Enterprise documentation

The credential

Enterprise uses its own key. It begins with ld_agent_. Send it as Authorization: Bearer ld_agent_… or as the X-API-Key header. A regular LawDiver API key, and the sign-in used by the conversational connector at /products/mcp, are rejected here.

The endpoint URL is issued with the credential and is not published on this site. The samples below write it as <enterprise endpoint URL>, or $LAWDIVER_ENTERPRISE_URL in a shell. Transport is Streamable HTTP. Each tool returns one JSON text block. A failed tool returns that same text channel with an error flag and a short correction, not a stack trace. A missing key is HTTP 401. A suspended account is HTTP 403. More than 2,000 calls in a minute is HTTP 429 with Retry-After.

GET <enterprise endpoint URL>/info and GET <enterprise endpoint URL>/toolspec.json list the server and the tool names without a key. Calling a tool requires the key.

Claude Managed Agents

Anthropic splits this into two steps. The agent definition names the server and does not contain the key. A vault holds the key, and each session points at that vault. The vault URL must be the same path as the agent URL. Host casing and a trailing slash are ignored. A different path is a different server, and the session then connects with no credential.

Create the agent with the Managed Agents beta header managed-agents-2026-04-01. Every declared server needs a matching mcp_toolset. Use a model your account is allowed to call. The toolset defaults to always_ask, which waits for approval before each call. Set the toolset permission policy when the agent should query without a person confirming each call.

curl https://api.anthropic.com/v1/agents \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "anthropic-beta: managed-agents-2026-04-01" \
  -H "content-type: application/json" \
  -d '{
    "name": "LawDiver research",
    "model": "claude-opus-5-5",
    "mcp_servers": [
      {
        "type": "url",
        "name": "lawdiver",
        "url": "<enterprise endpoint URL>"
      }
    ],
    "tools": [
      { "type": "agent_toolset_20260401" },
      { "type": "mcp_toolset", "mcp_server_name": "lawdiver" }
    ]
  }'

Register the key in a vault. mcp_server_url is the Enterprise URL, and token is the ld_agent_ key with no Bearer prefix. Anthropic adds the scheme when it connects.

curl "https://api.anthropic.com/v1/vaults/$VAULT_ID/credentials" \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "anthropic-beta: managed-agents-2026-04-01" \
  -H "content-type: application/json" \
  -d '{
    "display_name": "LawDiver Enterprise",
    "auth": {
      "type": "static_bearer",
      "mcp_server_url": "<enterprise endpoint URL>",
      "token": "ld_agent_…"
    }
  }'

Start a session with POST https://api.anthropic.com/v1/sessions and the same headers. The body names the agent, the environment, and the vault. A limited environment blocks MCP unless networking sets allow_mcp_servers: true or lists the endpoint’s host in allowed_hosts.

{
  "agent": "agent_…",
  "environment_id": "env_…",
  "vault_ids": ["vault_…"]
}

Session creation does not test the key. If the key is wrong, the session still starts and a session.error event arrives with mcp_authentication_failed_error and mcp_server_name lawdiver. A network failure is mcp_connection_failed_error. Tool output longer than about 100,000 characters is written to a file in the agent workspace. The model gets a preview and the path.

Tell the agent to call describe_schema before the first SQL, to count and join with query_corpus, and to open the opinions behind a number with search_cases and get_opinion_text. Cite only authorities those tools returned.

Other clients

Any client that speaks Streamable HTTP MCP uses the same URL and the same bearer key. The conversational Claude.ai connector is a different server. Do not paste an Enterprise key into that connector, and do not point a Managed Agent at https://lawdiver.com/mcp.

Claude Code

Add the server once. The header value includes the word Bearer and a space.

claude mcp add --transport http lawdiver-enterprise \
  $LAWDIVER_ENTERPRISE_URL \
  --header "Authorization: Bearer ld_agent_…"

The same entry in a project .mcp.json can read the URL and the key from the environment:

{
  "mcpServers": {
    "lawdiver-enterprise": {
      "type": "http",
      "url": "${LAWDIVER_ENTERPRISE_URL}",
      "headers": {
        "Authorization": "Bearer ${LAWDIVER_ENTERPRISE_KEY}"
      }
    }
  }
}

OpenAI

The Responses API takes a remote MCP tool. Send the header on every request. OpenAI does not store it.

{
  "type": "mcp",
  "server_label": "lawdiver",
  "server_url": "<enterprise endpoint URL>",
  "headers": {
    "Authorization": "Bearer ld_agent_…"
  },
  "require_approval": "never"
}

An application you run

Initialize, list tools, then call one. The body is JSON-RPC. Accept should include both application/json and text/event-stream.

curl -X POST $LAWDIVER_ENTERPRISE_URL \
  -H "Authorization: Bearer ld_agent_…" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "tools/call",
    "params": {
      "name": "describe_schema",
      "arguments": {}
    }
  }'

What every call shares

Every tool is read-only. A SQL statement may run for up to five minutes. If it does not finish, the error says to filter on opinion_id, court_id, or year and try again. Counts, sums, and other bigint or numeric columns come back as numbers when they fit in a safe integer; larger integers stay strings. columns is filled from the result metadata, so an empty result still names its columns. Long-running calls stream: the connection carries keep-alive comments while the statement runs, so a client that reads server-sent events sees no idle gap.

Case rankings and citation lookup use the decisions table. A citing-side join uses opinions, because a concurrence and a dissent are their own rows. citation_key is the lookup form of a reporter cite: 556 U.S. 662 is 556us662. Filter on that column. The display citation is for reading the result.

describe_schema

Call this before the first query_corpus. It returns every published table, with each column’s type, description, and example values.

The result is { tables: [{ name, oneRowPer, columns: [{ name, type, description, samples }] }] }. An unknown table name is an error that lists the real names.

query_corpus

One PostgreSQL SELECT or WITH … SELECT. Joins, subqueries, UNION, GROUP BY, aggregates, and window functions are allowed. The agent is expected to issue many small queries: a count, then a filter, then a lookup of the rows behind the count. A worked example, twelve circuits and four motions in one call, is on /products/enterprise/example.

The statement is rejected when it:

• is not a single SELECT or WITH … SELECT
• writes, or uses SELECT INTO, catalog tables, pg_ functions, comments, dollar-quoting, or quoted identifiers
• qualifies a table with a schema name — use the table names from describe_schema
• joins case_names or reporter_cites to the case tables — query those two on their own
• names a table that is not in the model — the error suggests the closest published name

A bad column comes back as text the agent can rerun, for example column cite does not exist; did you mean citation on decisions?

A successful result:

{
  "columns": ["year", "circuit", "opinions"],
  "rows": [{ "year": 2010, "circuit": "11", "opinions": 42 }],
  "rowCount": 1,
  "truncated": false,
  "elapsedMs": 840
}

truncated: true means more rows matched than maxRows. Raise the cap, add a filter, or call export_results for a file.

SELECT o.year, c.circuit, count(DISTINCT o.opinion_id) AS opinions
FROM citations x
JOIN opinions o ON o.opinion_id = x.citing_opinion_id
JOIN courts c ON c.court_id = o.court_id
WHERE x.cited_opinion_id = (
  SELECT opinion_id FROM decisions WHERE citation_key = '556us662'
)
AND o.year >= 2010
GROUP BY o.year, c.circuit
ORDER BY o.year, c.circuit

The published tables

opinion_id is the LawDiver id of an opinion. decision_id is the LawDiver id of the decision, shared by the majority, concurrences, and dissents. describe_schema returns this same list, including types and example values. Query case_names and reporter_cites in their own statements. Filter opinion_text by opinion_id.

export_results

Runs a query or a search and returns a signed download URL. The file lives for one hour. Supply sql or search. Supplying neither is an error.

The result:

{
  "url": "https://…",
  "rows": 1200,
  "columns": ["year", "circuit", "opinions"],
  "bytes": 18432,
  "truncated": false,
  "filename": "iqbal-by-circuit.csv"
}

search_cases

Federal and state case search. Jurisdiction is required. This is the tool for finding opinions. Counts, joins, and rankings belong in query_corpus.

Jurisdiction type is one of us_supreme_court, all_federal, federal_circuit, federal_district, one_state, one_state_plus_federal, all_states, all_states_and_federal. state is a two-letter USPS code when the type needs a state. circuit is 1–11, dc, or federal.

Each hit:

{
  "caseId": "109202100001835",
  "decisionId": "5109202100001835",
  "caseName": "Ashcroft v. Iqbal",
  "citation": "556 U.S. 662",
  "bluebookCitation": "Ashcroft v. Iqbal, 556 U.S. 662 (2009)",
  "parallelCitations": ["129 S. Ct. 1937"],
  "court": "Supreme Court of the United States",
  "courtAbbrev": "U.S.",
  "jurisdiction": "federal",
  "level": "supreme",
  "date": "2009-05-18",
  "year": 2009,
  "docketNumber": "07-1015",
  "published": true,
  "opinionType": "majority",
  "snippet": "…",
  "goodLaw": {
    "status": "good",
    "statusName": "good",
    "negative": false,
    "unknown": false,
    "negativeTreatmentCount": 0,
    "basis": null
  },
  "citedByCount": 18000,
  "topics": ["pleadings", "motions to dismiss"]
}

caseId is the LawDiver opinion id. decisionId is the LawDiver decision id. level is supreme, appellate, trial, bankruptcy, or specialty. goodLaw.status is good, caution, negative, no_negative, or unknown, the same labels as decisions.good_law_status; goodLaw.statusName is the finer corpus name (good, limited, negative_below_threshold, overruled, overruled_in_part, vacated, vacated_in_part, vacated_and_reversed, vacated_and_reversed_in_part, superseded, abrogated, questioned, or unknown). topics uses the same names as topics.name. The wrapper is { results, total, filteredOnPage?, limit, offset, note? }.

total is the number of hits after every filter when the tool knows it. With courts, courtLevel, judge, or minCitedBy, the response carries filteredOnPage: true; total is the filtered count when the ranked list was read to its end, and null with an explanatory note when the tool stopped after ten pages. Page with offset either way. With citesCase, total is the count of citing cases that pass the filters.

retrieve_case

Resolves a citation or a case name to one LawDiver case. Use the returned caseId with get_case, get_opinion_text, and the good-law tools.

Four shapes come back. status: found includes caseId, decisionId, caseName, citation, parallelCitations, court, date, goodLaw, citedByCount, and topics. A lookup by caseId also includes docketNumber. status: did_you_mean includes up to eight candidates, each with caseId, caseName, citation, court, and year. Call again with one of those ids. status: not_found includes the query and an explanation; a lone surname such as Iqbal is not enough to identify a case, so supply both parties or a reporter citation. status: error with retryable: true means the lookup hit its deadline and the case was not ruled out; retry.

get_case

The case card for a known LawDiver id. One argument, caseId.

Found: caseId, decisionId, caseName, citation, bluebookCitation, parallelCitations, court, courtAbbreviation, jurisdiction, date, year, docketNumber, published, goodLaw, citedByCount, topics. Missing: { status: "not_found", caseId }.

get_opinion_text

Opinion text, in pages. The body is plain text.

Found:

{
  "caseId": "109202100001835",
  "format": "text",
  "section": "all",
  "text": "…",
  "offset": 0,
  "nextOffset": 100000,
  "truncated": true
}

truncated: false and a null nextOffset mean this page is the end. Missing: { status: "not_found", caseId }.

get_good_law

The recorded good-law status for one case, the basis text, and the negative treatments that support that status. One argument, caseId.

Found:

{
  "caseId": "109202100001835",
  "status": "good",
  "negative": false,
  "unknown": false,
  "negativeTreatmentCount": 0,
  "basis": null,
  "negativeCitations": []
}

status is good, caution, negative, no_negative, or unknown on search hits, and the recorded treatment name on this call: good, no_negative, overruled, overruled_in_part, vacated, vacated_in_part, vacated_and_reversed, vacated_and_reversed_in_part, superseded, abrogated, questioned, limited, or negative_below_threshold. negative is true for overruled, overruled_in_part, vacated, vacated_in_part, vacated_and_reversed, vacated_and_reversed_in_part, superseded, abrogated, and questioned. Each negativeCitations row has citingCaseId, caseName, date, treatment, and rationale. citingCaseId is a LawDiver opinion id. Missing: { status: "not_found", caseId }.

goodlaw_check

The citing-case list. Rankings that need the whole graph, including most-cited and citation depth, belong in query_corpus on citations.

Found: { caseId, total, scanned, truncated, citingCases, note? }. total is the cited-by count. scanned is how many citing edges were read, newest first; truncated is true only for limit: unlimited on a case with more edges than the scan cap. If the database budget runs out the response is { status: "timeout", caseId, retryable: true, message }; retry with a smaller limit or use query_corpus. Each citing case has citingCaseId, caseName, citation, date, treatment, and pin. citingCaseId is a LawDiver opinion id. treatment is the treatment label, or not adverse when none was written. pin is reversal, mostRecent, or null. Negative order uses overruled, reversed, vacated, abrogated, superseded, overruled_in_part, questioned, criticized, limited, and distinguished.

check_citation

Verifies one citation, a batch, or citation-like lines in a passage. Supply one of citation, citations, or text. Supplying none is an error.

The result is { results: [...] }. Each row:

{
  "citation": "556 U.S. 662",
  "status": "verified",
  "normalized": "556 U.S. 662",
  "verdict": "valid",
  "explanation": "…",
  "caseId": "109202100001835",
  "caseName": "Ashcroft v. Iqbal",
  "parallelCitations": [],
  "goodLaw": { "status": "good", "statusName": "good", "negative": false, "unknown": false, "negativeTreatmentCount": 0, "basis": null },
  "pinCiteChecked": true
}

status is verified when the verdict is valid or section_verified, corrected when it is likely_valid, page_mismatch, or name_mismatch, error with retryable: true when the resolver hit its deadline before it could rule the citation in or out, and not_found otherwise. normalized is the corrected form when the resolver has one. goodLaw.status is good, caution, negative, no_negative, or unknown; goodLaw.statusName is the finer corpus name. A constitutional section such as U.S. Const. amend. XIV, § 1 verifies at the amendment; the explanation says the section itself was not separately checked.

retrieve_statute

Statute, regulation, constitution, or court-rule text by Bluebook citation or by an authority key from an earlier check. Citing opinions are not part of this body. Count them with query_corpus on statute_citations, filtering code.

The result has status, message, query, citationAsWritten, statute, verdict, and note. status is ok, not_found, unavailable, or not_a_statute. On ok, statute includes authorityKey, kind (statute, rule, constitution, or regulation), bluebook, heading, body, bodyChars, edition, editionYear, officialUrl, repealed, jurisdiction, and section. body is null when the section was confirmed and the full text is not held.

list_jurisdictions

No arguments. Returns the codes search_cases accepts, and the same circuit and court-level words used on the courts table.

{
  "types": ["all_states", "all_states_and_federal", "all_federal", "one_state", "one_state_plus_federal", "federal_circuit", "federal_district", "us_supreme_court"],
  "circuits": ["1", "2", "3", "4", "5", "6", "7", "8", "9", "10", "11", "dc", "federal"],
  "stateCodes": ["AL", "AK"],
  "states": [{ "code": "FL", "name": "Florida" }],
  "courtLevels": ["supreme", "appellate", "trial", "bankruptcy", "specialty"],
  "examples": [
    { "type": "all_states_and_federal" },
    { "type": "one_state", "state": "FL" },
    { "type": "federal_circuit", "circuit": "11" },
    { "type": "us_supreme_court" }
  ]
}

stateCodes and states are the full lists. The snippet above shows the shape, not every state.