RedisVL MCP

RedisVL includes an MCP server that exposes a Redis-backed retrieval surface through a small, deterministic tool contract. It is designed for AI applications that want to search or maintain data in one or more existing Redis indexes without each client reimplementing Redis query logic.

What RedisVL MCP Does

The RedisVL MCP server sits between an MCP client and Redis:

1. It connects to one or more existing Redis Search indexes.
2. It inspects each index at startup and reconstructs its schema.
3. It initializes vector capabilities only when the configured search or upsert behavior needs them.
4. It exposes stable MCP tools for discovery, search, and optionally upsert.

This keeps each Redis index as the source of truth for its search behavior while giving MCP clients a predictable interface.

How RedisVL MCP Runs

RedisVL MCP works with a focused model:

• One server process binds to one or several existing Redis indexes, each addressed by a logical id.
• The server supports stdio (default), Streamable HTTP, and SSE transports.
• Search behavior is owned by per-index configuration, not by MCP callers.
• Vector search and server-side embedding are optional capabilities configured explicitly per index.
• Upsert is optional and can be disabled globally with read-only mode or per index with a read_only flag.

A single-index server remains the simplest deployment: when exactly one index is configured, callers can omit the index selector entirely and every tool call targets that index. Multi-index support is fully formal — it adds discovery and explicit routing without changing the single-index contract.

Config-Owned Search Behavior

MCP callers can control:

• query
• limit
• offset
• filter
• return_fields

These request-time controls are still bounded by runtime config. In particular, deep paging is limited by a configured maximum result window, enforced as offset + limit.

On a multi-index server, callers also choose which index to target through an optional index argument (see Index Selection). Callers do not choose:

• whether retrieval is vector, fulltext, or hybrid
• query tuning parameters such as hybrid fusion or vector runtime settings

That behavior lives in the per-index server config under indexes.<id>.search. The response includes search_type as informational metadata, but it is not a request parameter.

Single and Multiple Index Bindings

The YAML config uses an indexes mapping. Each entry is a logical binding keyed by an id (for example knowledge or tickets) that points to an existing Redis index through redis_name. The mapping may contain one entry or several; each binding is inspected, validated, and given its own search config, runtime limits, and optional vectorizer independently at startup. Startup is all-or-nothing — if any binding fails to initialize, the server does not start.

A single-binding config is the simplest case and behaves exactly as before: the lone binding is the implicit target of every call. With multiple bindings the server stays a single process and endpoint, but callers select a binding per call.

Index Selection and Discovery

On a multi-index server, every tool call must say which logical index it targets:

• search-records and upsert-records accept an optional index argument naming the logical id.
• When exactly one index is configured, index may be omitted and resolves to that sole binding (backward compatible).
• When multiple indexes are configured, omitting index is an invalid_request; the caller must name one.
• An unknown logical id is an invalid_request.
• Both tools echo the resolved index in their response so clients can confirm routing.

Because a client cannot guess the configured logical ids, multi-index servers expose a list-indexes discovery tool. Clients should call list-indexes first to enumerate the available indexes and their filterable fields, then pass the chosen id as index on subsequent calls.

Schema Inspection and Overrides

RedisVL MCP is inspection-first:

• the Redis index must already exist
• the server reconstructs the schema from Redis metadata at startup
• runtime field mappings remain explicit in config

In some environments, Redis metadata can be incomplete for vector field attributes. When that happens, schema_overrides can patch missing attrs for fields that were already discovered. It does not create new fields or change discovered field identity.

Startup also validates that the inspected schema does not collide with MCP-reserved score metadata field names for the configured search mode.

Read-Only and Read-Write Modes

RedisVL MCP registers search-records and list-indexes by default (see Custom Tool Profiles for turning a built-in off deliberately).

Write availability is enforced at two levels:

• Global read-only mode disables writes across every binding. It is controlled by the CLI flag --read-only or the environment variable REDISVL_MCP_READ_ONLY=true.
• Per-index read-only disables writes for a single binding via indexes.<id>.read_only: true, while other bindings stay writable.

These combine into each binding’s effective write availability: a binding is read-only if global read-only is on or that binding sets read_only: true. The upsert-records tool is registered only when at least one binding is writable, so a fully read-only server does not advertise it at all. When the tool is registered, a write to a read-only binding is rejected with forbidden before any data is changed. list-indexes reports each binding’s effective write availability as upsert_available.

Use read-only mode when Redis is serving approved content to assistants and another system owns ingestion — globally when no binding should accept writes, or per index when only some indexes are writable.

Authentication and Authorization

The HTTP transports can require a JWT bearer token issued by an existing identity provider. The server validates the token signature, issuer, and audience, and can gate read vs write by scope or role claim. A custom tool profile can also scope every query to a tenant carried in the token; see Tenant Scoping From Token Claims. What the server does not do is map token claims to Redis ACL users or to separate indexes, which remains a gateway concern. The stdio transport is local and is never authenticated.

For configuration and the gateway boundary, see Authenticate RedisVL MCP.

Tool Surface

RedisVL MCP exposes up to three built-in tools, plus any configured custom tool profiles:

• list-indexes enumerates the configured logical indexes for discovery
• search-records searches a selected index using that index’s server-owned search mode
• upsert-records validates and upserts records into a selected writable index, embedding them only when that capability is configured

Any of the three can be turned off with server.builtin_tools, independently of whether custom tools are configured — useful for a server that should only ever read, or one that serves nothing but curated profiles:

server:
  builtin_tools:
    upsert-records: disabled

Only the three names above are accepted; anything else fails at startup rather than being silently ignored.

Disabling a built-in adjusts what the rest of the surface advertises, so the published contract never points at something the server withholds:

• list-indexes reports upsert_available: false for every binding when upsert-records is disabled, since a writable binding still cannot be written to through a tool that is not published.
• On a multi-index server with list-indexes disabled, every tool that requires an index — search-records and upsert-records alike — names the available index ids in its own description instead of deferring to a discovery tool that does not exist. That server still logs a startup warning naming the affected tools, because inlining the ids is a fallback rather than an endorsement of the shape.

A server whose tool set ends up unusable — no tools at all, or discovery disabled on a multi-index server — logs a warning at startup.

Tools register once per process. builtin_tools is re-read on restart, but the registered tool set is not rebuilt, so a stop/start against an edited config keeps the previous tools and logs a warning saying so. Start a new process to change the tool surface.

These built-in tools follow a stable contract (profiles differ where noted in Custom Tool Profiles — notably they accept the object filter form only):

• request validation happens before query or write execution
• the resolved logical index is echoed in every search-records and upsert-records response
• filters support either raw strings or a RedisVL-backed JSON DSL
• on a single-index server, search-records describes the inspected schema by advertising typed JSON DSL filter fields, object-filter exists support, and valid return_fields; on a multi-index server those hints are ambiguous, so the description instead directs clients to call list-indexes and pass index
• error codes are mapped into a stable set of MCP-facing categories

list-indexes

list-indexes returns one entry per configured binding so clients can route subsequent calls. Each entry reports:

• the logical id
• an optional description (only when configured)
• upsert_available, reflecting the binding’s effective write availability
• fields, the filterable fields discovered from the index
• limits, only the runtime limits that were explicitly configured

The discovery payload is deliberately minimal:

• the underlying Redis index name (redis_name) is never exposed
• the vector field and the configured embed-source text field are omitted from fields, since they are implementation inputs rather than fields a client filters on
• limits shows only explicitly set values (such as max_limit or max_upsert_records); defaults are not echoed

Custom Tool Profiles

The built-in tools expose the index generically, which leaves the model doing query engineering on every call: pick the index, understand the schema, build a filter, choose return fields. A profile moves those decisions into config. It is search-records with some arguments pre-filled and frozen and the rest still exposed, published under a name and description of your choosing.

Profiles are pure configuration. You add a custom_tools entry to the same YAML the server already loads and restart it; there is no Python to write.

custom_tools:
  - name: search-support-tickets
    based_on: search-records
    index: support_tickets
    description: >
      Search historical customer support tickets by semantic similarity.
      Use this to find prior resolutions for a customer problem.
    lock:
      return_fields: [subject, resolution, created_at]
      filter: { field: status, op: eq, value: resolved }
    params:
      limit: { expose: true, max: 20 }
      filter: { expose: true }

lock holds what the author decides; params holds what the model may still pass. Anything not listed in params stays exposed, so a profile that only locks a filter keeps the rest of the built-in’s contract. index is pinned by the top-level index: key rather than exposed as a param, and may be omitted only when exactly one index is configured.

params.limit.max bounds the result count. It applies whether the model names a limit or leaves it out — an omitted limit is capped rather than falling through to the binding’s default. An explicit request above the cap is rejected. When limit is hidden (expose: false), the cap becomes the fixed result count instead. The cap must not exceed the binding’s own runtime.max_limit, which is checked at startup. Note that it bounds page size, not total reachable data: offset is a separate argument, so paging is still possible up to the binding’s max_result_window.

suppress_schema_hints: true drops the auto-generated field hints from the tool description. Those hints enumerate filterable and returnable fields, which is noise once those arguments are locked — and by default they are appended only for arguments the model can still use.

Two things about filters are easy to conflate:

• lock.filter is an ordinary expression in the JSON filter DSL. Its and/or/not operators describe the locked filter’s own content.
• How the locked filter combines with a model-supplied one is fixed and not configurable: the executed query is always locked AND caller. A compound model-supplied expression renders parenthesized, so its or/not nests inside the locked AND and cannot reach the top level — the model can only narrow within the locked scope and never widen past it.

For that reason a profile accepts only the object form of a filter from the model. A raw filter string is rejected both by the advertised schema and by the tool itself, because strings bypass the DSL’s field validation and have no safe composition with a locked expression.

Structure is only half of it: the nesting guarantee holds only while every filter value stays inside its own clause. Text eq/ne values are handled by the Text filter itself, which renders them as a quoted phrase with any " or \ replaced by a space, so a parenthesis or a | the value carries is literal text rather than syntax. Text like values are patterns, so the library leaves them raw and this boundary escapes them instead — the delimiters that would close the clause are escaped, the pattern metacharacters are not. Tag values have their delimiters escaped and numeric values are type-checked. A caller filter that still renders as something able to break out is refused rather than combined.

Profiles resolve to a built-in call and nothing more, so they inherit the concurrency cap, request timeout, read-only policy, auth scoping, and error mapping already applied to search-records.

Because adding near-duplicate tools makes tool selection harder rather than easier, built-ins that curated profiles supersede can be turned off with server.builtin_tools (see Tool Surface).

Misconfiguration fails at startup rather than at the first call. Among the checks: a name colliding with a built-in or using a reserved redisvl-/redisvl_ prefix; a duplicate tool name; a missing or unknown index; a params key that is not a real argument; max on anything but limit, or a cap above the binding’s max_limit; hiding query; locking return_fields while also exposing them; and a locked filter or projection naming a field the bound index does not have. Unrecognized keys are rejected too, so a typo in lock fails loudly instead of silently producing a tool that reads as locked but enforces nothing.

Tenant Scoping From Token Claims

When several tenants share one index, separated by a field such as org_id, exposing search-records makes the tenant boundary depend on the model remembering to pass a filter. One forgotten filter is a cross-tenant read. A profile can remove that knob: lock.inject reads the tenant from the caller’s verified token and AND-combines it into every query the profile runs.

custom_tools:
  - name: search-customer-kb
    index: customer_kb
    description: Search this customer's knowledge base.
    lock:
      inject:
        - field: org_id                      # the tenant field in the index schema
          from: claim                        # the value comes from the verified token
          claim: "https://acme.example/org"  # the claim name your identity provider emits
          required: true

field and claim are independent names: field is what the index schema calls the tenant column, and claim is what the identity provider calls it. A token carrying "https://acme.example/org": "acme" makes every query from that caller run as @org_id:{acme} AND <everything else>.

The model cannot set the injected value. The field is absent from the tool’s input schema, a call that names it is rejected, and it is left out of the field hints appended to the tool description. It is not hidden outright: unless lock.return_fields excludes it, results carry the field, and list-indexes describes the whole schema. What the model sees there is only ever its own tenant. If the model filters on the field, its clause ANDs with the injected one, so it can narrow within its own tenant and naming another tenant matches nothing. The rest of the profile works as before, so a static lock.filter on another field and a model-supplied filter both still apply.

What Counts as a Usable Claim

The claim must be a single, non-empty string. Anything else refuses the request with a forbidden error, and no query runs:

A | inside a single string is accepted. It is escaped, so an identifier such as the Auth0 subject auth0|64f1c2 matches only its own documents.

A list is refused because of its type, not because of what it renders as. The union it produces is indistinguishable from one a caller could legitimately ask for, so no inspection of the finished query could catch it.

With several inject entries, every entry ANDs into the query, and one unusable claim refuses the whole request rather than narrowing by the entries that did resolve.

What Fails at Startup

Injection is checked at startup wherever the configuration alone can show it would not hold:

• authentication is not enabled, on any transport, including an unauthenticated loopback HTTP bind and any --allow-unauthenticated bind;
• authentication is configured but the server runs over stdio, which is never authenticated (checked when the server starts through rvl mcp or run_async; an embedder that calls startup() directly is not, and every call is then refused at request time instead);
• the Redis index is also reachable without the tenant scope: through search-records, through upsert-records unless every binding over it is read-only, or through another custom tool on it that does not inject exactly the same entries, since a tool scoped by another field, or by the same field from another claim, reads across the tenants this one separates. Bindings that share a redis_name count as one index;
• the injected field is absent from the bound index, is not a tag field, or is declared NOINDEX;
• an inject list is empty, names one field twice, or names a field that lock.filter also constrains;
• required is anything but true, or from is anything but claim.

An injected field must be a tag. Text equality is a phrase match over tokenised text, and text is tokenised on punctuation, so the phrase acme-corp would also match acme-corp-eu.

The tool set registers once per process. If a restart reloads a configuration that differs from the registered one, the server normally logs a warning and keeps the old tools. When injection is configured on either side of the change, startup fails instead, because keeping the old tenant scoping in force is not something a log line should report.

Threat Model

The guarantee is narrow: a client presenting a validly signed token cannot make the model widen or escape the tenant scope carried in that token. The trust boundary is the identity provider, not the MCP client, so the guarantee holds only while these hold:

• The token is genuinely verified. Use a real signing key and an asymmetric algorithm. The server refuses to start an injecting profile without authentication, but it does not check which algorithm you configured.
• The identity provider assigns the claim. If a tenant can mint its own token, or set the claim itself, nothing here stops it reading another tenant’s data.
• The config names every route to the tenant data. The startup check compares bindings by redis_name, so it does not see an index alias, or a second index built over the same key prefix. Either one is another route to the same documents, and a tool on it needs the same lock.inject.
• Only trusted ingestion writes the index. The server refuses upsert-records on a writable scoped index, because a write can retag another tenant’s document as the writer’s own. Whatever loads documents outside the server is inside the trust boundary.
• Every document carries exactly its tenant. Stamp the tenant field on each document, indexed as a tag, with the identifier exactly as the identity provider emits it. Redis normalises the stored value, not the claim: it splits it on the field’s separator (, by default), so a document stamped acme,victim belongs to both tenants; it trims surrounding whitespace; and on JSON storage it indexes every element of an array. A document without the field matches no tenant, so it is invisible rather than shared.
• Tenant identifiers differ by more than case. Tag fields fold case unless declared CASESENSITIVE, and the folding is Unicode-wide: Acme and acme are one tenant, and so are a Kelvin sign and K, or composed and decomposed forms of an accented letter. The server warns at startup when an injected field is not case-sensitive.

Where tenants share one index, the injected filter is the only isolation boundary. There is no Redis ACL or keyspace separation behind it, so a defect in filter combination or claim validation is a full cross-tenant read. If you need defence in depth, separate tenants at the Redis layer as well.

Listing the tenant claim under auth.required_claims is a cheap outer layer: the verifier then rejects a token that lacks the claim before any tool runs. That check confirms only that the claim is present. Its value is still validated by the profile on every call.

Why Use MCP Instead of Direct RedisVL Calls

Use RedisVL MCP when you want a standard tool boundary for agent frameworks or assistants that already speak MCP.

Use direct RedisVL client code when your application should own index lifecycle, search construction, data loading, or richer RedisVL features directly in Python.

RedisVL MCP is a good fit when:

• multiple assistants should share one approved retrieval surface
• you want search behavior fixed by deployment config
• you need a read-only or tightly controlled write boundary
• you want to reuse an existing Redis index without rebuilding retrieval logic in every client

For setup steps, config, commands, and examples, see Run RedisVL MCP.