prismal-server

Reference host for the prismal agent engine — a thin FastAPI/ASGI process that exposes prismal over REST · SSE · WebSocket (opt-in, host_ws_enabled) plus the inbound A2A endpoints, so non-Python clients and the front-end ecosystem can build against it.

prismal is a pure library (no web server by design). prismal-server is the missing process that binds a port, boots the engine once, and serves it — while reimplementing none of its logic.

Status

v0.2: usage, gateways and AG-UI. On top of the v0.1 host (streaming chat, inbound A2A, auth/tenancy seams, health, release plumbing), v0.2 adds per-tenant engine config and gateways, done.usage plus /v1/usage and /v1/budget, AG-UI 1.0 streaming, the A2A v1.0 dual stack, an opt-in WebSocket transport, a deterministic fake-LLM mode and read-only admin endpoints. Specs: specs/reference-host-bootstrap/ (v0.1) and specs/host-v0.2-usage-gateways-agui/ (v0.2). Release notes are in CHANGELOG.md.

Quickstart

uv pip install -e ".[dev]"
uvicorn prismal_server.app:app --reload
curl localhost:8000/healthz          # {"status":"ok"}
curl localhost:8000/readyz           # {"status":"ready"} once the runtime composes

Stream a chat turn (SSE):

tid=$(curl -sX POST localhost:8000/threads | python -c 'import sys,json;print(json.load(sys.stdin)["thread_id"])')
curl -N -X POST "localhost:8000/threads/$tid/messages" \
  -H 'content-type: application/json' \
  -d '{"content":"Hello!"}'

Or with Docker:

docker compose up --build
curl localhost:8000/healthz

The image bundles the engine, which pulls heavy ML deps (embeddings, vector store) — the first build is large and slow. See docs/running.md.

HTTP surface

Docs

The contract with the engine

The host calls the engine only through its public host read surface (HRS §3) plus the v0.1 entry points. Everything else it reads is a RuntimeContext method. A static import-guard test enforces an exact per-module symbol allowlist (tests/unit/test_import_guard.py); the rule and the review checklist live in CLAUDE.md.

• build_runtime(...) / build_test_runtime(...) → RuntimeContext (all ports, coordinated aclose())
• get_async_compiled_graph(...) → the graph to astream()
• A2AServerHandler + build_agent_card(...) → inbound A2A
• prismal.identity IdentityPort / OidcIdentityProvider → request → identity
• prismal.events.agui (AG-UI translator) and prismal.monitoring.get_run_summary (usage wire)

Development

uv pip install -e ".[dev]"
uv run pytest                      # unit suite (no live provider I/O)
uv run pytest -m contract          # opt-in: against a released engine
uv run ruff check . && uv run ruff format --check .
uv run mypy src

The unit suite runs fully offline against engine fakes (build_test_runtime); the contract suite boots a real build_runtime(), serves a real Agent Card, runs the v0.2 surfaces (native, AG-UI, usage, A2A v1) against the released engine in fake-LLM mode, and, when an LLM provider credential is present in the environment, round-trips a full chat turn over SSE.

Related repos

• prismal-ai/prismal — the engine (pip dependency prismal-ai, pinned >=3.22.1,<4).
• prismal-ai/prismal-sdk — thin client generated against this host's wire contract (/openapi.json + the A2A surface).

License

See LICENSE.