Ajay Krishnan
b4e863562c
context-docs was previously spawned per call as a fresh stdio container,
which meant every MCP request paid full cold-start cost (embedding model
load + Chroma open) and concurrent clients raced for the same Chroma
writer. The 50+ orphan container build-up I saw during the publish audit
was the visible symptom.
This refactor runs docs-mcp as one long-lived service:
- compose: docs-mcp leaves the 'mcp' profile, gets container_name,
restart: unless-stopped, healthcheck, and a host port (127.0.0.1:8776
by default). Runs as the host UID/GID so bind mounts don't end up
root-owned.
- docker image: adds mcp-proxy (0.12.0) and an entrypoint that fronts
llms-txt-mcp's stdio as Streamable HTTP. Reads sources from a flat
file mounted at /etc/context-kit/docs-sources.txt. Disables eager
preindex by default; callers refresh on demand via the docs_refresh
tool. Set CONTEXT_KIT_DOCS_PREINDEX=1 to restore eager behavior.
- bin/context-kit: 'start' brings up the docs service alongside SearXNG,
generates the sources file from CONTEXT_KIT_DOCS_SOURCES, and waits
for the HTTP endpoint to become ready (up to 180s for first-run model
download). 'docs' still works for stdio-only clients but is now a
thin mcp-proxy bridge onto the shared HTTP service. 'doctor' and
'status' both surface the new endpoint.
- install snippets: context-docs is now 'type: remote'/'type: http'
pointing at ${CONTEXT_KIT_DOCS_HTTP_URL}. HTTP-capable MCP clients
bypass the bridge entirely. snippets/*.json and the install command
output stay byte-identical.
- docs and .env.example updated for new vars (CONTEXT_KIT_DOCS_PORT,
CONTEXT_KIT_DOCS_HTTP_URL, CONTEXT_KIT_DOCS_PREINDEX) and the new
24h TTL default (down from 7d; the long-lived service makes shorter
defaults cheap).
Verified end-to-end:
- compose config -q, bash -n, sh -n all clean
- HTTP /status returns 200
- stdio bridge returns initialize + tools/list with the same 3 tools
(docs_sources, docs_refresh, docs_query)
- doctor passes all 10 checks including the new HTTP probe
- web-search and repomix MCP handshakes still work
- redaction-check clean
- install JSON valid for both targets + --absolute
2.1 KiB
Configuration
Configuration is via environment variables or a .env file in the repository
root. Start from .env.example.
Explicit environment variables win over .env values. The .env parser accepts
simple KEY=VALUE lines for CONTEXT_KIT_* variables only; it does not execute
shell code.
Core Variables
| Variable | Default | Purpose |
|---|---|---|
CONTEXT_KIT_DATA_DIR |
$HOME/.local/share/context-kit |
Persistent docs indexes and model cache |
CONTEXT_KIT_COMPOSE_PROJECT |
context-kit |
Docker Compose project and network prefix |
CONTEXT_KIT_SEARXNG_PORT |
8099 |
Localhost SearXNG port |
CONTEXT_KIT_DOCS_PORT |
8776 |
Localhost port for the long-lived docs-mcp HTTP service |
CONTEXT_KIT_DOCS_HTTP_URL |
http://127.0.0.1:${CONTEXT_KIT_DOCS_PORT}/mcp |
URL emitted into install snippets and used by the stdio bridge |
CONTEXT_KIT_DOCS_TTL |
24h |
Docs re-fetch cadence |
CONTEXT_KIT_DOCS_SOURCES |
config/sources.default.txt |
Space-separated source profile files |
CONTEXT_KIT_DOCS_MAX_GET_BYTES |
75000 |
Max bytes returned by docs retrieval |
CONTEXT_KIT_DOCS_EMBED_MODEL |
BAAI/bge-small-en-v1.5 |
SentenceTransformers embedding model |
CONTEXT_KIT_DOCS_PREINDEX |
0 |
Set to 1 to re-embed every source on container start |
TTL Guidance
24h is the default. Most reference docs do not need re-embedding more often,
and the shared service does not re-fetch sources until the TTL elapses.
Use shorter TTLs for fast-moving APIs:
CONTEXT_KIT_DOCS_TTL=6h bin/context-kit restart
Use longer TTLs for stable specs:
CONTEXT_KIT_DOCS_TTL=30d bin/context-kit restart
The docs-mcp container reads CONTEXT_KIT_DOCS_TTL at startup, so changes
require bin/context-kit restart. When freshness matters for one task, prefer
calling the docs_refresh MCP tool instead of lowering the global TTL.
Source Profiles
The docs MCP accepts one or more source files:
CONTEXT_KIT_DOCS_SOURCES="config/sources.default.txt config/sources.js.txt"
Each source file is plain text. Blank lines and # comments are ignored.