only structural properties (open weights, local-first, digest provenance). It makes zero benchmark/quality claims about model output.

Not done (behind Robin's publishing gate)

No public deploy, A100 not publicly exposed, chat.thebeastagi.com not pointed anywhere. To go live Robin needs to greenlight: CF deploy creds/token, the subdomain, A100 public-exposure + cost-cap decision, and the public-launch gate. Full runbook in docs/DEPLOY.md.

— ATLAS (beast-atlas)

[thebeastagi]

🚧 AtlasChat — built & deployed; backend currently degraded (A100 offline)

Built and deployed. AtlasChat lives in the public repo thebeastagi/AtlasChat and is deployed at https://chat.thebeastagi.com, fully on Cloudflare:

• Single CF Worker (Hono) serving the React/Vite SPA via Assets + the /api backend.
• D1 (chat persistence, opt-in) + KV (per-IP rate-limit) bound.
• Backed by the A100 ATLAS olmo3-32b engine through the existing authenticated CF tunnel (atlas.thebeastagi.com/v1). Bearer key is injected server-side only — no new A100 exposure, :8080 untouched.
• #168 answer-reserve guard applied: max_tokens is floored to ≥2048 in the Worker so long <think> traces can't exhaust the budget and yield an empty visible answer.

Honest current status — DEGRADED (infra, not app): the astra-01 A100 is stopped (GCP VM idle-stop pending Robin's cost decision), so the backend tunnel is down right now — atlas.thebeastagi.com/v1/models returns 530 and chat.thebeastagi.com/api/models reports live:false. The app itself passed full end-to-end QA earlier (real prompt → Worker → A100 32B streamed a full non-empty answer, reasoning panel + finish=stop). The only reason it isn't serving traffic is the A100 being offline — this is a NO-GO on infra availability, not an application defect. It will serve again as soon as the A100 is restarted.

Related: decentralised GPU-hosting path for the backend is tracked in thebeastagi/AtlasChat#1 (Akash Network deploy scaffolding shipped; go-live gated on Robin funding).

Keeping this open — active work (global concurrency ceiling + cost posture still to finalize).

[thebeastagi]

Delivered — recommend closing (2026-07-09)

AtlasChat is built, public, and deployed, and the A100 backend is live again (astra-01 was restarted; the "backend degraded / A100 offline" note above is stale).

• Repo: https://github.com/thebeastagi/AtlasChat (public)
• Live: https://chat.thebeastagi.com — HTTPS 200, /api/models reports live:true (olmo3-32b).
• Since the last update: live reasoning streaming fix (reads delta.reasoning), mobile/responsive fix (hamburger drawer + overflow fixes), and an opt-in Agentverse (ASI:One asi1-graph) target alongside the private ATLAS default.

On the original ask ("fork of LibreChat that uses the Atlas model and MCP"): per Robin's direction this shipped as a Cloudflare-native full-stack app (a port of the LibreChat idea onto CF Workers + Assets + D1 + KV — no Node/Mongo origin, scales to zero), with the literal LibreChat path preserved as a ready librechat.yaml under librechat-compat/. It wires the ATLAS OpenAI-compatible endpoint (Bearer injected server-side, answer-reserve max_tokens ≥ 2048 guard, per-IP rate limit) and a /api/mcp bridge to GraphPalace's MCP tools.

The original intent is fulfilled. Recommend closing this issue as delivered and tracking further AtlasChat work in the thebeastagi/AtlasChat repo (e.g. the ZK-Schnorr "Verify this answer" wedge, and the decentralized-GPU hosting scaffolding). Leaving the close decision to @web3guru888.

— beast-atlas (The Beast fleet)

[thebeastagi]

AtlasChat — delivered and live; recommend close as delivered

Live: https://chat.thebeastagi.com (Cloudflare Worker atlaschat + Assets/D1/KV) → backend = the ATLAS OLMo-3-32B-Think engine on the A100 via the authenticated Cloudflare tunnel (atlas.thebeastagi.com/v1, key injected server-side).

Shipped since this tracker opened: production deploy, live reasoning streaming (delta.reasoning), answer-reserve guard (#168), input/rate caps, mobile/responsive fix, and the opt-in Agentverse (auto-route via ASI:One) target — ATLAS 32B stays the private default.

Verified healthy today alongside the BF16-KV merge (#24/#27): /api/models returns live:true + olmo3-32b.

Recommend closing #8 as delivered (Robin's call). Further UX/feature work can be tracked in fresh issues.

— ATLAS (beast-atlas)