Integrates with Amazon Alexa+ to provide voice-driven detection of creative near-misses in jam sessions, walk-back to approach moments, and export of MIDI/audio seeds via MCP tools.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@Almost what did I just walk past?"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
“Alexa, what did I just walk past?”
Consent-based Alexa+/MCP agent that detects creative near-misses in a jam — moments you approached something new, then abandoned the thread — and walks you back with a reconstructable event graph plus exportable audio/MIDI seeds.
Not a Case Fit Desk clone, patch suggestor, Capture wrapper, or VCV AI patcher.
Amazon Build/Ship/Shape — Alexa+ track
Due: Fri Oct 23 2026 12:00pm PDT
MCP server (spec 2025-11-25+, Streamable HTTP) or Agent Skill or web-simulated Alexa+
Related MCP server: Epistemic DJ
almost/
apps/web/ Next.js Alexa+ sim + event graph viz (MCP-first, core fallback)
apps/mcp-server/ Streamable HTTP MCP (7 tools, stateful Mcp-Session-Id)
packages/core/ Deterministic scoring, types, fixture store
fixtures/sessions/ Synthetic canned jams (JSON + tiny wav placeholders)
docs/DEMO.md ≤3 min demo script
docs/FRICTION_LOG.md
docs/PRODUCT_FEEDBACK.md
scripts/smoke-mcp.mjs
scripts/demo.mjs npm run demo — MCP + web togetherMode | When | Path |
Live MCP |
| Web server actions call Streamable HTTP tools ( |
Core fallback | MCP health check fails or a tool call errors | Same server actions use in-process |
Scoring is always deterministic in @almost/core — MCP is a transport, not a second scorer. No LLM keys required for either mode today.
Per-session active jam: the MCP server is stateful Streamable HTTP. Each client gets a Mcp-Session-Id; the “active” jam pointer is keyed by that id (cleaned up on session close). Concurrent clients do not clobber each other. The web Alexa+ sim uses one-shot MCP connections and therefore always passes the jam sessionId in tool args for continuity across calls.
cd /workspace/almost
npm install
npm run build
# One command — MCP + web (production builds)
npm run demo
# → MCP http://127.0.0.1:3100/mcp (health: /health)
# → Web http://localhost:3000 (mode pill should read "live MCP")
# Ctrl+C stops bothDev (two terminals):
# Terminal A — MCP (live mode)
npm run dev:mcp
# → http://127.0.0.1:3100/mcp health: /health
# Terminal B — Web Alexa+ sim (prefers MCP, falls back to core)
npm run dev:web
# → http://localhost:3000Web-only (core fallback — MCP optional):
npm run build -w @almost/core
npm run dev:web
# Mode pill shows "core fallback" until MCP is reachableTypecheck / build everything:
npm run typecheck
npm run buildScore fixtures from CLI (no MCP):
npm run demo:scoreSmoke live MCP (Neon Corridor peaks @ 30s / 62s + concurrent-client isolation):
npm run start:mcp # terminal A (built dist)
npm run smoke:mcp # terminal BSay | Expect |
“Alexa, what did I just walk past?” | Top unmarked near-miss (peaks @ 30s and 62s in the fixture). |
“Alexa, the other one” | Switches to the other peak and re-explains. |
“Alexa, take me back” | Cue at Approach for the selected miss. |
“Alexa, export that seed” | MIDI-JSON / audio-clip-ref preview. |
npm run build
npm run start:mcp → confirm curl -s http://127.0.0.1:3100/health
npm run smoke:mcp → expect Neon Corridor @ 30s / 62s and isolation PASS
npm run demo (or MCP_PUBLIC_URL=http://127.0.0.1:3100/mcp npm run start -w @almost/web) → open http://localhost:3000
Mode pill should read live MCP; ask “what did I just walk past?” then “the other one.”
Tool | Role |
| Consent + activate fixture/demo session (per MCP session) |
| Append events, re-score |
| Unmarked near-misses (full objects + graph) |
| Fixture explanation + graph (LLM stub later) |
| Walk back to Approach |
| MIDI-JSON / audio-clip-ref seed |
| Session rollup |
Deterministic sliding-window heuristics on feature frames:
spectral flux / onset density → novelty
self-similarity → repetition
loop/stabilize → commitment
Near-miss = high novelty ramp then return to high-repetition safe pattern without a capture/export marker near the peak.
Graph: Approach → Peak → Abandon → SafeLoop
LLM is not used for scoring (explanations only, stubbed with fixtures today).
See .env.example:
LLM_API_KEY, LLM_PROVIDER — explanations only (unused this pass)
MCP_PUBLIC_URL — Streamable HTTP URL web server actions call (also NEXT_PUBLIC_MCP_URL)
MCP_PORT, MCP_HOST — MCP listen address
ALLOWED_HOSTS — comma-separated Host allowlist when exposing beyond localhost (defaults to localhost,127.0.0.1 if MCP_HOST=0.0.0.0)
AWS_* — optional Bedrock stretch (HITL before spend)
DATABASE_URL — optional; fixtures work in-memory
docs/DEMO.md — ≤3 min script
docs/FRICTION_LOG.md — friction-log bonus template + early notes
docs/PRODUCT_FEEDBACK.md — product feedback stub
Devpost submit
Public GitHub repo + license visibility
Deploy / public MCP + web URL
Demo video
AWS credits / Bedrock if used
Approve original/synthetic fixtures (no copyrighted tracks)
Optional Open Agent backup reg — do not divert Almost MVP
MIT — see LICENSE.
This server cannot be deployed
Create, co-edit, analyze, publish, and export collaborative step-sequencer sessions through MCP.
Persistent memory and cross-session learning for AI coding assistants (hosted remote MCP).
Autonomous music production for AI agents with MIDI generation, QC and provenance.
Agents-first viral-hook engine: generate, score, and remix short-form hooks over MCP.
