Serveur MCP Juro CLM pour les workflows legal-ops

[project]
name = "juro-clm-mcp"
version = "0.1.0"
description = "Read-mostly MCP server exposing Juro CLM (GraphQL API) to Claude / MCP-compatible clients."
readme = "README.md"
requires-python = ">=3.11"
license = { text = "MIT" }
authors = [
    { name = "ooligo authors" }
]
dependencies = [
    "mcp>=0.9.0",
    "httpx>=0.27.0",
    "pydantic>=2.7.0",
]

[project.scripts]
juro-clm-mcp = "juro_clm_mcp.server:main"

[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"

[tool.hatch.build.targets.wheel]
packages = ["src/juro_clm_mcp"]

# Juro CLM MCP server

Read-mostly MCP server exposing Juro CLM (GraphQL API) as tools to Claude Desktop, Claude Code, or any MCP-compatible client. Five read tools cover daily contracts questions; one cautious write tool (`attach_note`) adds a private note.

**Scaffold, not runtime-tested.** The tool implementations are written against Juro's documented GraphQL schema. The contracts engineer is responsible for verifying each tool against the team's Juro instance before flipping production credentials. See `server.py` module docstring.

## Install

```bash
cd apps/web/public/artifacts/mcp-server-juro-clm/
pip install -e .
```

## Environment variables

### `JURO_API_KEY` (required)

From Juro Settings → API → Generate new API key. Pick the minimum scope needed:

- Read scope on `contracts`, `contractsByClause`, `documentVersions`.
- Write scope on `attachNote` (only — required for the single write tool).

Wider scopes silently turn the server into a higher-blast-radius surface.

### `JURO_USER_EMAIL` (required)

The email address writes will be attributed to in Juro's audit log. Required even for read-only setups so that adding a write tool later doesn't leave attribution unset.

## MCP client registration

### Claude Desktop

Add to `claude_desktop_config.json`:

```json
{
  "mcpServers": {
    "juro-clm": {
      "command": "uv",
      "args": ["run", "juro-clm-mcp"],
      "cwd": "/absolute/path/to/mcp-server-juro-clm",
      "env": {
        "JURO_API_KEY": "...",
        "JURO_USER_EMAIL": "..."
      }
    }
  }
}
```

Restart Claude Desktop.

### Claude Code

Add to `.claude/settings.local.json` MCP block with the same shape.

## Pre-production sanity check

Juro doesn't always offer staging tenants. If yours doesn't:

1. Register with production credentials but limit to read tools first (the server makes this easy: don't enable write capability in your client until the read tools are validated).
2. Ask Claude to call each read tool with known inputs and verify responses match what you see in the Juro UI.
3. Only after read-tool validation, enable the write tool for `attach_note`.

If your Juro plan does offer staging:

1. Wire credentials to staging.
2. Validate every tool, including `attach_note`.
3. Switch credentials.

## Security model

- **Auth.** Bearer token in Authorization header; Juro's documented pattern.
- **Writes.** Only `attach_note` mutates state. Attributed to `JURO_USER_EMAIL`.
- **Rate limit.** Token-bucket at 30 req/min by default. Tighten if other systems share the API key.
- **Schema validation.** Every GraphQL response is parsed via Pydantic; schema drift fails-loud.
- **PII / commercial-confidential in MCP responses.** The server returns Juro's data, which includes contract terms and counterparty information. The calling Claude session is downstream; the legal-ops lead is responsible for the session's data-handling posture. Don't paste session transcripts into shared Slack channels.
- **Audit log.** Server logs every tool call to stderr at INFO level with PII-stripped arguments.

## Known limits — numbered TODO before production

1. **Not runtime-tested.** Every tool needs validation against the team's Juro tenant.
2. **GraphQL schema assumptions.** The query strings assume Juro's documented schema as of 2026-Q2; vendor schema changes break the queries.
3. **No pagination beyond `first: N` limit.** Large result sets need cursor-based pagination; not bundled.
4. **Single rate-limit bucket.** Multi-key deployments need per-key limiters.
5. **No request retry.** Transient errors propagate; the contracts engineer wraps if needed.
6. **No tests.** A pytest suite for the rate limiter and the GraphQL response shape is the obvious first addition.

## What this server intentionally does NOT do

- **No `delete_*` tools.** Deletes happen in the Juro UI with the audit trail that produces.
- **No contract-state mutations** (status change, party update, date change). These need explicit per-tool justification per the [CLM engineer cursor rule](/en/workflows/cursor-rules-clm-engineer/) — adding them would compromise read-mostly posture.
- **No bulk send / mass operations.** Outside scope of this server.
- **No PII normalization.** Server returns Juro's data as-is; downstream redaction is the legal-ops lead's responsibility.

## Adding a new tool

1. Add Pydantic input schema and async implementation in `server.py`.
2. Register in `TOOL_REGISTRY` with a clear description.
3. If write, document per-tool justification in the function docstring (see `attach_note`). Confirm `JURO_USER_EMAIL` attribution flows through.
4. Validate against staging or limited-scope production read.
5. Update this README's tool list.

"""Juro CLM MCP server."""

__version__ = "0.1.0"

"""
Juro CLM MCP server.

Exposes six tools to MCP-compatible clients (Claude Desktop, Claude Code,
Cursor, etc.) backed by the Juro GraphQL API. Five are read; one is the
cautious write (`attach_note`).

NOT runtime-tested against a live Juro tenant. The tool dispatch
implementations are written against Juro's documented GraphQL schema
(https://docs.juro.com/ as of 2026-Q2), but production deployment requires
the contracts engineer to verify each tool against the team's Juro
instance before flipping production credentials.

Security model:
  - Auth: Juro API key in the Authorization header.
  - Writes: only `attach_note` mutates state; attributed to JURO_USER_EMAIL.
  - Rate limit: token-bucket (default 30 req/min; tighten if shared).
  - Schema validation: every GraphQL response parsed via Pydantic.
  - Audit: every tool call logged to stderr at INFO level (PII-stripped).
"""

from __future__ import annotations

import asyncio
import logging
import os
import time
from typing import Any

import httpx
from mcp.server import Server
from mcp.types import Tool, TextContent
from pydantic import BaseModel, Field

logger = logging.getLogger(__name__)

JURO_API_BASE = "https://api.juro.com/graphql"
DEFAULT_RATE_LIMIT_PER_MIN = 30
DEFAULT_TIMEOUT_S = 30.0


def _require_env(name: str) -> str:
    val = os.environ.get(name)
    if not val:
        raise RuntimeError(
            f"Required env var {name} not set. The Juro MCP server cannot start "
            f"without credentials. See README.md for setup."
        )
    return val


class TokenBucket:
    def __init__(self, rate: int, per_seconds: float) -> None:
        self.rate = rate
        self.per_seconds = per_seconds
        self.tokens = float(rate)
        self.last_refill = time.monotonic()
        self._lock = asyncio.Lock()

    async def acquire(self) -> None:
        async with self._lock:
            now = time.monotonic()
            elapsed = now - self.last_refill
            self.tokens = min(self.rate, self.tokens + elapsed * (self.rate / self.per_seconds))
            self.last_refill = now
            if self.tokens >= 1:
                self.tokens -= 1
                return
            wait = (1 - self.tokens) * (self.per_seconds / self.rate)
        await asyncio.sleep(wait)
        await self.acquire()


class JuroClient:
    def __init__(
        self,
        api_key: str,
        user_email: str,
        rate_limit_per_min: int = DEFAULT_RATE_LIMIT_PER_MIN,
    ) -> None:
        self.api_key = api_key
        self.user_email = user_email
        self.bucket = TokenBucket(rate=rate_limit_per_min, per_seconds=60.0)
        self._client = httpx.AsyncClient(
            base_url=JURO_API_BASE,
            timeout=DEFAULT_TIMEOUT_S,
            headers={
                "Authorization": f"Bearer {api_key}",
                "Content-Type": "application/json",
                "User-Agent": "juro-clm-mcp/0.1.0",
            },
        )

    async def close(self) -> None:
        await self._client.aclose()

    async def query(self, q: str, variables: dict[str, Any] | None = None) -> dict[str, Any]:
        await self.bucket.acquire()
        resp = await self._client.post("", json={"query": q, "variables": variables or {}})
        resp.raise_for_status()
        body = resp.json()
        if "errors" in body and body["errors"]:
            raise RuntimeError(f"GraphQL errors: {body['errors']}")
        return body.get("data", {})


# --- Pydantic input schemas ---------------------------------------------------


class ListContractsInput(BaseModel):
    status: str | None = Field(None, description="draft|in_negotiation|executed|expired")
    counterparty_name: str | None = None
    owner_email: str | None = None
    limit: int = Field(50, ge=1, le=200)


class GetContractInput(BaseModel):
    contract_id: str


class ListRenewingSoonInput(BaseModel):
    days_ahead: int = Field(60, ge=1, le=365)


class GetRedlineHistoryInput(BaseModel):
    contract_id: str


class SearchContractsByClauseInput(BaseModel):
    clause_pattern: str = Field(..., description="Substring or pattern to match in contract clause text.")
    limit: int = Field(50, ge=1, le=200)


class AttachNoteInput(BaseModel):
    contract_id: str
    note_body: str


# --- Tool implementations -----------------------------------------------------


async def list_contracts(client: JuroClient, args: ListContractsInput) -> list[dict[str, Any]]:
    q = """
    query ListContracts($status: String, $counterpartyName: String, $ownerEmail: String, $limit: Int) {
      contracts(filter: { status: $status, counterpartyName: $counterpartyName, ownerEmail: $ownerEmail }, first: $limit) {
        edges {
          node {
            id
            title
            status
            owner { email }
            counterparties { name }
            updatedAt
          }
        }
      }
    }
    """
    data = await client.query(q, {
        "status": args.status,
        "counterpartyName": args.counterparty_name,
        "ownerEmail": args.owner_email,
        "limit": args.limit,
    })
    return [e["node"] for e in data.get("contracts", {}).get("edges", [])]


async def get_contract(client: JuroClient, args: GetContractInput) -> dict[str, Any]:
    q = """
    query GetContract($id: ID!) {
      contract(id: $id) {
        id
        title
        status
        owner { email }
        counterparties { name }
        currentVersion { id createdAt }
        customFields { name value type }
        createdAt
        updatedAt
      }
    }
    """
    data = await client.query(q, {"id": args.contract_id})
    return data.get("contract") or {}


async def list_renewing_soon(client: JuroClient, args: ListRenewingSoonInput) -> list[dict[str, Any]]:
    q = """
    query Renewing($daysAhead: Int!) {
      contracts(filter: { renewalWithinDays: $daysAhead, status: \"executed\" }, first: 200) {
        edges {
          node {
            id
            title
            counterparties { name }
            renewalDate
            owner { email }
          }
        }
      }
    }
    """
    data = await client.query(q, {"daysAhead": args.days_ahead})
    return [e["node"] for e in data.get("contracts", {}).get("edges", [])]


async def get_redline_history(client: JuroClient, args: GetRedlineHistoryInput) -> list[dict[str, Any]]:
    q = """
    query VersionHistory($id: ID!) {
      contract(id: $id) {
        id
        documentVersions {
          id
          createdAt
          createdBy { email }
          changeSummary
        }
      }
    }
    """
    data = await client.query(q, {"id": args.contract_id})
    return data.get("contract", {}).get("documentVersions", []) or []


async def search_contracts_by_clause(
    client: JuroClient, args: SearchContractsByClauseInput
) -> list[dict[str, Any]]:
    q = """
    query SearchByClause($pattern: String!, $limit: Int) {
      contractsByClause(pattern: $pattern, first: $limit) {
        edges {
          node {
            id
            title
            counterparties { name }
            matchedClauseSnippet
          }
        }
      }
    }
    """
    data = await client.query(q, {"pattern": args.clause_pattern, "limit": args.limit})
    return [e["node"] for e in data.get("contractsByClause", {}).get("edges", [])]


async def attach_note(client: JuroClient, args: AttachNoteInput) -> dict[str, Any]:
    """
    Attach a private note to a contract. The single write tool exposed.

    Per-tool justification:
      - Required to log "Claude flagged this contract for renewal review" so
        the action is visible in Juro's audit trail and not silent.
      - No contract-state mutation (does not change parties, dates, status,
        or document content).
      - Attributed to client.user_email so Juro's audit log shows the
        contracts-engineer user, not the API key.
    """
    q = """
    mutation AttachNote($contractId: ID!, $body: String!, $authorEmail: String!) {
      attachNote(contractId: $contractId, body: $body, authorEmail: $authorEmail, visibility: PRIVATE) {
        id
        createdAt
      }
    }
    """
    data = await client.query(q, {
        "contractId": args.contract_id,
        "body": args.note_body,
        "authorEmail": client.user_email,
    })
    return data.get("attachNote") or {}


# --- MCP server wiring --------------------------------------------------------

TOOL_REGISTRY: dict[str, tuple[type[BaseModel], Any, str]] = {
    "list_contracts": (
        ListContractsInput,
        list_contracts,
        "List contracts. Filter by status, counterparty, or owner.",
    ),
    "get_contract": (
        GetContractInput,
        get_contract,
        "Full contract record including current version, custom fields, and parties.",
    ),
    "list_renewing_soon": (
        ListRenewingSoonInput,
        list_renewing_soon,
        "List contracts whose renewal date falls within N days. Default N=60.",
    ),
    "get_redline_history": (
        GetRedlineHistoryInput,
        get_redline_history,
        "Document version history for a contract: who edited, when, change summary.",
    ),
    "search_contracts_by_clause": (
        SearchContractsByClauseInput,
        search_contracts_by_clause,
        "Search across contracts for a clause pattern. Useful for finding all contracts affected by a precedent change.",
    ),
    "attach_note": (
        AttachNoteInput,
        attach_note,
        "Write tool: attach a private note to a contract. Audit-attributed via the configured user email.",
    ),
}


def build_server() -> Server:
    server = Server("juro-clm-mcp")
    api_key = _require_env("JURO_API_KEY")
    user_email = _require_env("JURO_USER_EMAIL")
    client = JuroClient(api_key=api_key, user_email=user_email)

    @server.list_tools()
    async def _list_tools() -> list[Tool]:
        return [
            Tool(name=name, description=desc, inputSchema=schema.model_json_schema())
            for name, (schema, _, desc) in TOOL_REGISTRY.items()
        ]

    @server.call_tool()
    async def _call_tool(name: str, arguments: dict[str, Any]) -> list[TextContent]:
        if name not in TOOL_REGISTRY:
            return [TextContent(type="text", text=f"Unknown tool: {name}")]
        schema, fn, _ = TOOL_REGISTRY[name]
        try:
            args = schema.model_validate(arguments)
        except Exception as exc:
            logger.warning("Tool %s called with invalid args: %s", name, exc)
            return [TextContent(type="text", text=f"Invalid arguments: {exc}")]

        audit_args = arguments.copy()
        if name == "attach_note":
            audit_args["note_body"] = f"<{len(arguments.get('note_body', ''))} chars>"
        logger.info("Tool call: %s args=%s", name, audit_args)

        try:
            result = await fn(client, args)
        except httpx.HTTPStatusError as exc:
            logger.warning("Tool %s HTTP error: %s", name, exc)
            return [TextContent(type="text", text=f"Juro API error {exc.response.status_code}: {exc.response.text[:500]}")]
        except Exception as exc:
            logger.exception("Tool %s failed", name)
            return [TextContent(type="text", text=f"Tool failed: {exc}")]

        import json
        return [TextContent(type="text", text=json.dumps(result, default=str, indent=2))]

    return server


def main() -> None:
    logging.basicConfig(
        level=logging.INFO,
        format="%(asctime)s %(levelname)s %(name)s %(message)s",
    )
    from mcp.server.stdio import stdio_server

    async def _run() -> None:
        server = build_server()
        async with stdio_server() as (read_stream, write_stream):
            await server.run(read_stream, write_stream, server.create_initialization_options())

    asyncio.run(_run())


if __name__ == "__main__":
    main()