Configuration

Every setting is optional; the defaults work for local development. Settings are read from environment variables prefixed MANNA_, or from a .env file in the working directory. Unknown MANNA_* variables are ignored. The source of truth is src/manna/config.py.

Variable	Default	What it does
MANNA_HOST	0.0.0.0	HTTP bind address (HTTP mode only)
MANNA_PORT	8000	HTTP listen port (HTTP mode only)
MANNA_LOG_LEVEL	INFO	DEBUG, INFO, WARNING, or ERROR; logs are JSON lines on stderr
MANNA_ARCHIVES	(unset)	Comma-separated archive short names to activate; unset means every non-paused archive (Archives)
MANNA_ALLOWED_HOSTS	(unset)	Comma-separated hostnames the server may fetch (exact or subdomain match). Unset allows any public host; private, loopback, and link-local targets are refused regardless (Security)
MANNA_TAP_SYNC_TIMEOUT_SECONDS	20.0	Client timeout for a TAP /sync request; in mode="auto" a timeout promotes the query to an async job
MANNA_COUNT_ASYNC_BUDGET_SECONDS	15.0	count_observations_near_target: how long to poll an async count job before returning a pending envelope with the job_url
MANNA_COUNT_ASYNC_POLL_INTERVAL_SECONDS	1.0	Interval between those polls
MANNA_ASYNC_WAIT_SECONDS	20.0	How long get_async_job_status waits server-side for a terminal phase when the call omits wait_seconds; also the wait after an async submission in run_adql_query before the promotion envelope is returned
MANNA_ASYNC_WAIT_MAX_SECONDS	30.0	Ceiling applied to any caller-supplied wait_seconds. Keep it under the MCP client’s per-call timeout (60 s in the MCP SDKs)
MANNA_INLINE_ROW_LIMIT	200	Max rows returned inline before a TAP result is promoted to async or a cone/SIA result is truncated
MANNA_INLINE_BYTE_LIMIT	49152	Same, in bytes of the serialized envelope (48 KiB)
MANNA_REGISTRY_DESCRIBE_BYTE_LIMIT	49152	Above this, describe_ivoa_service degrades from per-column detail to a table catalog (names, descriptions, column counts)

Presets by model backend

The inline caps default to sizes that suit a small-context local model, roughly 64K to 128K tokens. Raise them for a frontier model:

Backend	MANNA_INLINE_ROW_LIMIT	MANNA_INLINE_BYTE_LIMIT
Small local model (e.g. gpt-oss-120b on vLLM)	200	49152
Claude (200K+)	2000	262144

Transport

There is one transport flag, on the command line rather than in the environment: manna --stdio serves MCP over stdio for clients that launch the server themselves. Without it the server binds MANNA_HOST:MANNA_PORT and serves Streamable HTTP at /mcp/, plus /health and /ready.

Example .env

The repository ships .env.example with every variable and a comment. Copy it to .env and edit.