ragleap-integrations

An MCP (Model Context Protocol) client that turns allowlisted tools on remote MCP servers into ragleap_tools.Tool objects, so MCP tools and native tools share one shape.

pip install ragleap-integrations

What this is (and isn't)

v0.3.0 is an MCP client over Streamable HTTP only. The owner configures servers and an exact server.tool allowlist; the model only supplies a tool's JSON arguments, never a URL, a token or a tool outside the allowlist. It does not own a tool-calling loop (that is ragleap-agents' job).

Not supported in this version: stdio (it launches a subprocess), the deprecated HTTP+SSE transport, sampling/elicitation/roots (a server asking for client input gets a clear "unsupported" result), resources, prompts, subscriptions and OAuth. Connectors, code execution and HTTP fetch are deliberately not here: each needs its own security design pass.

Quickstart

from ragleap_integrations import McpConfig, McpServerConfig, make_mcp_tools

config = McpConfig(
    servers=[McpServerConfig("deepwiki", "https://mcp.deepwiki.com/mcp")],  # token="..." if a server needs one
    allowed_tools=["deepwiki.read_wiki_structure"],
)
tools = make_mcp_tools(config)  # discovers the allowlisted tools once (tools/list)

openai_tools = [t.to_openai_schema() for t in tools]  # or t.to_gemini_schema()
result = tools[0].call(repoName="sqlite/sqlite")
print(result.success, result.result)  # True {"text": "..."}

Nothing is read from the environment: your application builds the config. Tools are exposed as <server>__<tool> (characters outside letters, digits, _ and - become _, at most 64 characters).

To keep server-controlled text out of the model's context entirely, give the descriptions and schemas yourself; nothing is sent over the network at setup:

from ragleap_integrations import McpToolSpec

spec = McpToolSpec(
    server="deepwiki", name="read_wiki_structure",
    description="List the documentation topics of a GitHub repository.",
    parameters={"type": "object", "properties": {"repoName": {"type": "string"}}, "required": ["repoName"]},
)
tools = make_mcp_tools(config, specs=[spec])

Easy mode: connect()

from ragleap_integrations import connect

conn = connect("https://mcp.example.com/mcp", allow=["list_orders", "get_order"], token=MY_TOKEN)
print(conn.report())   # which tools are ready, which were skipped and why (with the setting to change)

provider_tools = conn.toolkit.to_openai_schema()    # also to_gemini_schema() / to_anthropic_schema()
result = conn.toolkit.call(name_from_model, arguments_from_model)   # never raises

allow is an exact list of tool names (no wildcard). name (default "mcp") prefixes the exposed names (mcp__get_order). The toolkit checks a model's arguments against the server's schema before anything is sent. For several servers, build an McpConfig and call connect_config(config).

Governance: approval flags and audit events

A server's tool annotations (readOnlyHint, destructiveHint, ...) are untrusted hints, so they can only make policy stricter. Every tool carries tool.metadata["mcp"] with requires_approval, true unless you list the tool in pre_approved, or trust a server's readOnlyHint for that server.

def approve(server, tool, arguments):
    return ask_a_person(f"Run {server}.{tool} with {arguments}?")   # must return exactly True

conn = connect(URL, allow=["list_orders", "delete_order"], approve=approve,
               pre_approved=["list_orders"], on_event=print)

Without approve the flag is advice for your own approval gate (for example in ragleap-agents). With it, a flagged call runs only if approve returns True; an exception or any other value denies the call, the server is not contacted, and the result is "MCP call was not approved". on_event receives one small dict per discovery and per call (server, tool, ok, error, approval, duration); arguments and results never appear in it. See docs/design/mcp-governance.md.

Multi-tenant use: McpHub

One owner-level McpConfig, one Connection per tenant. The tenant is a Scope: with pip install "ragleap-integrations[security]" it is ragleap-security's Scope; without it a small built-in Scope with the same fields and key format. The model never sees or sets the scope.

from ragleap_integrations import McpConfig, McpHub, McpServerConfig, RateLimit, Scope

hub = McpHub(
    McpConfig(servers=[McpServerConfig("crm", "https://mcp.example.com/mcp")],
              allowed_tools=["crm.get_order", "crm.update_order"]),
    token_provider=lambda scope, server: vault.token_for(scope.tenant, server),
    allow_for=lambda scope: ["crm.get_order"] if scope.tenant == "trial" else None,
    limit=RateLimit(60, 60),   # 60 calls a minute per tenant
    audit=audit_log,           # anything with append(event, data, scope)
)
conn = hub.connect(Scope(tenant="acme", subject="u1"))
conn.toolkit.call(name_from_model, arguments_from_model)  # same Connection as connect()

• Each scope has its own client, tool snapshot and token; nothing mutable is shared between scopes.
• token_provider(scope, server) is called for every request and nothing is cached by this package. None sends no token (the static token is not used then); a failing provider gives the constant text "MCP token is unavailable".
• allow_for(scope) can only narrow the owner's allowlist. policy_for(scope) picks a per-tenant McpPolicy; its approve_scoped(scope, server, tool, arguments) wins over approve, and without a scope it denies.
• RateLimit is per tenant (all of a tenant's scopes share it) and is checked before approval and before any network call.
• Events carry tenant and scope; arguments and results are never in them. audit receives every event with the scope.
• hub.connect() needs an exact Scope (a string, a dict, None or a subclass raises ScopeError). This package does not decide who the caller is.

Design and limits: docs/design/multi-tenant.md.

Trust model

• Everything a server says is untrusted text: tool descriptions, input schemas and results flow into a model's context. This package does not screen them for prompt injection. The structural limits are the allowlist, a snapshot of the tool list taken once at setup (a server that changes a description later has no effect), length caps, and optional owner-supplied specs.
• Server-supplied error text is never put in a result; failures carry a constant message plus, at most, an HTTP status or a JSON-RPC error code.
• Network access is bounded in code: https only, no credentials in the URL, no IP-literal hosts, DNS resolved once with every address required to be public (so DNS rebinding cannot swap in an internal address), the connection pinned to that IP with TLS (minimum 1.2) verified against the hostname, no redirects, no compressed responses, a response-size cap, and a wall-clock deadline that also covers the TLS handshake and the headers.
• A bearer token, if configured, is sent only to its own server URL and is never logged.

Protocol

The current revision (2026-07-28) is tried first: every request is its own POST carrying MCP-Protocol-Version, Mcp-Method and Mcp-Name headers and the version, client info and capabilities in params._meta; parameters marked x-mcp-header in a tool's schema are mirrored into Mcp-Param-* headers. A 4xx without a recognized modern error (-32020, -32021, -32022) identifies a legacy server, and the client then uses the legacy initialize handshake with protocol version 2025-06-18 only (an explicit allowlist, not "whatever the server answers"). The era found is remembered per server.

Limits and defaults

Known limitations

• This package does not decide who the caller is: tenant isolation is only as strong as the Scope you pass and the token your token_provider returns.
• Rate windows are kept per tenant until hub.forget(tenant=...), the limit is per process, and allow_for and policy_for are read once per scope.
• On a legacy server every tool call performs a full initialize handshake (three POSTs) and never closes the session.
• A HeaderMismatch error is reported, not retried after re-reading the tool list.
• Only text content is returned; other content types are counted and omitted.
• ragleap_tools.Tool.call in ragleap-tools 0.4.0 cannot take an argument literally named self.
• Not safe to share one client between threads while calls are running.
• Server-supplied text is untrusted and not screened for prompt injection: tool descriptions and schemas, tool results, and the text of a tool-level error (it becomes ToolResult.error, capped at max_result_chars like a result).
• A server whose tool list is larger than the response cap (1 MiB by default; 256 KiB in 0.1.0, which a real 15-tool server's 359 KB list exceeded) makes discovery fail with "response too large; raise max_response_bytes" until the limit is raised.

Verification status

• 331 tests (the 275 earlier ones unchanged plus 56 new in 0.3.0: scope and token isolation, allowlists, policies, limits, events, audit, threads; 327 passed and 4 skipped without ragleap-security); 40 of 40 deliberate breaks of the new code caught.
• Live-checked on 2026-10-26 with the repository source of the 0.3.0 branch on Python 3.10.12 against a third party's public mock MCP auth server, using scripts/tenant_live.py (repository only, not in the wheel): two real tokens were minted; each tenant's requests carried only its own token (compared by hash); a made-up token got HTTP 401 and the real tenants kept working; after one tenant's token was changed only that tenant was refused (the token is resolved per request); RateLimit(2, 60) denied the third call locally and only two tools/call requests reached the server; no event carried a token. The mock server cannot tell tenants apart, so separation is shown from the requests that were sent.
• 275 tests (67 of them new in 0.2.0: governance, events and connect(); 53 of 53 deliberate breaks of the new code caught), including a real local TLS server with a throwaway certificate for the transport (hostname verification, IP pinning, size cap, truncated bodies, and the deadline against slow-drip bodies, slow-drip headers and a stalled handshake).
• Live-checked on 2026-10-05 against DeepWiki's public MCP server (https://mcp.deepwiki.com/mcp, no authentication): the server answered the modern request with HTTP 400 and a JSON-RPC -32600 error, the client fell back to the legacy handshake, negotiated 2025-06-18, listed tools and read a repository's documentation structure. This verifies the legacy path, the fallback decision, event-stream responses and the pinned TLS transport against a real server.
• Live-checked on 2026-10-12 against two public Google Cloud MCP endpoints (https://developerknowledge.googleapis.com/mcp and https://pubsub.googleapis.com/mcp) on the modern 2026-07-28 path, using the published 0.1.0 wheel in a clean virtual environment (Python 3.10.12). Read-only: tools/list only, no credentials, no tools/call. The server accepted the modern request, replied with plain application/json (non-stream), and the client discovered the tools and read their input schemas (3 tools from one server). A second server's tool list was 359,293 bytes, above the default max_response_bytes (256 KiB): discovery failed with "response too large" until the limit was raised, and two of its tools were skipped because their input schemas exceed max_schema_chars. Both servers answered an unsupported protocol version with HTTP 400 and -32600 and no supported-version list; neither implements server/discover (this client never calls it).
• Live-checked on 2026-10-16 against https://pubsub.googleapis.com/mcp using the published 0.1.1 wheel in a clean virtual environment (Python 3.10.12), with no limit overrides. Read-only: tools/list only, no credentials, no tools/call. The 359,293-byte list of the 15-tool server, which 0.1.0 refused at its 256 KiB default, is now accepted with the default max_response_bytes (1 MiB), and the client recorded the modern era. Four of the 15 tools have an input schema above the default max_schema_chars (10,000): update_subscription (33,502 characters), create_subscription (32,989), update_topic (23,537) and create_topic (22,835); the other eleven are 2,631 or smaller. With the first three tools allowlisted, discovery registered one and skipped create_topic and update_topic, as designed. As an experiment only, max_schema_chars=40000 (an override; the default is unchanged) registered all 15 tools. Not verified: that schemas of this size are usable by a model.
• Live-checked on 2026-10-17 against a third party's public test servers (https://mcpplaygroundonline.com/mock-mcp-servers; their own description of what each server does is not independently verified), using the repository source at commit 912d100 (0.1.1) on Python 3.10.12 and scripts/live_check_mock.py (repository only, not in the wheel): (a) a modern-path tools/call succeeded against the stateless server, both pinned with ?rev=2026-07-28 and unpinned, and the tool itself reported the modern era and revision 2026-07-28; (b) bearer-token authentication: a call without a token was refused (HTTP 401) and the same call with a freshly minted token succeeded; (c) against the same server pinned to the legacy revision (?rev=2025-11-25) the client fell back to the legacy handshake and a tool call succeeded, the tool reporting the legacy era (the reply that triggered the fallback was not recorded); (d) a tool that reports an error returned success=False with the server's message in error and nothing raised.
• Live-checked on 2026-10-25 with the published 0.2.0 wheel on Python 3.10.12 against Google Cloud's Pub/Sub MCP endpoint (https://pubsub.googleapis.com/mcp, modern path, tools/list only, no credentials, no tools/call), using scripts/connect_live.py (repository only, not in the wheel): connect() with the default limits registered 11 tools and skipped the four whose input schemas exceed max_schema_chars. All 15 tools carry the four boolean hints. With trust_annotations=False all 11 are flagged requires_approval; with trust_annotations=True exactly six are not (the get_ and list_ tools for topics, subscriptions and snapshots, which declare readOnlyHint: true and destructiveHint: false). In both runs the flags equal an independent calculation from the raw hints, and each connect() produced one discover event and no call event.
• Not live-verified: x-mcp-header mirroring (none of the four tools on the complex-schema server carries the annotation), pagination (none of the servers checked paginated) and this client's own handling of a -32022 reply (it never received one live; a raw legacy initialize sent to the modern-only server got HTTP 400 and -32022 with the supported versions in error.data.supported, which is where the client looks). They follow the public specification and are covered by tests against fakes only; treat them as best-effort until confirmed live.
• Not live-verified (new in 0.2.0): the approve hook, its denial text and the per-call audit events against a real server (they need a real tools/call, and the Pub/Sub server requires credentials), and connect(token=...) against a real authorization server.
• Not live-verified (new in 0.3.0): approve_scoped, allow_for and policy_for against a real server, a real ragleap-security AuditLog fed by a live run, concurrent tenants against a real server, a server that really separates tenants' data, and the published wheel (the live check ran on the branch source and is repeated with the wheel after release).

License

MIT