Delete the uptrace DSN parser and the {endpoint}/v1/* concatenation that
produced the dead export, and build the OTLP exporters with no endpoint,
headers, or protocol so the SDK reads OTEL_EXPORTER_OTLP_* itself.
The SDK concatenates as naively as we did: measured against a local sink,
a ?query base sends POST /?query/v1/metrics and a #fragment base sends
POST /. So one check stays -- a query or fragment on
OTEL_EXPORTER_OTLP_ENDPOINT fails at startup. Per-signal variables are
used verbatim and need none.
A protocol variable is the second startup check. The exporter picks its
transport at build time and this binary ships OTLP/HTTP alone, so
OTEL_EXPORTER_OTLP_PROTOCOL=grpc kept exporting over HTTP with nothing in
the log -- the same silence this change removes.
service.name stays a default rather than policy: an attribute set on the
resource builder wins over the SDK detectors that read OTEL_SERVICE_NAME
and OTEL_RESOURCE_ATTRIBUTES, so anwesen supplies its own only for keys
the environment leaves alone.
Telemetry-off is our own check on the endpoint variables: with none set
the exporter would still build and aim at the SDK default localhost:4318.
--uptrace-dsn, --otlp-endpoint, and --otlp-header stay parsed but hidden,
so a deployment upgrading with them set fails naming the OTEL_
replacement instead of going quiet.
Assumed hidden clap stubs are the right migration shape; clap rejecting
the flags as unknown would say nothing about the replacement. Flag if a
plain unknown-argument error is wanted instead.
Assumed rejecting http/json alongside grpc is right: neither has a
transport in this build. Flag if a build with both features is wanted
instead.
|
||
|---|---|---|
| .github/workflows | ||
| deploy | ||
| scripts | ||
| src | ||
| tests | ||
| .gitignore | ||
| Cargo.lock | ||
| Cargo.toml | ||
| LICENSE | ||
| README.md | ||
Anwesen
Anwesen (German): the premises, the estate -- and, in Heidegger, a coming-to-presence. A read-only HTTP daemon that brings a markdown vault into presence over the network.
Anwesen serves an Obsidian-style markdown vault read-only over HTTP. It walks the notes in place, parses their YAML frontmatter, and lets remote clients query notes by frontmatter fields, read individual notes, and list folders -- with no database, no copy, and no migration of the underlying directory. The same query and merge engine is also available offline as a one-shot CLI.
Why
Obsidian is the odd one out among knowledge stores in not shipping a native HTTP API. Notion, Confluence, Wiki.js, Outline and the rest expose one out of the box; an Obsidian vault is just a directory of markdown files on disk. Anwesen brings that vault up to the same bar without converting it into something else: the files stay exactly where Obsidian wrote them and remain editable in Obsidian with no coordination. Programs that need the vault's contents stop re-implementing "walk the directory, parse the YAML, filter the notes" -- they ask Anwesen.
What it does
Obsidian vault -> Anwesen (walk + frontmatter index) -> HTTP/JSON -> consumers
Anwesen answers three kinds of question over HTTP:
- Give me this one note --
GET /notes/<path> - List the contents of this folder --
GET /notes/<folder>/ - Find every note whose frontmatter matches these predicates --
GET /query?...
The frontmatter index is built once at startup and kept current by watching the vault directory. The index lives in memory; a restart rebuilds it, and there is nothing on disk to corrupt or migrate.
The same query-and-merge engine also runs offline, with no server: anwesen merge walks a directory, evaluates a query, and writes the merged markdown to stdout (see Local generation).
Quick start
Build (Rust, 2024 edition, stable toolchain):
cargo build --release
Run:
anwesen serve --vault /path/to/vault
Anwesen binds 127.0.0.1:8080 by default, walks the vault once, then serves the API and watches for changes. Stop it with SIGINT/SIGTERM; there is nothing on disk to clean up.
Check a vault before serving it:
anwesen doctor --vault /path/to/vault
Build one file out of many notes, without starting the daemon:
anwesen merge --vault /path/to/vault --query 'tags=adr&__anw-order=title' > ADRs.md
CLI
anwesen serve --vault <path> [--bind <addr:port>] [--log-level <level>]
[--otlp-slow-request-ms <n>]
anwesen doctor --vault <path>
anwesen merge --vault <path> --query <query-string>
anwesen version
| Flag | Env var | Default | Meaning |
|---|---|---|---|
--vault <path> |
ANWESEN_VAULT |
required | Path to the vault root. |
--bind <addr:port> |
ANWESEN_BIND |
127.0.0.1:8080 |
Listen address for serve. |
--log-level <level> |
ANWESEN_LOG_LEVEL |
info |
error, warn, info, debug, or trace. |
--query <query-string> |
-- | required for merge |
A /query query string: frontmatter predicates plus __anw- controls. |
--otlp-slow-request-ms |
ANWESEN_OTLP_SLOW_REQUEST_MS |
500 |
Requests at or over this duration, or answering 5xx, also export a span. |
Every flag has a matching ANWESEN_<UPPER> environment variable. CLI flags win
over env vars.
--bind and --otlp-slow-request-ms apply to serve only. Where telemetry is
exported is configured entirely through the standard OTEL_ variables below.
serve-- run the daemon: walk the vault, build the index, watch for changes, serve the API.doctor-- walk the vault once and report what would stop clean ingestion: unreadable files, unparseable YAML, path collisions on the HTTP surface, and frontmatter type drift (the same key carrying incompatible types across notes). Read-only; non-zero exit if any issue is found.merge-- one-shot local generation: walk the vault, evaluate--query, and write the merged markdown document to stdout. No server, no HTTP. See Local generation.version-- print version and exit.
Telemetry
anwesen exports OTLP metrics for every request, and a span for requests at or
over --otlp-slow-request-ms or answering a 5xx. Export is configured through
the standard OpenTelemetry environment variables, which the SDK reads directly:
| Variable | Meaning |
|---|---|
OTEL_EXPORTER_OTLP_ENDPOINT |
Base URL. The SDK appends /v1/metrics and /v1/traces. |
OTEL_EXPORTER_OTLP_METRICS_ENDPOINT |
Full metrics URL, used as given. Overrides the base for metrics. |
OTEL_EXPORTER_OTLP_TRACES_ENDPOINT |
Full traces URL, used as given. Overrides the base for traces. |
OTEL_EXPORTER_OTLP_HEADERS |
Export headers, key=value comma-separated. |
OTEL_EXPORTER_OTLP_PROTOCOL |
http/protobuf, the default and the only value this build speaks. |
OTEL_SERVICE_NAME, OTEL_RESOURCE_ATTRIBUTES |
Override the anwesen service identity. |
The per-signal OTEL_EXPORTER_OTLP_METRICS_PROTOCOL and
OTEL_EXPORTER_OTLP_TRACES_PROTOCOL are read the same way.
With none of the three endpoint variables set, telemetry is off: no exporter is built and the request middleware is not installed.
Two settings fail at startup rather than export nowhere in silence:
- A query or a fragment on
OTEL_EXPORTER_OTLP_ENDPOINT. The base URL must be a base URL; the SDK appends the signal path after whatever it is given. - A protocol other than
http/protobuf. The binary ships the OTLP/HTTP transport alone, sogrpcwould keep exporting over HTTP with nothing in the log. Point the endpoint at the collector's HTTP port, not its gRPC one.
uptrace:
OTEL_EXPORTER_OTLP_ENDPOINT=https://api.uptrace.dev
OTEL_EXPORTER_OTLP_HEADERS=uptrace-dsn=https://TOKEN@api.uptrace.dev?grpc=4317
Paste the DSN from Project Settings into the header verbatim, tail and all -- it is a credential there, not an address. The endpoint is the host alone.
Removed in 0.3.0: --uptrace-dsn, --otlp-endpoint, --otlp-header and their
ANWESEN_ variables. Passing any of them fails at startup with the OTEL_
replacement to use.
HTTP API
All endpoints are GET and return JSON unless noted.
GET /notes/<path> -- read one note
Returns { path, frontmatter, body, last_modified, etag, size }, where body is the markdown with the frontmatter block stripped.
Accept: text/markdownreturns the raw file verbatim, frontmatter intact.If-None-Match: "<etag>"returns304 Not Modifiedwhen the note is unchanged. ETags are strong -- the BLAKE3 hash of the file's raw bytes -- so they are stable across restarts and unaffected by identical-byte rewrites (format-on-save, sync clients,git checkoutof the same revision)...segments in the path are rejected with400.
GET /notes/<folder>/ -- directory listing
Note the trailing slash; that is how a folder index is distinguished from a note read. Returns the immediate children only (non-recursive), each with name, type (file/dir), last_modified, and size.
GET /query?... -- frontmatter query
Returns { results, total, truncated }. Each query-string key=value is a predicate on a frontmatter field; different keys combine with AND.
GET /query?tags=anwesen&kind=adr
Field operators (suffix on the key):
| Suffix | Example | Meaning |
|---|---|---|
| (none) | tags=python |
Exact match on scalars; "contains" if the field is a list. |
__in |
tags__in=python,go |
Any of the listed values matches. |
__all |
tags__all=python,fastapi |
List contains all listed values. |
__not |
status__not=draft |
Negation. |
__exists |
deprecated__exists=false |
Presence / absence of the key. |
__regex |
title__regex=^PDR-\d+ |
Regex on a scalar (anchor it for speed). |
__prefix |
path__prefix=Projects |
String prefix. |
__gt __gte __lt __lte |
date__gte=2026-01-01 |
Ordered comparison; numbers and ISO dates. |
ISO-8601 dates and RFC 3339 datetimes are coerced to typed dates at read time, so range operators work on them. Nested keys use dots: author.name=.... Unknown operators return 400 rather than being silently ignored. Unanchored substring (__contains) is intentionally not provided; use __regex.
Control parameters (__anw- prefix; configure the query rather than constrain matches):
| Parameter | Default | Meaning |
|---|---|---|
__anw-recursive=<bool> |
true |
Recurse into subdirectories. |
__anw-path=<prefix> |
vault root | Restrict matches to a path prefix. |
__anw-limit=<n> |
no limit | Cap the result list; total still reports the full match count. |
__anw-order=<key>[:asc|:desc] |
path order | Order fragments (merge mode only). |
__anw-kind=<key> |
off | Refuse a mixed merge unless every matched note shares one value for the key (merge mode only). |
By default /query returns metadata only; fetch bodies with /notes/<path>.
Markdown-merge mode
Accept: text/markdown on /query returns the bodies of all matched notes concatenated into one markdown document, each fragment preceded by an HTML-comment source marker and separated by a blank line. This is the single-request path for building one file out of many notes:
<!-- source: ADR-001 Language and Foundation Libraries.md -->
...body...
<!-- source: ADR-002 Filesystem Change Tracking.md -->
...body...
__anw-order sets fragment order; __anw-kind guards against merging notes that disagree on a key (returns 400 naming the offenders). Both evaluate over the full match set before __anw-limit truncates.
The same merged document is available offline, without the daemon, via anwesen merge (see Local generation).
GET /health -- liveness and index freshness
Returns vault path, note count, last index/event timestamps, watcher state, an in-flight-rescan flag, and supervisor restart counters. Always returns 200 while the process is up; for a hard liveness signal use TCP connectivity.
Status codes
200 ok | 304 ETag matched | 400 bad request | 404 not found | 500 internal | 503 index not ready (retry shortly).
Local generation
anwesen merge produces the markdown-merge document on the command line, with no server and no HTTP round-trip. It walks the vault, evaluates the query, and writes the merged document to stdout:
anwesen merge --vault /path/to/vault --query 'tags=adr&__anw-order=title&__anw-kind=kind'
The --query string is the exact /query grammar: frontmatter predicates plus the __anw- controls, including __anw-order for fragment order and __anw-kind for the homogeneity guard. The output is byte-identical to the HTTP merge mode for the same vault and query -- both run the same engine. A __anw-kind violation exits non-zero and names the offending values on stderr; an empty match set writes nothing and exits 0.
This is the materialization path: build a CLAUDE.md, a skill bundle, or any single file assembled from many notes, driven from a script or a one-off shell.
Design notes
- In place, read-only. Anwesen reads the same directory Obsidian writes to and never writes back. The vault stays editable in Obsidian with no coordination, and there is no write API by design.
- Frontmatter is the index. Filtering is server-side and first-class; clients never walk-and-parse. Note bodies are served verbatim but not full-text indexed.
- No authentication. Anwesen trusts every request it accepts and binds
127.0.0.1by default. Put a reverse proxy (nginx, caddy, warpgate) in front for TLS and access control. The contract is simple: if you can reach Anwesen, you can read everything it indexes. - Standalone. Anwesen is its own repository with its own release cycle, reusable by any consumer -- not a sub-package of the first thing that used it.
Implementation
Rust (2024 edition) on Tokio: axum/tower for HTTP, clap for the CLI, tracing for logs, serde (with serde_yaml and serde_json) for the data model, notify for filesystem watching, and BLAKE3 for ETags. The frontmatter index is evaluated in memory; there is no external search engine and no on-disk index.
A Note on Agentic Engineering
LLM agents were heavily involved in writing this code. Every change has nonetheless been reviewed by a human, who made a best effort to understand, test, and validate it before it landed. Where the agents and the human disagreed, the human had the final say -- and bears responsibility for the result, bugs included. Read it with the same scrutiny you'd apply to any code: trust, but verify.
License
BSD 3-Clause. See LICENSE.