Redra

Redra MCP

Redra is an open-source Model Context Protocol server that searches current class-action settlement records. It works with any MCP client and supports two deployment choices:

1. Use the hosted Redra MCP and dataset (recommended).
2. Self-host both the MCP and a local copy of Redra's published dataset.

Redra identifies possible keyword matches. It does not determine legal eligibility, submit claims, or provide legal advice.

Data source and license

During the technical beta, local mode defaults to the public SettleSignal JSON feed at https://settlesignal.com/data/settlements.json. SettleSignal identifies its Hugging Face dataset card as an official profile, and that card licenses the feed's public fields under CC BY 4.0. Redra preserves source attribution and links while normalizing fields and adding derived lifecycle and quality metadata.

Redra also supports its independently assembled, content-addressed open and upcoming publication format. That publication remains in a gated shadow beta; it is not the fresh-install default until its launch checks pass. Every imported independent record must have independent discovery lineage and Tier A or B provenance, and a failed validation leaves the prior local snapshot untouched.

open records have a currently usable filing endpoint. upcoming records require evidence of a future claim window but have no current claim form. They are exposed only as a watchlist and are excluded from the open count and claimable-money total. The normalized factual compilation is published under CC BY 4.0. Search results preserve official source links and carry source and change attribution. See NOTICE.md.

The provider is explicit and is never inferred from a downloaded payload. Set REDRA_DATASET_PROVIDER=independent to consume Redra's publication, or leave it as settlesignal during the transition. REDRA_DATASET_URL can override the selected provider's default URL.

No settlement data is committed to this repository.

Hosted service (public preview)

The project currently operates this optional public service:

• Website: https://redra.ai/
• Documentation: https://redra.ai/docs
• Hosted MCP: https://mcp.redra.ai/mcp
• Privacy policy: https://redra.ai/privacy

The operational database behind the hosted MCP is private, has no public Fly route, and is reachable only by the hosted MCP over authenticated internal networking. The MCP currently requires no user API key. It does not impose a per-IP quota, because hosted AI clients can send many unrelated users through shared network addresses. Availability is provided on a best-effort basis without an uptime, completeness, or fitness-for-eligibility guarantee.

The Redra application does not persist prompts, profile data, or MCP tool arguments. The hosted MCP receives only tool arguments, not an agent's complete conversation. Hosting infrastructure may retain ordinary operational metadata, such as timestamps and request paths, in short-lived access logs. The hosted MCP stores each network-source IP address with a cumulative count of MCP work units for aggregate usage measurement; items in a batch count individually. A network source may be shared provider infrastructure and is not treated as a Redra user identity or quota key. Search terms and tool arguments are not associated with that counter. Self-hosted Redra does not enable this hosted usage counter.

Option 1: hosted MCP and dataset (recommended)

Connect a Streamable HTTP MCP client to the URL published by the Redra service, currently:

https://mcp.redra.ai/mcp

No local process or dataset is needed.

Option 2: self-hosted MCP and dataset

Python 3.11 or newer is required. Clone and install Redra:

git clone https://github.com/sacrosaunt/Redra.gitcd Redrapython -m venv .venvsource .venv/bin/activatepip install -e .

The project pins the current stable v1 line of the official Python MCP SDK with mcp>=1.27,<2. Version 2 was still prerelease when this release was built.

Download the settlement dataset and start the local MCP:

export REDRA_DATABASE_PATH="$PWD/data/settlements.db"redra-mcp dataset updateredra-mcp serve --transport stdio

The update is transactional: a failed download or invalid source payload leaves the previous SQLite dataset intact. To keep a local dataset fresh automatically:

redra-mcp dataset watch --interval-hours 24

This performs an update immediately and then once every 24 hours. It is optional; self-hosters can instead invoke dataset update from their existing cron system.

To expose the local MCP over Streamable HTTP:

export REDRA_PORT=8000redra-mcp serve --transport streamable-http

The MCP endpoint is /mcp and binds to 127.0.0.1 by default.

Network safety: Redra's HTTP transport does not provide authentication or TLS. Do not bind it to 0.0.0.0 or publish port 8000 directly unless access is protected by a trusted authenticated reverse proxy, firewall, or private network.

Docker

Start the server and its once-daily updater:

docker compose up --build

The updater and MCP share the SQLite volume. The MCP remains available if a refresh fails, and the previous valid snapshot stays in place. Compose publishes the MCP only on the host loopback interface at 127.0.0.1:8000.

MCP tools

• search_settlements: keyword and structured-filter search.
• search_settlements_batch: runs up to 50 independent searches in one MCP call without changing the logical-AND meaning of each query's keywords. Records are deduplicated across queries, and matched_query_indices preserves every search angle that found each record.
• get_settlement: fetches one record and its official source links, with agent guidance to verify the complete eligibility terms when browsing is available.
• get_settlements: fetches up to 20 finalist records together before the agent composes its answer.
• get_dataset_info: reports source, license, freshness, and counts.

The hosted provider also returns its cached claimable-money headline in dataset info. It is calculated only from records explicitly marked for inclusion in the claimable total. Upcoming records never contribute. Self-hosted providers may return additional provider-specific aggregate metrics in the same extra object.

Search behavior

Every item in keywords is required (logical AND). Agents should use independent query objects in search_settlements_batch for unrelated companies, products, or alternative terms. This preserves each query's meaning while avoiding a long series of MCP round trips.

Batch search caps the unique records returned across all queries with max_total_results, which defaults to 50 and may be set as high as 100. The cap does not reduce the number of searches performed. Query summaries report each query's total match count, sampled count, returned IDs, and omitted sampled count; the unique records appear once in the top-level items list.

For broad eligibility scans, the MCP directs agents to search expansively across plausible brands, aliases, parent companies, subsidiaries, services, purchases, fees, and incidents recalled from the context the client makes available. It also considers eligibility-relevant demographic angles such as state, age group, occupation, student or veteran status, parent or guardian status, and housing or household situation.

Recalled or inferred associations are query candidates, not facts or eligibility decisions, and queries must not contain identifying information. A category-level or demographic association alone is not enough to call a result a match.

For each plausible result, agents with web access are directed to inspect the official settlement or claim page for the class definition, qualifying dates, products or services, geographic limits, exclusions, proof requirements, and deadline. Agents should compare only confirmed terms with context the user has made available. Before composing the final answer, agents retrieve the complete stored record for every finalist with get_settlement or get_settlements. Plausible results are presented as concise lead cards showing why each surfaced, confirmed matching terms, important unknowns, deadline, payout, proof requirement, and the official link when those fields are available. Evidence strength is ranked before urgency, with labels such as Notice found, Strong possible match, and Needs your confirmation; imminent deadlines remain prominent.

A direct notice is strong evidence but does not prove every class condition. Agents ask a focused non-sensitive follow-up before recommending action when one unknown could change the recommendation. A candidate mentioned during progress receives an explicit final disposition, and purely category-level or demographic candidates are kept out of the action list unless the user requests an exhaustive trace.

Agents search and investigate before asking for more information. When non-sensitive answers would materially clarify a lead, they prefer the client's native structured-question or question-card interface. Clients without that interface fall back to asking the same focused questions directly in chat. Questions should explain why the answer matters, offer a Not sure or Skip choice where appropriate, and never request identifying or sensitive information.

Agents without browsing must say that they could not independently verify the complete terms, avoid guessing, and direct the user to the official link.

Claim status defaults to open. The public MCP supports only open and upcoming. If a broad open scan produces no credible leads or only weak leads, agents may run focused upcoming searches for the strongest user-specific angles. Upcoming records are not yet claimable, are presented as a separate watchlist, and are excluded from current claim counts and money totals. Historical, payment, and uncertain lifecycle states remain internal rather than being exposed as public search options.

The provider's opaque verification tier is not a search filter. It is retained in results as source_verification_status for transparency. Redra also returns objective source_kind, source_checked_at, claim_url_available, claimability, and quality_flags. A source label such as Verified record is converted to an open or closed lifecycle when its deadline makes that determination possible.

Use the exact settlement_type filter instead of keywords when searching by type: class_action_settlement, consumer_product_settlement, data_breach_settlement, financial_fee_settlement, government_refund, other_consumer_compensation, privacy_settlement, regulatory_compensation_program, or state_ag_refund. Keywords intentionally do not search the settlement_type field.

Other filters include state, proof requirement, deadline range, and result limit. An empty applicable_states list is treated as nationwide.

Environment variables

Per-IP rate limiting is disabled by default. It should not be used as a per-user quota behind ChatGPT, Claude, or another hosted client because unrelated users may share provider egress addresses. Operators that understand their direct network topology can still enable it explicitly. Arbitrary X-Forwarded-For values are never trusted. Local stdio mode is never rate-limited.

Only enable REDRA_TRUST_FLY_HEADERS when the process is actually behind Fly's trusted proxy path. Enabling it behind an arbitrary proxy can allow clients to spoof the address used for rate limiting.

The hosted MCP instead uses bounded request bodies, a concurrency cap, a short bounded cache for identical normalized searches, and a centralized emergency ceiling on aggregate work units. Batch items count individually toward that ceiling. Cache keys are per-process keyed digests and cached values contain public settlement results rather than submitted query arguments. The emergency ceiling is a service-wide circuit breaker, not a user quota. Its dataset service is not exposed to public callers and uses separate internal concurrency protection.

Privacy and safety

Do not submit names, email addresses, street addresses, account numbers, health details, claim identifiers, or other identifying information. Search with brand, product, employer, service, incident, date-range, and state information. Always direct users to the official administrator before they provide claim information. The hosted service is designed not to persist prompts, profiles, or tool arguments, but self-host the MCP and dataset when control over the complete request path is required.

Development

pip install -e '.[dev]'pytest

Please report vulnerabilities privately as described in SECURITY.md.