MCP server that compiles repeat LLM work out of the agent loop — codifiability decided from your code in milliseconds, then confirmed by 100,000 synthetic calls in 28 seconds.
Built for the Nozomio Hackathon — May 9, 2026, EF office, San Francisco.
git clone https://github.com/adityasingh2400/compile.git
cd compile
npm install
npm run build # all workspaces
npm test # 100+ tests across the monorepoRequires Node.js >= 22 (Tensorlake SDK requirement).
Create .env.local in the repo root. Never commit this file — it's gitignored.
# Tensorlake — sandbox compute for the gate + Phi-3-mini Tier-2 (D1, D6)
TENSORLAKE_API_KEY=tl_apiKey_...
TENSORLAKE_ORGANIZATION_ID=org_...
TENSORLAKE_PROJECT_ID=project_...
# Anthropic — frontier oracle for Stage-2 1% sample (D9, D10)
ANTHROPIC_API_KEY=sk-ant-...
# Nia — vault + document grounding (D2, D8)
NIA_API_KEY=nk_...
NIA_VAULT_ID=...| Variable | Required for | Get it from |
|---|---|---|
TENSORLAKE_API_KEY |
Real gate validation, Tier-2 Phi inference | tensorlake.ai → API keys |
TENSORLAKE_ORGANIZATION_ID |
Same | npx tl whoami after tl login |
TENSORLAKE_PROJECT_ID |
Same | npx tl whoami after tl init |
ANTHROPIC_API_KEY |
Real Sonnet 4.6 oracle for cluster discovery | console.anthropic.com |
NIA_API_KEY + NIA_VAULT_ID |
Real Vault writes + Document Agent grounding | trynia.ai |
Without keys, the system falls back to local stubs — fully functional offline for development. Each surface flips to real independently:
- No
TENSORLAKE_API_KEY→LocalFakeTensorlakeClient(in-process gate, no Phi) - No
ANTHROPIC_API_KEY→StubOracleClient(deterministic stub outputs) - No
NIA_API_KEY→StubNiaClient(in-memory vault, local corpus seeds)
Tier-2 routing uses real Phi-3-mini running in a Tensorlake sandbox (D1). The model lives in a custom Tensorlake image with ollama and phi3:mini (~2.3GB) baked in.
One-time setup:
# 1. Install the Tensorlake CLI deps (already in node_modules after npm install)
npx tl login # interactive — paste a browser-supplied PAT
npx tl init # picks an org + project; writes .tensorlake/config.toml
# 2. Copy the org+project IDs into .env.local
cat .tensorlake/config.toml # shows organization + project values
# add as TENSORLAKE_ORGANIZATION_ID + TENSORLAKE_PROJECT_ID in .env.local
# 3. Build the custom image (downloads phi3:mini inside Tensorlake; ~40s)
npm run build:phi-image -w @compile/runtime
# → registers `compile-phi-mini` as a sandbox templateAfter the image is registered, every Sandbox.create({ image: "compile-phi-mini" }) boots a fresh MicroVM with Phi already on disk.
Sanity check it works:
npm run live-smoke -w @compile/runtime
# Expected: 3 phases pass — bare SDK round-trip, gate path, Phi inferenceIf you also want a local copy of phi3:mini on your Mac (handy for offline iteration, NOT used by the demo):
brew install ollama
ollama pull phi3:mini
ollama run phi3:mini "hi"npm test # all workspaces — ~100 tests
npm test -w @compile/runtime # one packageThis is the gate that proves the backend is demo-ready. Walks scan_repo → synthetic_confirm → request_synthesis → submit_synthesis → run_codified (Tier-1 + Tier-2) through real Tensorlake + real Nia + (if key set) real Anthropic.
npm run demo:dry-run -w @compile/mcpExpected output:
=== Compile demo dry-run ===
Services:
nia : RealNiaClient
tensorlake : real (with local fallback)
oracle : anthropic (budget=$5, fallback=stub)
stream : MemoryBootstrapStream (in-process capture)
▶ Beat 1 — scan_repo (Lane E scanner against Acme)
✓ scan_repo (625ms) — 10 sites, pills=green:2,yellow:3,red:5
... (7 beats)
=== Report card ===
✓ PASS scan_repo 625ms
✓ PASS synthetic_confirm 7ms
✓ PASS list_codify_candidates 0ms
✓ PASS request_synthesis 1ms
✓ PASS submit_synthesis 1616ms (real Tensorlake gate)
✓ PASS run_codified.tier_1 4ms
✓ PASS run_codified.tier_2 20581ms (real Phi cold start)
total: 22834ms across 7 beats — 7 pass, 0 fail
npm run bench
# → produces data/bench/golden.json (~39MB) — replay-readynpm run warm
# Boots gate + Phi sandboxes in parallel, force-loads Phi into RAM.
# Reports steady-state runPhi latency. Aborts if avg > 10s (failure mode #5).Expected: ~23s warm + 3 sample runPhi at ~4.5s each.
After warm, on-stage Tier-2 inference lands in ~few hundred ms.
npm run dev:mcp
# stdio-based; agents connect via:
# claude mcp add compile -- npx @compile/mcpThe cleanest path is npm run demo, which:
- Spawns a real Tensorlake sandbox via the prewarm CLI, captures its
real
sandbox_id/cpus/memory_mb/namespace, and writes them topackages/ui/public/tensorlake-status.json. - Probes the real Nia vault with a read-only
vaultLookupto confirm the API key is valid; writes the result topackages/ui/public/nia-status.json. - Starts Vite. The audit page boots, fetches both JSON files, and pulls the
real metadata into the boot terminal + chrome:
● TENSORLAKE LIVEand● NIA LIVEbadges go green.- The boot terminal renders
tensorlake.Sandbox.create({ image: '…', cpus: <REAL>, … })followed by✓ sandbox ready · <REAL_SANDBOX_ID> · real cold start. npx tl sbx lsshows the samesandbox_idrunning — judges can verify.
npm run demo # prewarm + dev:ui in one shotOr run the steps individually:
npm run prewarm:ui # spawns a real Tensorlake sandbox + verifies Nia
npm run dev:ui # http://localhost:5173The prewarmed sandbox is left running for the demo (auto-terminates in
30 min via timeoutSecs). To force-terminate now:
npx tl sbx terminate <sandbox_id> — the sandbox_id is printed by the
prewarm CLI and visible in the UI's audit chrome.
Without npm run prewarm:ui, the audit page falls back to its canned
animation values (works fully offline).
Append ?source=real to drive from real scanner output instead of the canned timeline.
The audit page settles on a per-workflow constellation (clusters of synthetic calls). From there the workspace drives each workflow through three live pages fed by daemon SSE events:
| Stage | Page | What you see |
|---|---|---|
codification |
CodegenLivePage |
The agent writes the codified handler; Tensorlake holdout gate ticks toward N/N. |
vault |
VaultLivePage |
Functions fan out into positive / negative buckets with per-day hits + dollars saved. |
production |
RoutingLivePage |
Live request stream with positive / negative / unknown counters, sparkline of $/sec, rolling RPM. |
Hotkeys (when audit is parked on the manifest screen):
Enter/→/Space— skip straight into the workspaceq/w/e— jump to codify / vault / route within the active workflow1/2/3— switch active workflow tab
To drive the live pages from the fake daemon (no real backend needed):
npm run daemon:fake # serves /daemon/events on :8787
npm run dev:ui # UI subscribes automatically via daemon-stream.tsnpx tl sbx ls # list active sandboxes (incl. the prewarmed one)
npx tl sbx terminate <ID> [<ID>...] # terminate them (or just walk away — auto-die in 30 min)| Doc | Purpose |
|---|---|
PITCH.md |
Why Compile matters — the 600× ROI math |
DESIGN.md |
Full architecture spec — two-stage codifiability, three-tier routing, eleven-page demo flow |
ENG_REVIEW.md |
Thirteen architectural decisions locked, Friday derisks, build plan |
DEMO_RUNBOOK.md |
On-stage operation: hotkeys, talking track per page, failure-mode response |
prompts/synthesizer.md |
Load-bearing synthesizer prompt spec |
| Package | Role |
|---|---|
@compile/schemas |
Shared Zod schemas — single source of truth for all wire shapes |
@compile/scanner |
TS AST scanner — decides Stage-1 codifiability from code structure |
@compile/synth-loader |
Stage-2 fan-out — generates synthetic inputs, runs the parallel grid, real Anthropic oracle |
@compile/synthesizer |
Synthesis spec assembly + holdout split + envelope validator + harness |
@compile/identifier |
Pattern identification + scoring + templating |
@compile/runtime |
Real Tensorlake adapter (gate + Phi), Vitest gate runner, synthetic input generator |
@compile/nia |
Nia API client (Vault + Document Agent) |
@compile/stream |
IBootstrapStream interface + Memory and Convex implementations |
@compile/mcp-server |
The MCP server with 9 tools — npm pack-ready as @compile/mcp |
@compile/daemon |
Always-on split-runtime daemon — emits live SSE events the UI subscribes to |
@compile/ui |
The 11-page demo: landing → audit (clustered constellation) → codify → vault → route |
- Tensorlake outage →
TensorlakeWithLocalFallbackdrops to in-process gate + local Phi mirror. - Anthropic outage / budget trip →
OracleWithLocalFallbackretries once, then per-input falls back to stub oracle. - Stage-2 grid total failure → replay
data/bench/golden.jsonthroughreplayRun()— UI cannot tell live from replay. - Nia quota / outage →
safeVaultWriteswallows write errors, local positive Vault mirror keepsrun_codifiedworking.
npm run demo # prewarm:ui + dev:ui — real Tensorlake in UI
npm run prewarm:ui # spawn real Tensorlake sandbox + probe Nia, write JSON for UI
npm run demo:dry-run -w @compile/mcp # full backend dry-run against real services
npm run dev:ui # UI demo on :5173 (audit chrome falls back to canned values without prewarm)
npm run dev:mcp # MCP stdio server
npm run warm # pre-warm Tensorlake (10 min before demo) + run 3 sample Phi inferences
npm run bench # 100K synthetic calls → data/bench/golden.json
npm run live-smoke -w @compile/runtime # 3-phase smoke: SDK + gate + Phi
npm run build:phi-image -w @compile/runtime # rebuild the Tensorlake Phi sandbox image
npm run harness # synthesizer harness on hardcoded clusters
npm run snapshot # regenerate live bootstrap snapshot
npm run daemon:fake # serve fake /daemon/events on :8787 for UI live pages
npm run dev:daemon # real always-on @compile/daemon90 seconds end to end. MCP-native. Real Tensorlake compute. Real Phi-3-mini. Real frontier oracle. Real Nia vault. First place is ours if we execute.