Agience Chorus

Operators, by domain. Chorus is the tekton standard library: seven MCP services, each holding the operators of one domain, served from one host or deployed one at a time.

A tekton is a chorus member — a condensor whose tools are organons, invoked by condensation. Each declares a module-level PERSONA in its own server.py, and src/agience_chorus/personas.py discovers them and hands the roster to the host:

Tektons are reached over the wire — HTTP and MCP — and are independently deployable, so an operator runs only the ones they need. Each constructs its own server auth and signs with its own identity, and no tekton imports another.

Running it

pip install agience-chorus                  # every tekton
pip install "agience-chorus[iris,seraph]"   # only what those two need
python -m agience_chorus.server             # every tekton on one process, :8082

MCP_HOST (default 0.0.0.0), MCP_PORT (default 8082) and LOG_LEVEL (default INFO) configure the host. uvicorn agience_chorus.server:app resolves too — the app is built at import time. CHORUS_CRYSTALS=aria,seraph boots a node holding a subset.

The import name is agience_chorus, and the tektons are subpackages of it — agience_chorus.aria, .astra, .iris, .lumen, .ophan, .sage, .seraph.

The extras select dependencies, not modules. agience-chorus[ophan] installs all seven tektons and adds Stripe; Python has no mechanism for an extra to exclude code. What it saves is real, all the same: a node running only iris and seraph otherwise installs Stripe, HuggingFace datasets, pypdf, sympy and numpy to run none of them. Whether an operator may discharge is still the capability gate's answer, not the installer's.

For a checkout, pip install -e ".[all,dev]" — which is what requirements.txt resolves to.

Running the tests

pytest -q src tests --rootdir=src

Both paths, and --rootdir=src, are load-bearing. Each tekton carries its own pyproject.toml, so without the flag pytest picks one of them as the rootdir and src/conftest.py — which wires the organons — never loads; the symptom is a phantom _op_retrieve is None rather than an error naming the cause.

pytest tests/ on its own does not work, and the way it fails is misleading: it stops at a collection error on import seraph.server, which reads like a broken test. The import is correct (seraph is its own distribution, agience-server-seraph); what is missing is the environment the invocation above sets up. Verified 2026-09-16 — under the full command that file collects and its twenty tests pass. Reach for the whole command before concluding anything from a subset that fails.

The bundles

src/agience_chorus/bundles/ holds the operator payloads, one JSON per group, built from the sources named in src/agience_chorus/seraph/bundle_spec.json. Each carries a sha256 that the mesh publishes.

They live inside the package so the wheel carries them: the runtime executes the payload, so an installed chorus without them could not discharge an operator at all.

The runtime reads the payload, not the source file. An edit to a tekton that is not rebuilt has no effect, and the published sha then disagrees with the tree. src/tests/test_bundles_match_source.py fails when they drift.

Layout

Model-free

A tekton's answer path is grounded operators over the artifact graph, so nothing imports a judgement made elsewhere that a caller cannot re-derive. Reaching a model stays legitimate as a deliberate act through a tekton that offers it, where the call is explicit and the provenance records who chose it. resolve_llm_credentials in src/agience_chorus/seraph/server.py raises and names the rule in the @mcp.tool description a caller reads.

Security issues: email connect@agience.ai rather than opening a public issue.

Dual-licensed — see LICENSE, COMMERCIAL_LICENSE.md, NOTICE and CLA.md.

Declaration of generative AI use

The author used Anthropic's Claude Opus (versions 4.8, 5, and 5.5) in the preparation of this work. Its contribution was to write code, and to generate and validate content. The ideas, the construction and the claims are the author's. No other generative AI tool was used. The author reviewed and edited all output and takes full responsibility for the content of this publication.