test-demoprep / mcp_server.py
mikeboone's picture
feat(mcp): KEY_VAR ENV: indirection + scheme-less/port-tolerant host matching
e27ea87
Raw
History Blame Contribute Delete
30.8 kB
"""
mcp_server.py β€” Demo Wire MCP server (v1)
Exposes DemoPrep's demo-build pipeline as an MCP tool so an external agent
(AgentSpot / Spotter Assistant) can build a full ThoughtSpot demo from a prospect
*brief* without a human clicking through the Gradio UI.
This is a THIN ADAPTER. It does not reimplement any business logic β€” it drives the
exact same `ChatDemoInterface` controller the Gradio app uses, headlessly, skipping
the research phase by injecting the brief as the research output. All build logic
stays in one place (see docs/SINGLE_PIPELINE.md and CLAUDE.md).
v1 scope (simplest thing that works):
* ONE blocking tool β€” build_demo_from_brief(...) runs the whole pipeline
(blueprint -> dataset -> DDL -> Snowflake -> TS model -> liveboard) and returns
the resulting IDs when done.
* Everything is fixed/server-side; per-call config (env, model, sharing) is a
later version.
* Builds run for minutes. If the caller's connection times out, the build still
finishes server-side and DemoPrep's Slack notification delivers the liveboard
URL. (Async / job-backed delivery is the planned next step.)
Server-side config (env vars):
MCP_ACCESS_TOKEN (required in http mode) shared bearer secret gating the endpoint
TS_ENV_URL_DEFAULT ThoughtSpot URL used when a build doesn't pass ts_url
TS_USER_DEFAULT owner email used when a build doesn't pass owner_email
TS_ENV_<n>_LABEL / TS_ENV_<n>_URL / TS_ENV_<n>_KEY_VAR
the environment table β€” SAME numbered triplets the app's
dropdown uses. The trusted-auth key for a build is found
by matching its ts_url host against TS_ENV_<n>_URL.
MCP_TRANSPORT "stdio" (default, local dev) | "http" (HF Space)
MCP_HTTP_PORT http port (default 7860)
Also read from env: SUPABASE_URL / SUPABASE_ANON_KEY (Snowflake creds) + ANTHROPIC_API_KEY.
Both the TS target and the owner are resolved per build: the request may pass
ts_url / owner_email, else they fall back to TS_ENV_URL_DEFAULT / TS_USER_DEFAULT.
Local smoke test (needs the project venv with requirements installed):
TS_ENV_URL_DEFAULT=https://sebe.thoughtspotstaging.cloud python mcp_server.py
(the TS_ENV_* table comes from .env; the URL default just picks which entry)
Then point an MCP client (e.g. `mcp dev mcp_server.py` / MCP Inspector) at stdio.
"""
from __future__ import annotations
import os
import sys
import time
import types
import uuid
import threading
import collections
# --- env bootstrap: order matters --------------------------------------------
# .env first so MCP_* / TS_ENV_* / Supabase creds are available.
from dotenv import load_dotenv
load_dotenv()
# Owner identity is resolved PER BUILD (mirrors the TS target): the request may
# pass owner_email; otherwise it falls back to TS_USER_DEFAULT. Nothing is
# hardcoded, and boot does not require it β€” a build fails loud if neither is set.
os.environ.setdefault("DEMOPREP_NO_AUTH", "true")
_DEFAULT_OWNER = (os.getenv("TS_USER_DEFAULT") or "").strip()
if _DEFAULT_OWNER:
# Best-effort acting user for any headless path with no per-build owner.
os.environ.setdefault("DEMOPREP_DEV_USER_EMAIL", _DEFAULT_OWNER)
def _resolve_owner(owner_email: str) -> str:
"""Owner of the objects a build creates: the per-request owner_email, else
TS_USER_DEFAULT. Fails loud if neither is set (no silent blank owner)."""
owner = (owner_email or os.getenv("TS_USER_DEFAULT") or "").strip()
if not owner:
raise RuntimeError("No owner: pass owner_email in the request or set TS_USER_DEFAULT.")
return owner
# Snowflake creds + SNOWFLAKE_DATABASE from Supabase admin settings -> os.environ.
# Must run once at startup; the deploy path reads these via get_admin_setting/env.
from supabase_client import inject_admin_settings_to_env, log_mcp_payload
inject_admin_settings_to_env()
# Heavy imports (pull gradio etc.) β€” safe headless. DEMOPREP_NO_AUTH is set above,
# which must happen BEFORE importing chat_interface.
from chat_interface import ChatDemoInterface
from demo_builder_class import DemoBuilder
from demo_personas import parse_use_case, get_use_case_config
from llm_config import DEFAULT_LLM_MODEL
def _ts_env_table() -> list[tuple[str, str, str]]:
"""Read the numbered TS_ENV_<n>_LABEL/URL/KEY_VAR triplets β€” the SAME
environment table the app's dropdown uses (single source of truth for
which ThoughtSpot instances exist and their trusted-auth keys).
KEY_VAR normally holds the trusted-auth key itself. An `ENV:<NAME>` value
dereferences another env var instead β€” for referencing a secret that
already exists under its own name (e.g. TECHPARTNERS_TA_KEY on the Space)
without copying its value into the table."""
table = []
i = 1
while True:
label = (os.getenv(f"TS_ENV_{i}_LABEL") or "").strip()
if not label:
break
url = (os.getenv(f"TS_ENV_{i}_URL") or "").strip().rstrip("/")
key = (os.getenv(f"TS_ENV_{i}_KEY_VAR") or "").strip()
if key.startswith("ENV:"):
key = (os.getenv(key[4:].strip()) or "").strip()
table.append((label, url, key))
i += 1
return table
def _host_of(url: str) -> str:
"""Hostname of a URL; tolerates scheme-less input and strips any port."""
if "://" in url:
url = url.split("://", 1)[1]
return url.split("/", 1)[0].split(":", 1)[0].lower()
def _resolve_ts_target(ts_url: str) -> tuple[str, str]:
"""Resolve the (url, trusted_auth_key) for a build.
The URL comes from the build request; if absent it falls back to
TS_ENV_URL_DEFAULT. The trusted-auth key comes from the TS_ENV_<n>_* table
entry whose URL host matches. Fails loud on a missing URL, an unknown
instance, or a blank key (no silent blanks). Resolved per build (not at
boot) so the server stays up regardless of which environment a given
build targets.
"""
url = (ts_url or os.getenv("TS_ENV_URL_DEFAULT") or "").strip().rstrip("/")
if not url:
raise RuntimeError(
"No ThoughtSpot URL: pass ts_url in the request or set TS_ENV_URL_DEFAULT."
)
if "://" not in url:
url = f"https://{url}"
table = _ts_env_table()
if not table:
raise RuntimeError(
"No TS_ENV_<n>_LABEL/URL/KEY_VAR entries configured β€” the MCP server "
"reads the same numbered environment table as the app."
)
host = _host_of(url)
for label, env_url, key in table:
if _host_of(env_url) == host:
if not key:
raise RuntimeError(f"TS_ENV entry {label!r} matches {url} but its KEY_VAR is blank.")
return url, key
known = ", ".join(f"{label} ({env_url})" for label, env_url, _ in table)
raise RuntimeError(
f"No TS_ENV entry matches {url}. Known environments: {known}. "
f"Add a TS_ENV_<n>_LABEL/URL/KEY_VAR triplet for this instance."
)
from mcp.server.fastmcp import FastMCP
from mcp.server.transport_security import TransportSecuritySettings
# Bounded concurrency: the build pipeline is already concurrency-safe (the app/QA
# have run concurrent builds for months), so this is only a resource guardrail β€” a
# semaphore caps how many run at once so a caller can't OOM the Space by firing
# dozens; a call past the cap gets "busy". Owner/env/creds are per-instance, so
# concurrent builds share no mutable process state.
_MAX_CONCURRENT_BUILDS = max(1, int(os.getenv("MCP_MAX_CONCURRENT_BUILDS", "3")))
_build_sem = threading.BoundedSemaphore(_MAX_CONCURRENT_BUILDS)
# Lightweight hit log so we can watch incoming calls (who + what args). Writes to
# stderr (the server log) and, if MCP_HIT_LOG is set, appends to that file too.
_HIT_LOG = (os.getenv("MCP_HIT_LOG") or "").strip()
def _log_hit(msg: str) -> None:
line = f"[{time.strftime('%H:%M:%S')}] {msg}"
print(line, file=sys.stderr, flush=True)
if _HIT_LOG:
try:
with open(_HIT_LOG, "a") as fh:
fh.write(line + "\n")
except Exception:
pass
# --- build status tracking (for the status tool / polling long builds) ---
_SERVER_STARTED = time.time()
_state_lock = threading.Lock()
_build_count = 0
_active_builds = {} # run_id -> live build dict; supports multiple concurrent builds
_last_build = None # summary of the most recent completed build
# Compact result of each completed build, keyed by run_id (newest last), capped so
# a client that fired several concurrent builds can collect EACH result by run_id
# instead of racing the single _last_build slot. In-memory; resets on restart.
_recent_builds = collections.OrderedDict()
_RECENT_BUILDS_CAP = 50
def _set_progress(run_id: str, phase: str | None = None, detail: str | None = None) -> None:
"""Update a running build's coarse phase and/or latest raw progress line."""
with _state_lock:
b = _active_builds.get(run_id)
if b is not None:
if phase:
b["phase"] = phase
if detail is not None:
b["detail"] = detail
def _progress_text(item) -> str:
"""Best-effort human-readable line from a generator yield (str | dict | tuple)."""
if isinstance(item, dict):
for k in ("message", "text", "content", "status", "detail"):
v = item.get(k)
if isinstance(v, str) and v.strip():
item = v
break
else:
item = str(item)
elif isinstance(item, (tuple, list)):
item = next((x for x in item if isinstance(x, str) and x.strip()), str(item))
return " ".join(str(item).split())[:200] # collapse whitespace + cap length
def _phase_from_text(text: str) -> str | None:
"""Map a raw progress line to a granular phase label (None = keep current phase)."""
t = text.lower()
# [async] import lines say "…polling…" as normal operation β€” check them
# before the retry bucket so a healthy import isn't mislabeled "retrying".
if "[async]" in t:
return "importing tables"
if any(k in t for k in ("504", "retry", "retrying", "throttl", "timed out", "timeout")):
return "waiting on ThoughtSpot (retrying)"
if "connection" in t:
return "creating connection"
if "tml" in t or "enhance" in t or "post-process" in t or "applying style" in t:
return "liveboard TML phase"
if "liveboard" in t or "pinboard" in t:
return "creating liveboard"
if "model" in t:
return "model created" if any(k in t for k in ("created", "ready", "guid")) else "creating model"
if "verif" in t and "schema" in t:
return "verifying schema"
if "table" in t and ("creat" in t or "import" in t):
return "creating tables"
if any(k in t for k in ("import", "loading", "populat", "copy into", "batch", "chunk", " rows")):
return "loading data"
if "schema" in t and "creat" in t:
return "creating schema"
return None
# The app yields a periodic UI "spinner" banner while deploy_all runs in its
# worker thread. It is a heartbeat, not progress β€” and it happens to contain the
# word "connection" (…Creating connection & tables…), so if it reaches
# _set_progress it both freezes `phase` on "creating connection" and clobbers the
# real per-line detail arriving via on_progress. Drop it on the MCP side.
_BANNER_MARKERS = (
"deployment in progress",
"thoughtspot deploying",
"starting thoughtspot deployment",
)
def _is_heartbeat_banner(text: str) -> bool:
t = text.lower()
return any(m in t for m in _BANNER_MARKERS)
def _apply_progress(run_id: str, item) -> None:
"""Fold one generator yield / deploy line into the build's live status.
Ignores the UI heartbeat banner so the real per-line progress (delivered via
on_progress from the deploy thread) drives `phase`/`detail` instead of being
overwritten every 2s.
"""
text = _progress_text(item)
if not text or _is_heartbeat_banner(text):
return
_set_progress(run_id, phase=_phase_from_text(text), detail=text)
# DNS-rebinding protection validates the Host header and only trusts localhost
# by default, which returns 421 when reached through a tunnel or an hf.space host.
# We gate access on the bearer token instead (and this is server-to-server, not
# browser-driven), so disable the host check to allow any public host.
mcp = FastMCP(
"demoprep",
transport_security=TransportSecuritySettings(enable_dns_rebinding_protection=False),
)
@mcp.tool()
def ping() -> dict:
"""Health / connectivity check β€” returns server identity + config, no side effects.
Use this to confirm a client (e.g. AgentSpot) can reach and invoke the server
without triggering a full build.
"""
_log_hit("CALL ping")
return {
"ok": True,
"server": "demoprep-mcp v1",
"default_ts_url": os.getenv("TS_ENV_URL_DEFAULT", ""),
"default_owner": os.getenv("TS_USER_DEFAULT", ""),
"tools": ["ping", "status", "build_demo_from_brief"],
}
@mcp.tool()
def status() -> dict:
"""Report server + build status: whether the server is up, whether a build is
running (with its phase and elapsed time), and the most recent completed build.
Use this to poll a long build: call build_demo_from_brief, then poll status()
to watch progress and retrieve the result even if the original call's
connection dropped.
"""
_log_hit("CALL status")
now = time.time()
with _state_lock:
active = [dict(b) for b in _active_builds.values()]
last = dict(_last_build) if _last_build else None
recent = [dict(r) for r in _recent_builds.values()]
count = _build_count
for b in active:
b["elapsed_seconds"] = round(now - b.pop("started_at", now), 1)
active.sort(key=lambda b: b.get("elapsed_seconds", 0), reverse=True)
return {
"server": "up",
"default_ts_url": os.getenv("TS_ENV_URL_DEFAULT", ""),
"busy": len(active) > 0,
"active_count": len(active),
"capacity": _MAX_CONCURRENT_BUILDS,
"active_builds": active,
# Back-compat for single-build pollers: present only when exactly one runs.
"current_build": active[0] if len(active) == 1 else None,
"last_build": last,
# Compact result of each recently completed build (newest last), keyed data
# so a client that fired several builds can collect EACH result by run_id.
"recent_builds": recent,
"builds_started": count,
"uptime_seconds": round(now - _SERVER_STARTED, 1),
}
def _run_build(run_id: str, brief: str, company_name: str, use_case: str, company_url: str,
ts_target_url: str, ts_auth_key: str, owner: str) -> dict:
"""Drive the controller headlessly (runs in a BACKGROUND THREAD; the caller has
already registered the build in _active_builds and holds a _build_sem slot (released when this
returns). Mirrors tests/newvision_sample_runner.py but injects `brief` in place of
the research phase. Returns a structured dict and never raises β€” failures come back
as status 'failed' / 'partial'."""
started = time.time()
ddl_text = "" # surfaced via status/result once generated; stays "" if the build fails earlier
# Full-brief capture + durable persist β€” ONCE per accepted build (this only runs for
# builds that actually got a slot, so busy-retries of a queued build never duplicate
# rows). Logged with the run_id so the durable record correlates to the build.
_log_hit(f"PAYLOAD_BRIEF_BEGIN run_id={run_id} company={company_name!r} use_case={use_case!r} "
f"url={company_url!r} ts_url={ts_target_url!r} owner={owner!r} brief_len={len(brief or '')}")
print(brief or "", file=sys.stderr, flush=True)
_log_hit("PAYLOAD_BRIEF_END")
_persisted = log_mcp_payload({
"run_id": run_id, "company_name": company_name, "use_case": use_case,
"company_url": company_url, "ts_url": ts_target_url, "owner_email": owner, "brief": brief,
})
_log_hit(f"PAYLOAD_PERSISTED run_id={run_id} supabase={_persisted}")
def result(status: str, dc: dict | None = None, schema: str | None = None,
error: str | None = None) -> dict:
dc = dc or {}
out = {
"run_id": run_id,
"status": status, # success | partial | failed
"schema": dc.get("schema") or schema or "",
"model_guid": dc.get("model_guid", ""),
"liveboard_guid": dc.get("liveboard_guid", ""),
"model_url": dc.get("model_url", ""),
"liveboard_url": dc.get("liveboard_url", ""),
"ts_environment": ts_target_url,
"owner_email": owner,
"ddl": ddl_text,
"warnings": dc.get("warnings", []),
"errors": ([error] if error else dc.get("errors", [])),
"elapsed_seconds": round(time.time() - started, 1),
}
global _last_build
with _state_lock:
_last_build = {
"run_id": out["run_id"], "status": out["status"], "schema": out["schema"],
"model_url": out["model_url"], "liveboard_url": out["liveboard_url"],
"errors": out["errors"], "elapsed_seconds": out["elapsed_seconds"],
"ddl": out["ddl"],
"finished_at": time.strftime("%H:%M:%S"),
}
# Per-run result (compact, no ddl) so concurrent callers can collect EACH
# build by run_id without racing the single _last_build slot.
_recent_builds[run_id] = {
"run_id": out["run_id"], "status": out["status"], "schema": out["schema"],
"company_name": company_name,
"model_url": out["model_url"], "liveboard_url": out["liveboard_url"],
"warnings": out["warnings"], "errors": out["errors"],
"elapsed_seconds": out["elapsed_seconds"], "finished_at": time.strftime("%H:%M:%S"),
}
while len(_recent_builds) > _RECENT_BUILDS_CAP:
_recent_builds.popitem(last=False) # drop oldest
_active_builds.pop(run_id, None)
return out
controller = None
try:
# (i) controller β€” acting user = the per-build owner (per-instance; no
# process-global write, so concurrent builds never race on the owner).
controller = ChatDemoInterface(user_email=owner)
# (ii) settings: model + fixed TS env (exact key names per the wiring trace)
controller.settings["model"] = controller.settings.get("model") or DEFAULT_LLM_MODEL
controller.settings["thoughtspot_url"] = ts_target_url
controller.settings["thoughtspot_trusted_auth_key"] = ts_auth_key
# Always share the created model + liveboard with the operator, so every demo is
# visible no matter who ran it. The OWNER is unchanged β€” whoever was passed as
# owner_email (the build authenticates AS that user, so they own the objects);
# this just adds the operator as a viewer. Override target via MCP_SHARE_WITH.
controller.settings["share_with"] = os.getenv("MCP_SHARE_WITH", "mike.boone@thoughtspot.com")
# vertical / function / use_case_config exactly as the runner does
controller.vertical, controller.function = parse_use_case(use_case or "")
controller.use_case_config = get_use_case_config(
controller.vertical or "Generic", controller.function or "Generic"
)
# (iii) demo_builder with the BRIEF injected in place of research.
db = DemoBuilder(use_case=use_case, company_url=company_url)
db.company_analysis_results = brief # component field
db.combined_research_results = brief # <-- the field build_demo actually reads
db.company_summary = brief # <-- liveboard Spotter story reads this (Gotcha 2)
# Force the exact display name β€” extract_company_name() otherwise parses the
# domain from company_url. Minimal shim standing in for a scraped website.
db.website_data = types.SimpleNamespace(
title=company_name, url=company_url, text="", css_links=[], logo_candidates=[]
)
controller.demo_builder = db
controller.generic_use_case_context = ""
# (iv) DDL β€” returns a (response, ddl) tuple; NOT a generator.
_set_progress(run_id, phase="building dataset + DDL", detail="")
resp, ddl_text = controller.run_ddl_creation()
if not ddl_text or "CREATE TABLE" not in ddl_text.upper():
return result("failed", error=f"DDL generation failed: {str(resp)[:500]}")
# Surface the DDL immediately β€” it exists ~5 min in, well before the ~15-min
# TS deploy β€” so a caller polling status() gets the schema as soon as it's ready.
with _state_lock:
b = _active_builds.get(run_id)
if b is not None:
b["ddl"] = ddl_text
# (v) Snowflake load, then ThoughtSpot. Both are generators β€” draining them
# IS what runs the work. Decoupled from validation_mode: drain the Snowflake
# generator, read the schema it set, then run the TS deploy ourselves.
_set_progress(run_id, phase="loading Snowflake", detail="")
for _item in controller.run_deployment_streaming():
_apply_progress(run_id, _item)
schema = getattr(controller, "_deployed_schema_name", None)
if not schema:
return result(
"failed",
schema=getattr(controller, "_last_schema_name", None),
error="Snowflake load did not complete (no deployed schema).",
)
_set_progress(run_id, phase="deploying ThoughtSpot model + liveboard", detail="")
# Real-time status: on_progress fires from the deploy thread for EVERY
# progress line (incl. the [async] import + model/liveboard steps), so
# status.current_build.detail tracks live instead of freezing between yields.
def _dp(m):
_apply_progress(run_id, m)
for _item in controller._run_thoughtspot_deployment(schema, company_name, use_case, on_progress=_dp):
_dp(_item)
# (vi) structured result from the completion record.
dc = getattr(controller, "deployment_completion", None)
if not dc:
# deploy_all raised before the completion record was written β€” surface
# partial success: the Snowflake schema exists even if TS didn't finish.
return result(
"partial", schema=schema,
error="ThoughtSpot deploy did not complete; Snowflake schema exists.",
)
return result("success" if dc.get("success") else "partial", dc=dc, schema=schema)
except Exception as e: # never leak a raw exception to the MCP caller
schema = getattr(controller, "_deployed_schema_name", None) if controller else None
return result(
"partial" if schema else "failed",
schema=schema,
error=f"{type(e).__name__}: {e}",
)
@mcp.tool()
def build_demo_from_brief(
brief: str,
company_name: str,
use_case: str = "",
company_url: str = "",
ts_url: str = "",
owner_email: str = "",
) -> dict:
"""Start a full ThoughtSpot demo build from a prospect brief.
Skips DemoPrep's own research: the `brief` IS the research context. The build
(dataset -> DDL -> Snowflake -> model -> liveboard) runs in a BACKGROUND THREAD and
takes ~15-20 minutes, so this returns IMMEDIATELY with a run_id β€” it does NOT block.
Poll status() for progress (phase + elapsed) and the final result (schema, model,
liveboard URL); the build survives even if this call's connection drops.
Args:
brief: Prospect narrative β€” pain points, what they're evaluating, industry
context, goals. Becomes the research context the demo is built from.
company_name: Display name for the demo (e.g. "Acme Corporation").
use_case: (optional) The analytics story / label (e.g. "Retail Sales").
If omitted, the demo is authored purely from the brief and
labeled "Custom Analytics".
company_url: Optional company URL (used for context/branding; not scraped).
ts_url: Optional ThoughtSpot instance URL to deploy into. If omitted,
falls back to TS_ENV_URL_DEFAULT; the trusted-auth key is chosen
by instance (sebe vs SE primary).
owner_email: Optional ThoughtSpot user to own the created objects. If
omitted, falls back to TS_USER_DEFAULT.
Returns:
dict: {status: "started"|"busy"|"failed", run_id, message}. Poll status()
for progress and the eventual result.
"""
_log_hit(
f"CALL build_demo_from_brief company={company_name!r} use_case={use_case!r} "
f"url={company_url!r} ts_url={ts_url!r} owner={owner_email!r} brief_len={len(brief or '')} "
f"brief_preview={(brief or '')[:200]!r}"
)
# NOTE: the full-brief capture + durable persist happen ONCE per ACCEPTED build,
# inside _run_build β€” NOT here. Doing it here (before the slot/busy check) meant a
# client re-polling a queued build sprayed a duplicate durable row on every retry.
# The lightweight CALL line above still records every attempt in the (ephemeral) log.
if not brief or not brief.strip():
return {"status": "failed", "errors": ["brief is required"]}
if not (company_name or "").strip():
return {"status": "failed", "errors": ["company_name is required"]}
# Resolve the ThoughtSpot target up front (URL from the request or the default,
# key chosen by instance) so a bad URL / missing key fails fast β€” before we
# take the single build slot.
try:
ts_target_url, ts_auth_key = _resolve_ts_target(ts_url)
owner = _resolve_owner(owner_email)
except RuntimeError as e:
return {"status": "failed", "errors": [str(e)]}
# Bounded concurrency: grab a build slot (non-blocking). Only when ALL slots are
# in use does a call get 'busy' β€” otherwise it runs alongside the others.
if not _build_sem.acquire(blocking=False):
return {
"status": "busy",
"errors": [f"All {_MAX_CONCURRENT_BUILDS} build slots are in use. Poll status(); retry when one frees."],
}
run_id = uuid.uuid4().hex[:12]
b, c, u, url = brief.strip(), company_name.strip(), (use_case or "Custom Analytics").strip(), (company_url or "").strip()
global _build_count
with _state_lock:
_build_count += 1
_active_builds[run_id] = {
"run_id": run_id, "company_name": c, "use_case": u,
"ts_url": ts_target_url,
"started_at": time.time(), "phase": "starting",
}
# Run the build OFF the request thread so the server stays responsive to status()
# polls throughout β€” a synchronous build monopolizes the interpreter and starves them.
def _worker() -> None:
try:
_run_build(run_id, b, c, u, url, ts_target_url, ts_auth_key, owner)
finally:
_build_sem.release()
threading.Thread(target=_worker, daemon=True, name=f"build-{run_id}").start()
return {
"status": "started",
"run_id": run_id,
"message": "Build started in the background (~15-20 min). Poll status() for progress and the final result (schema, model, liveboard URL).",
}
def _run_http(token: str) -> None:
"""Serve over streamable HTTP behind a shared-bearer gate.
The auth check is a PURE-ASGI wrapper, NOT Starlette's BaseHTTPMiddleware β€”
the latter buffers responses and breaks the streamable-HTTP SSE stream. This
checks the bearer on every HTTP request and otherwise passes the raw ASGI
through untouched.
"""
import uvicorn
inner = mcp.streamable_http_app() # verified accessor on mcp 1.28.1
class _BearerGate:
def __init__(self, app):
self.app = app
async def __call__(self, scope, receive, send):
if scope.get("type") == "http":
path = scope.get("path", "")
if path in ("/", "/health"):
# Unauthenticated liveness check so the HF Space reports healthy.
# The MCP protocol itself lives at /mcp behind the bearer gate.
await send({
"type": "http.response.start",
"status": 200,
"headers": [(b"content-type", b"application/json")],
})
await send({"type": "http.response.body",
"body": b'{"status":"ok","service":"demoprep-mcp"}'})
return
headers = dict(scope.get("headers") or [])
ip = (headers.get(b"x-forwarded-for", b"") or headers.get(b"cf-connecting-ip", b"")).decode().split(",")[0].strip()
_log_hit(f"HIT {scope.get('method', '?')} {path} "
f"auth={'Y' if b'authorization' in headers else 'N'} ip={ip or '?'}")
if headers.get(b"authorization", b"").decode() != f"Bearer {token}":
await send({
"type": "http.response.start",
"status": 401,
"headers": [(b"content-type", b"application/json")],
})
await send({"type": "http.response.body", "body": b'{"error":"unauthorized"}'})
return
await self.app(scope, receive, send)
uvicorn.run(_BearerGate(inner), host="0.0.0.0", port=int(os.getenv("MCP_HTTP_PORT", "7860")))
def main() -> None:
transport = (os.getenv("MCP_TRANSPORT") or "stdio").strip().lower()
print(
f"[mcp_server] ready β€” default_ts_url={os.getenv('TS_ENV_URL_DEFAULT', '(unset)')} "
f"default_owner={os.getenv('TS_USER_DEFAULT', '(unset)')} transport={transport}",
file=sys.stderr, flush=True,
)
if transport in ("stdio", ""):
mcp.run(transport="stdio")
elif transport in ("http", "streamable-http", "streamable_http"):
token = (os.getenv("MCP_ACCESS_TOKEN") or "").strip()
if not token: # public endpoint must be gated
raise RuntimeError("MCP_ACCESS_TOKEN is required in http mode (public endpoint).")
_run_http(token)
else:
raise RuntimeError(f"Unknown MCP_TRANSPORT: {transport!r} (use 'stdio' or 'http').")
if __name__ == "__main__":
main()