Code mirror, nothing runs here. These files are a copy of mcp/aureon_public from Aureon-OS at commit cf35ec95. They cover stages 1 and 2 of 4, and stages 3 and 4 are closed. The server needs a full Aureon-OS checkout to run, because it sits inside the mcp/ package there, so downloading this folder alone won't start anything. To run it, clone Aureon-OS and follow the README in mcp/aureon_public. No Hugging Face Space is running yet. The text below is the unchanged README from GitHub, so where it says "not deployed", "does not create a Hugging Face repository" or "the same pin as the Hugging Face Spaces", read it as written before this mirror existed. The Space header settings were removed here because this is a plain model repo.

aureon-public

This is a partial release covering stages 1 and 2 of 4. Stages 3 and 4 stay closed. That includes miner, queen, exchanges, orders, and keys. COVERAGE_PLAN.md records the counts: listed symbols 20083, public entry points 4388, reachable now 69, tested 69, refused 4319. Stage 1 reachable now is 30 and stage 1 tested now is 30. Stage 2 reachable now is 39 and stage 2 tested now is 39. 30 + 39 = 69 reachable. 30 + 39 = 69 tested.

Miner is stage 3: 15 listed, 0 public entry points, reachable 0, tested 0, refused 0. Queen is stage 4: 1348 listed, 566 public entry points, reachable 0, tested 0, refused 566. Exchanges are stage 4: 643 listed, 0 public entry points, reachable 0, tested 0, refused 0. Accounting is stage 3 with reachable 0. Trading is stage 4 with reachable 0.

The measured tests are the call sweep. It calls every entry point that was still reachable, with no arguments, and calls each stage-2 entry point again with the real arguments in stage2_arguments.py. Connections, process starts, and writes outside the sandbox are logged: the committed record says sockets blocked, processes blocked, and writes outside the sandbox blocked. A call that opens a connection, starts a process, writes outside the sandbox, or loads queen, miner, swarm_motion, inhouse_ai, ollama_config, or an HTTP client is refused. Other arguments have not been swept.

The Space entry is python -m aureon_public.space. It sets AUREON_PUBLIC_REQUIRE_PIN=1 and AUREON_OBSERVER_MODE=dry_run, and it refuses to start when AUREON_PUBLIC_SWEEP is set. The image installs requirements.txt (mcp==1.30.0,<2 and aureon pin 8aa6a2d971cc529484e2a5e261e28afafab4e4f4) and then the mcp package with --no-deps, the same runtime install as measured/aureon_public_stage1_run.sh. The image does not install test requirements and does not contain AUREON_PUBLIC_API_KEY.

AUREON_PUBLIC_OPEN=1 is set in the image, so the read-only stage-1 and stage-2 surface is served on port 7860 without a key. Set AUREON_PUBLIC_API_KEY in the Space secret settings to require Authorization: Bearer. A missing or wrong bearer then receives HTTP 401. With neither the key nor AUREON_PUBLIC_OPEN=1, the process binds to 127.0.0.1 only and refuses 0.0.0.0. The image keeps AUREON_PUBLIC_HOST=0.0.0.0 whenever it is in open mode or keyed mode. The MCP SDK turns DNS-rebinding protection on for a 127.0.0.1 bind and then answers HTTP 421 to Host: *.hf.space. A 0.0.0.0 bind accepts that Host header, which is what Hugging Face's proxy sends. The public HTTP server uses MCP SDK 1.30 stateless JSON responses (stateless_http and json_response). GET /mcp returns HTTP 405. That long-lived event stream is not used, and it does not take an in-flight call slot. AUREON_PUBLIC_MAX_INFLIGHT (default 4, the same size as the public worker pool) caps concurrent in-flight POST /mcp requests. Each tool call starts a worker process. A POST over that cap returns HTTP 503 with error busy. AUREON_PUBLIC_RATE_LIMIT (default 120) and AUREON_PUBLIC_RATE_WINDOW_S (default 60) limit requests per client. AUREON_PUBLIC_TRUSTED_PROXY_HOPS defaults to 1. When X-Forwarded-For is present and has at least that many values, the client is the value that many hops from the right. With the default, that is the last value, including a header whose only value is the client address. A fake first hop in front of that value does not get its own bucket. Zero trusted proxy hops ignores the header and uses the socket peer. The real Hugging Face header shape has to be confirmed on the private Space before this default is treated as settled. AUREON_PUBLIC_FORWARDED_DIAG=1 records hop count and whether each hop is an IPv4 address, an IPv6 address, or something else. It does not store the addresses. Open mode buckets by that address only. A bearer is the client id only when it matches AUREON_PUBLIC_API_KEY. The tracked-id table is capped by AUREON_PUBLIC_RATE_MAX_CLIENTS (default 4096) and drops expired windows. AUREON_PUBLIC_MAX_BODY_BYTES (default 1048576) counts bytes as they arrive, including a chunked body with no content-length. Every body message is counted until more_body is false or the client disconnects. Empty chunks do not stop that count, and the unread remainder is not passed through. AUREON_PUBLIC_REQUEST_TIMEOUT_S (default 30) limits a POST and also bounds that body read. A rate limit returns HTTP 429. A body over the cap returns HTTP 413. An idle or unfinished body returns HTTP 408 with error body_timeout. A POST that passes the timeout returns HTTP 504.

This scaffold is not deployed and it does not create a Hugging Face repository. The measured session was taken without Docker. .github/workflows/aureon-public-space.yml builds this image on a pull request and runs python -m aureon_public.space --check. AUREON_REF defaults to 0390194b80cddb66c8cfd052efe3aa4f9d1545d6, the merge commit of #225. That tree contains the aureon pin. Do not point AUREON_REF at cf43882e as a standing default. The first Hugging Face build log must show the AUREON_REF build arg. Uvicorn is started with proxy_headers off and an empty forwarded_allow_ips, so X-Forwarded-For is read only by this server's own middleware.

MCP lets outside models call Aureon's tools. No outside model or API becomes Aureon's reasoning. The cognition runs in Aureon's own code.

This process is a read-only FastMCP server named aureon-public. It exposes the milestone-1 tools, resources, and prompts. It does not place orders, read keys, start circulation, self-code, self-repair, or write the repository or state/.

aureon is pinned to commit 8aa6a2d971cc529484e2a5e261e28afafab4e4f4, the same pin as the Hugging Face Spaces. A local checkout run uses its own commit, not the pin. The MCP SDK requirement is mcp==1.30.0,<2 in requirements.txt and pyproject.toml. That installs 1.30.0 and excludes every 2.x release. This server imports FastMCP from mcp.server.fastmcp. The check that MCP 2 renamed that class to MCPServer is in COVERAGE_PLAN.md, with the sources. Migration to SDK v2 is a later decision.

Any public hosting, including this aureon-public Space scaffold, must set AUREON_PUBLIC_REQUIRE_PIN=1. That refuses to start unless aureon_source is kind pinned at 8aa6a2d971cc529484e2a5e261e28afafab4e4f4. Local runs of python -m aureon_public leave the variable unset and keep the visible aureon_source label. python -m aureon_public.space calls enforce_public_pin() before it binds.

The operator process is separate: python -m aureon_operator. It stays off unless AUREON_OPERATOR_MCP=1, speaks stdio only, and refuses live orders. It does not import aureon.trading or the other blocked packages.

repo_root() and aureon_source() raise FileNotFoundError when the process is not inside a checkout that contains huggingface/hnc-field/guard.py and aureon/core/hnc_field.py. This server needs that checkout. It does not invent a 'no checkout' status.

A layer counts as full depth only where its real entry points are called. Each figure is reachable public functions and methods out of the public functions and methods in the pinned docs/mcp/inventory.json.

Package Reachable Public entry points
accounting 0 0
alignment 11 24
analytics 0 167
appliance 0 7
approval 0 42
atn 0 28
autonomous 0 0
bio 0 206
bots 0 0
bots_intelligence 0 0
bridges 0 0
briefing 0 34
code_architect 0 0
cognition 1 14
command_centers 0 0
connectors 0 0
conversion 0 0
core 0 452
data_feeds 0 118
decoders 4 80
evaluation 6 10
exchanges 0 0
gates 3 12
generated 4 4
governance 0 105
grants 0 44
harmonic 0 278
identity 0 0
inhouse_ai 0 0
integrations 0 65
intelligence 0 274
launch 0 0
miner 0 0
monitors 0 158
observability 0 9
observer 24 63
operator 0 0
organism 0 1
plumber 0 271
portals 0 51
portfolio 0 0
queen 0 566
s51 1 24
saas 0 0
scanners 0 0
search 0 5
simulation 0 200
strategies 0 0
swarm 15 41
swarm_motion 0 10
trading 0 0
unified_margin_trader 0 0
utils 0 455
vault 0 195
wisdom 0 375

The extra package tools are helpers. A helper call is not that layer's cognition.

The order for opening the rest of the inventory is COVERAGE_PLAN.md. Alignment is the first package in that order.

body_call runs one inventoried function or method. body_surface reads a module or class and does not construct it. body_depth reports the same ratio as this table. body_registry names the servers. The operator tool is body_call_operator. Live orders stay refused. A checkout file is imported only when its git blob matches pin 8aa6a2d9. The plan and the call rules are in docs/mcp/REGISTRY.md.

Licence: MIT, the same licence as the repository LICENSE.

Two brain paths

hnc_reality_step takes enable_multiverse and reports that value in the output.

The class default HarmonicRealityField() sets enable_multiverse=True (multiverse on). This tool's default is that class default.

Do not mix the two committed paths in huggingface/hnc-field/replay/brain.json:

Path enable_multiverse Coherence in the file
Class default, multiverse on (multiverse_extension) true 0.857719007627168
Multiverse off (hnc) false 0.9331334813083556

Those figures are cited from that file. They are not a track record. Trade readiness is HOLD.

Which Auris

Four meanings stay separate.

  • The conjecture is the five symbolic-life markers in LambdaEngine. It is not called.
  • The nine nodes are QueenHiveMind.read_auris_nodes and AURIS_NODES. That is auris_nodes and aureon://auris/nodes.
  • The persona on QueenHiveMind is not scored.
  • The coherence evaluator is QueenHiveMind.get_auris_coherence. That is auris_coherence.

Only the nine nodes and get_auris_coherence are real functions exposed here. If any of the nine nodes has no reading, auris_coherence returns status no_data, coherence null, and the missing node names. The Auris-to-Lambda link is HOLD.

lambda_engine_step returns status unavailable when readings are missing or fewer than two. It does not fill the result with 0.

Install

From a checkout of this branch:

python3 -m venv .venv-aureon-public
source .venv-aureon-public/bin/activate
pip install -e mcp

pip install -e mcp installs this package and the pinned aureon commit. The install is editable so the server can still see huggingface/ replay files in the checkout.

Run

stdio, which is what desktop clients use:

python -m aureon_public

Optional streamable HTTP on 127.0.0.1:8077 at path /mcp:

python -m aureon_public --transport streamable-http --host 127.0.0.1 --port 8077

The HTTP URL is http://127.0.0.1:8077/mcp.

Claude Desktop

Put this in the Claude Desktop MCP config. Use the venv interpreter and the checkout path.

{
  "mcpServers": {
    "aureon-public": {
      "command": "/absolute/path/to/.venv-aureon-public/bin/python",
      "args": ["-m", "aureon_public"]
    }
  }
}

Cursor

The same stdio block goes in Cursor's MCP settings (mcp.json):

{
  "mcpServers": {
    "aureon-public": {
      "command": "/absolute/path/to/.venv-aureon-public/bin/python",
      "args": ["-m", "aureon_public"]
    }
  }
}

Any other MCP client

Any client that speaks MCP stdio can use the same command. A client that speaks streamable HTTP can use http://127.0.0.1:8077/mcp after the HTTP command above. No API key is read. No outside model becomes the reasoning: the client only receives the tool results computed in this process.

What is not on this server

Orders, exchange keys, aureon.exchanges, aureon.trading, unified_margin_trader, aureon.bots, aureon.bots_intelligence, aureon.bridges, aureon.scanners, circulation start, self-code, and self-repair are not tools. The operator Flask routes in aureon/bio/mcp_transport.py are a different process and are not widened here.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support