Connectors

A connector turns a source into a file-like URI tree. Once you register one, the same commands you use on a local folder work everywhere — search, grep, ls, tree, cat — whether the bytes live in a GitHub repo, a Postgres table, a Slack channel, or an S3 bucket:

mfs search "release checklist" github://owner/repo
mfs ls slack://acme/channels
mfs cat postgres://prod-db/public/tickets/rows.jsonl --locator '{"id": 12345}'

This page covers what's available and the mechanics shared across all connectors — the lifecycle, the credential model, and the [[objects]] config for structured sources. Each connector has its own page with its URI shape, credentials, and quirks; the catalog below links to them.

Connectors are optional extras

file is always available. The other connectors load lazily and are skipped on a server that doesn't have their dependencies installed. A scheme being listed here means it's in the codebase — not that it's installed on every server. Always probe on the target server before a large sync.

The catalog

Local & web

Connector	Use it for
file	Local directory trees — code, docs, PDFs. No credentials.
web	Crawl HTTP(S) pages into searchable markdown.
s3	Objects under an S3 / R2 / GCS / MinIO bucket prefix.
gdrive	Google Drive files visible to a user OAuth token.
notion	Notion pages and database records shared with an integration.

Databases & warehouses

Connector	Use it for
postgres	Postgres table rows + schema, incremental sync.
mysql	MySQL table rows + schema in one database.
mongo	MongoDB documents, one record each.
bigquery	BigQuery table rows via Application Default Credentials.
snowflake	Snowflake table rows, key-pair / password / PAT auth.

Issues, code & work tracking

Connector	Use it for
github	Repo code, plus optional issues and pull requests.
jira	Jira issues by project (Cloud or Server).
linear	Linear issues across teams.

CRM & support

Connector	Use it for
hubspot	HubSpot CRM objects — contacts, companies, deals, tickets.
zendesk	Zendesk tickets, comments, users, organizations.

Chat, mail & collaborative docs

Connector	Use it for
slack	Slack channel threads and the user directory.
discord	Discord guild channel messages and active threads.
gmail	Gmail threads under chosen labels.
feishu	Feishu / Lark chats and docx documents.

File-like vs structured sources

Connectors come in two shapes, and the difference decides how much config you write:

• File-like sources (file, web, github code, s3, gdrive, notion pages) expose real files. They're classified by type and indexed automatically — no per-object config needed.
• Structured sources (the databases, plus issue/CRM/chat connectors) expose rows, records, and message threads as .jsonl objects. To make a row searchable, MFS needs to know which fields carry text. The public SaaS connectors ship built-in presets so they work out of the box; the database connectors need an [[objects]] block per table (see below).

Lifecycle

mfs connector probe <target> --config ./connector.toml   # check creds + reachability
mfs add <target> --config ./connector.toml               # register + index (returns a job)
mfs job show JOB_ID                                       # watch it reach succeeded

Task	Command
Probe without registering	mfs connector probe <target> --config <file>
Add / register and index	mfs add <target> --config <file>
Skip the confirm prompt	mfs add <target> --config <file> --yes
Re-sync (incremental)	mfs add <target> again
Re-sync from a date	mfs add <target> --since <date> (only gdrive and feishu)
Force a full re-index	mfs add <target> --force-index
List connectors	mfs connector list
Inspect one	mfs connector inspect <target>
Update config	mfs connector update <target> --config <file>
Remove	mfs connector remove <target> (or mfs remove <target>)

For an external target, mfs add first runs a zero-billing estimate — it discovers the object count and dry-runs the chunker/tokenizer (no embedding API calls) and asks you to confirm before queueing. Pass --yes to skip it.

Remove is destructive inside MFS

Removing a connector never touches the source system, but it does drop the connector's MFS metadata, artifacts, and index. Keep the confirmation prompt on unless you're scripting deliberately.

Credentials: references, not values

A connector's TOML never holds a raw secret. It carries a reference, and the server resolves it when it builds the plugin — so the real value lives on the server, never in the TOML or the database:

token = "env:SLACK_BOT_TOKEN"                       # from a server env var
dsn = "env:PG_PROD_DSN"
credential_ref = "file:/etc/mfs/snowflake/rsa_key.p8"   # file contents (k8s/docker secret)

• env:VAR_NAME — read from the server process environment (not your client shell) when the connector is constructed.
• file:/abs/path — the file's contents; the path must be absolute.

Set env vars in the environment that starts mfs-server, and keep credential files outside the repo with tight permissions. See Auth and secrets for the full boundary.

[[objects]] for structured rows

Database connectors need an object rule per table so MFS knows which columns are searchable text and which key reopens an exact record:

[[objects]]
match = "/public/tickets"               # connector-relative object path
text_fields = ["title", "description"]  # embedded, searchable content
locator_fields = ["id"]                 # the key for `cat --locator`
metadata_fields = ["status", "updated_at"]  # returned alongside hits

The public SaaS connectors (GitHub issues, Jira, Slack, Linear, Zendesk tickets, Gmail, Discord, Feishu) ship presets, so [[objects]] is optional there — use it only to override the default fields. See Search and browse for how these become content, locator, and metadata.fields in a result.

After ingest

Every scheme answers the same commands:

mfs search "billing escalation" zendesk://acme --top-k 10
mfs search "rate limit retry policy" --all
mfs tree postgres://prod-db -L 3
mfs cat github://owner/repo/src/main.rs --range 100:150

See Search and browse for the retrieval loop.