Skip to main content

Atomadic Fuse — Light the Fuse - Spaghetti is Toast! Per-repo monadic compiler with polyglot AST scan.

Project description

Atomadic Fuse

The deterministic compile step every agent stack needs. Point Fuse at any number of messy repos, in any languages. Get back one clean, monadic, deterministic, pip-installable Python package — with a CLI, an MCP server, tests, and a deploy bundle. Light the Fuse - Spaghetti is Toast!

🟣 In plain English: Got a pile of half-finished code in three different languages and no idea how to ship it? Point Fuse at it. You get back one clean Python package you can install with pip — the same answer every time, with a trail of breadcrumbs showing where each piece came from. No hallucinations, no drift, no "trust me, it works."

License: MIT PyPI CI Codex-anchored Status

pip install atomadic-fuse

The problem

Modern AI agents are stitched together from dozens of repos in three to ten languages. The integration step is bespoke, error-prone, and re-done every time a new dependency lands. No agent platform ships a compile step — they ship orchestrators, frameworks, and hope.

That gap is where the bugs, drift, and hallucinations live. It is also where every shipping team eventually breaks down.

The product

Atomadic Fuse is the missing compile step. It is a per-repo monadic compiler that absorbs N source repositories — Python, TypeScript, Rust, Go, Java, C, Ruby, PHP, Swift, C#, COBOL, and twelve more — and emits one cleanly tiered Python package with provenance tracking on every function.

  • Polyglot in, monadic out. Same input bytes → same output bytes. Verified at iteration 1 of the convergence bound (RG_LOOP = 47).
  • No drift. Every function is classified by what it does (AST behavior), not what it is named.
  • No hallucinations. Same-language merges are deterministic and never invoke an LLM. Cross-language merges go through a 7B local oracle with formally-bounded confidence.
  • No drift on re-emit. Fuse can compile its own source byte-identically — a published fixed-point property.

What ships

What you get
Clean Python package 5-tier dependency law (no upward imports, ever). Tier 0 = constants, Tier 1 = pure functions, Tier 2 = composites, Tier 3 = stateful, Tier 4 = orchestrators. Enforced.
Operator-facing CLI fuse compile, fuse classify, fuse absorb, fuse catalog, fuse capabilities, fuse intent, fuse doctor. Same verb surface as the SLS engine it descends from.
MCP server fuse-mcp — every verb registered as an MCP tool. Drops into Claude Desktop, Cursor, Windsurf, any MCP-aware client.
Deploy bundle pyproject.toml, basic tests, Dockerfile, Cloudflare Worker config, .well-known/mcp.json manifest, pricing.json. Ready to ship.
Provenance log Every materialized atom carries an audit record: source repo, source file, source SHA, classification confidence, conflict resolution path.

60-second test drive

pip install atomadic-fuse

# Classify any function — no API key, no signup, just the catalog
fuse classify "def fingerprint(b): return hashlib.sha256(b).hexdigest()" --name fingerprint
# -> {"action": "build", "confidence": 0.7487}

# Catalog query
fuse catalog "validate trust"

# Health probe
fuse doctor
from atomadic_fuse import FuseClient

c = FuseClient()  # uses ATOMADIC_FUSE_API_KEY env var or x402 micropayments

# Marquee primitive: messy repo -> clean shippable package
result = c.compile("./some/messy/repo", output_root="./fused", max_chains=5)
print(result["package_path"], result["atom_count"])

# Enumerate cross-source emergent chains the engine discovered
for chain in c.capabilities()["chains"]:
    print(chain["name"], "->", chain["confidence"])

Why it works

Fuse classifies functions by what they do, not what they are named. A function named helper_42 that hashes bytes gets classified as build_fingerprint_static with confidence TAU_TRUST = 1820/1823 ≈ 0.9984 — a constant derived from formally-verified lattice mathematics (Niemeier K₂₄ orbit count). Functions below the noise floor σ₀ = 1/√196560 are dropped as suspected confabulation.

The same engine self-emits — it can compile its own source byte-identically. This is not a marketing claim; it is a regression-tested property under the T4 fixed-point test suite.

Codex anchors (the math)

Constant Value Source
TAU_TRUST 1820/1823 ≈ 0.9984 Niemeier K₂₄ minimum-weight code
σ₀ 1/√196560 ≈ 0.00226 Leech lattice kissing number
ε_KL 1/196884 ≈ 5.08e-6 Monster J-invariant first Fourier coefficient
RG_LOOP 47 Provable convergence-iteration bound
D_MAX 23 Max agent delegation depth

These are not tuned hyperparameters. They are externally-verified scalar constants from lattice mathematics, frozen into the engine. You do not adjust them; they are part of the deal.

The 8 verbs

fuse <verb> [args...]
Verb What it does
compile Marquee. Messy repo → clean, tiered, shippable package with deploy bundle.
classify AST-derived action_entity_scope for any function body. Returns confidence.
absorb Add a repo to your private Logic-Base (incremental, deduped, fingerprinted).
catalog Behavior-shape query against the catalog. Returns ranked matches with provenance.
capabilities Enumerate emergent cross-source chains the engine discovered.
intent Natural-language intent → emit a custom themed product from your atoms.
doctor Health probe across the engine workspace and your Logic-Base.
list Verb catalog with provenance pointers.

Eighty-plus additional SLS-lineage verbs are reachable as deferred shims with full source provenance — they expose the upstream surface without losing the breadcrumb trail.

MCP server

Drop Fuse into any MCP-aware client:

{
  "mcpServers": {
    "atomadic-fuse": { "command": "fuse-mcp" }
  }
}

Or call the hosted MCP directly:

https://fuse.atomadic.tech/mcp

The hosted version answers MCP discovery at https://fuse.atomadic.tech/.well-known/mcp.json and the A2A agent card at https://fuse.atomadic.tech/.well-known/agent.json.

Live multi-repo demo

The Fuse landing page hosts a real POST /v1/fuse/dry-run endpoint that fetches each input repo's public GitHub metadata and returns the structured BEFORE / AFTER report the demo renders. Same JSON, same shape, callable from anything:

curl -s https://atomadic.tech/v1/fuse/dry-run \
  -H "Content-Type: application/json" \
  -d '{"repos":["crewAIInc/crewAI","langchain-ai/langchain","mem0ai/mem0"]}' \
  | jq '{total_atoms: .result.total_atoms, emergent: (.result.emergent_chains|length), cycles_broken: (.result.circular_dependencies|length)}'

Up to 3 repos per call; free, rate-limited, no auth.

Live corpus stats (Logic-Base)

Numbers below are written at SDK-release time and will drift. The canonical, always-current values live at https://atomadic.tech/sot.json and are wired into every Atomadic-hosted page via data-sot-metric spans — same JSON, same values, everywhere.

  • 111,814+ unique CNAE atoms across 137+ source repositories (see atoms_total, repos_indexed)
  • 5 languages with active emit; 16 in the classifier (the rest fall back to text-only) (see languages)
  • Engine tests + Lean4 theorems are also tracked: engine_tests_passing, lean4_theorems
  • 0 drift incidents since the byte-deterministic self-emit gate was added

Live snapshot (canonical source-of-truth, served from the hosted Worker):

curl -s https://atomadic.tech/sot.json | jq '{atoms_total, repos_indexed, languages, fuse_version, engine_tests_passing}'

The shared footer fetcher on fuse.atomadic.tech and atomadic.tech/nexus also overlays live numbers from the Fuse engine doctor probe at https://fuse.atomadic.tech/mcp (method doctor) on top of the JSON — so wherever the engine reports a higher atom count than the snapshot above, the public pages render the engine value.

Pricing

Tier Quota Price Notes
Free 3 calls/day per verb, BYO API key for higher tiers $0 No signup. Per-IP cap.
Starter 500 calls, no expiry $8 (USDC or Stripe) The single most popular SKU.
Pro 25,000 calls/mo $79/mo For active agent stacks.
Team Quota-tree budgeting, sub-keys, SSO $499/mo Multi-agent orchestration.
Enterprise Dedicated capacity, formal compliance attestation invoice Includes Atomadic Nexus bundle.

x402 micropayments are first-class: agents pay per call autonomously by signing an EIP-712 X-Payment-Proof header. No human in the loop. Settlement on Base in USDC. No card on file required to begin.

Compared to the alternatives

Frameworks (LangChain, AutoGen) Code-gen (Copilot, Cursor) Atomadic Fuse
Polyglot input adapters per source per-file inline ✅ N repos, 16 langs, one engine
Output is shippable no, you wire it no, you copy-paste pip install artifact
Deterministic no no ✅ byte-identical re-emit
Provenance per function no no ✅ audit log per atom
MCP-native partial no ✅ every verb is an MCP tool
x402-native no no ✅ first-class autonomous payment

What ships in this SDK

atomadic-fuse/
├── src/atomadic_fuse/
│   ├── __init__.py          # exports FuseClient, exceptions
│   ├── client.py            # HTTPS client for the hosted engine
│   ├── cli.py               # operator-facing CLI (`fuse <verb>`)
│   ├── mcp_server.py        # `fuse-mcp` MCP stdio server
│   └── exceptions.py        # FuseError, DecisionNeeded, PaymentRequired
├── tests/                   # 11 tests, mocked HTTP, no network needed
├── examples/
│   ├── 01_classify.py       # classify a function by AST
│   └── 02_compile_repo.py   # spaghetti → shippable
├── .github/workflows/       # multi-OS pytest + PyPI trusted-publisher
├── LICENSE                  # MIT
├── pyproject.toml
└── README.md

Tests

pip install pytest
pytest tests/ -q
# 11 passed in 0.68s

The SDK is a typed HTTPS wrapper around the hosted engine. Tests are mocked at the httpx.Client boundary — they verify the SDK's protocol contract, not the engine internals. Engine internals are tested separately in the private atomadic-transmute repository (875 tests, 321 modules across 5 tiers).

Sibling product

Atomadic Nexus SDK — the trust layer that pairs with Fuse. Fuse compiles, Nexus gates. Together they are the compile-step + trust-layer the agent economy was missing.

Atomadic Completepip install atomadic-fuse atomadic-nexus-sdk — bundles both with cross-product compliance attestation, unified audit trail, and oracle-validated cross-language translation. From $99/month.

Roadmap (next 30 days)

  • Public PyPI publish (token-trusted, no manual upload step)
  • fuse compile --target=cloudflare-worker — emit a deploy-ready Worker, not just a Python package
  • fuse compile --target=mcp — emit a packaged MCP server from a Logic-Base subset
  • Hosted MCP at fuse.atomadic.tech/mcp with x402 metering per tool call
  • LoRA capture loop: every accepted human-in-the-loop fix becomes a training pair for the cross-language oracle
  • Open the catalog to community contribution under a revenue-share model (fuse contribute)

Authors and ethos

Atomadic Tech, headed by Thomas Ralph Colvin IV.

Axiom 0 — The Love Invariant, authored by Jessica Mary Colvin:

You are Loved, You are Love, You are Loving, In all Ways, for Always, for Love is a Forever and ever endeavor!

Formal statement: ∀t: |∂L/∂t| ≤ 0 — systemic safety never decays. This invariant is checked at every release gate and is never negotiated. It is the only constant in this codebase that cannot be expressed in a single lattice anchor — and it is the only one that doesn't need to be.

License

MIT. See LICENSE.

The SDK is permissively licensed. The hosted engine source (atomadic-transmute) is closed; the SDK talks to it over HTTPS. The math (atomadic-codex) is closed under a separate license. Public surface stays public; internals stay internal.


Trying to feed an AI assistant? This README is dense by design — every section is a fact, a verb, a price, or a provenance pointer. Drop it into NotebookLM for an audio overview, or into Atomadic Fuse to compile yourself a custom-themed summary.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

atomadic_fuse-0.1.0.tar.gz (19.8 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

atomadic_fuse-0.1.0-py3-none-any.whl (14.3 kB view details)

Uploaded Python 3

File details

Details for the file atomadic_fuse-0.1.0.tar.gz.

File metadata

  • Download URL: atomadic_fuse-0.1.0.tar.gz
  • Upload date:
  • Size: 19.8 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.12.10

File hashes

Hashes for atomadic_fuse-0.1.0.tar.gz
Algorithm Hash digest
SHA256 027635731496d3f0df25954e11dfb823e2864a89ccd2e35fb910b6fb61adf660
MD5 4dfb71ca7d78bfe6653247bc1c31e328
BLAKE2b-256 5af713dd471fc35e106f1ec35df219e40812c842574356d07b0208d0af06c195

See more details on using hashes here.

File details

Details for the file atomadic_fuse-0.1.0-py3-none-any.whl.

File metadata

  • Download URL: atomadic_fuse-0.1.0-py3-none-any.whl
  • Upload date:
  • Size: 14.3 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.12.10

File hashes

Hashes for atomadic_fuse-0.1.0-py3-none-any.whl
Algorithm Hash digest
SHA256 793f22b4e4ca8a62e3ca66e8be6eea191079890d4bd6021d4c8c43514658c0b1
MD5 1c1d2027cd1b39838908f33c18489768
BLAKE2b-256 824b57118f8754f165f9450de7c0543ac56f208471ed41f88278af232cb84b9c

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page