Scopes and authorization

Every tool in Secronyx is registered with exactly one scope label. Authorization happens at two points: when the server starts, a scope policy decides which tools are registered at all; on the HTTP transport, every tools/call is then checked against the scopes carried by the authenticated identity. This page lists the scopes, explains both stages, and shows what a denial looks like.

The scope list

There are nineteen scopes. The counts below come from starting the current build with --scopes <name> --enable-sensitive and reading tools/list; the authoritative per-tool listing is the Tool reference.

In total 661 tools are registered with --enable-sensitive, and 630 without it. An earlier document described a scope named sbom; the code registers those tools under software.

Stage one: registration policy

ScopePolicy in internal/mcp/authz.go is installed before any tool is registered:

type ScopePolicy struct {
    // AllowedScopes, when non-empty, is the only set of scopes whose tools
    // are registered. Empty means every scope except "sensitive".
    AllowedScopes []string

    // EnableSensitive permits registration of tools in the "sensitive"
    // scope. It is required even when AllowedScopes lists "sensitive".
    EnableSensitive bool
}

The command line maps onto it as follows:

A tool the policy refuses is simply never registered. It is absent from tools/list on every transport, including stdio, and calling it returns the ordinary tool-not-found error, so a client cannot tell the difference between a tool that does not exist and one that was withheld:

{"jsonrpc":"2.0","id":1,"error":{"code":-32602,"message":"Tool not found","data":"get_env_vars"}}

At start-up the server logs how many tools the policy withheld, and a separate warning when sensitive tools are on:

Scope policy: 31 tools not registered (sensitive=false, scopes="")
WARNING: sensitive-scope tools are enabled

Two consequences worth knowing. First, a scope name that matches nothing registers nothing: --scopes bogus starts a server with zero tools and logs Scope policy: 661 tools not registered. Second, --scopes core,sensitive without --enable-sensitive registers the eight core tools only.

# Core metrics only
secronyx --scopes core

# Core, logs and hooks
secronyx --scopes core,logs,hooks

# Everything, including sensitive tools (both are required)
secronyx --scopes core,logs,hooks,sensitive --enable-sensitive

# The same through the environment, for a service unit
SECRONYX_SCOPES=core,logs SECRONYX_ENABLE_SENSITIVE=1 secronyx

On stdio the registration policy is the only control. Nothing in the stdio protocol carries a caller identity, so the process that spawns the server is trusted to the extent of whatever was registered. The Command line page covers the --query mode, which does not go through the registry at all.

Stage two: per-request checks

When the HTTP transport authenticates a request it places an Identity in the request context:

type Identity struct {
    Subject  string
    ClientID string
    Scopes   []string
    // Method names the authenticator that established the identity.
    Method string
}

tools/call then runs IdentityAllows(identity.Scopes, tool.Scope) before the handler executes. A grant matches when it is:

• the wildcard *;
• the tool's scope name itself, for example logs; or
• the namespaced form mcp:tools:<scope>, for example mcp:tools:logs, so an authorization server whose scope namespace is shared with other resources can issue distinct values.

The bare audience value mcp:tools grants nothing on its own. An identity with an empty scope list can authenticate but can call nothing. tools/list is filtered with the same function, so a caller sees only what it may invoke; the .well-known/oauth-protected-resource document advertises scopes_supported as the scopes that map to registered tools.

A request with no identity in its context (stdio, or HTTP with no authentication configured, which the listener guardrails permit only on loopback or with --allow-unauthenticated) is governed by the registration policy alone.

Where identity scopes come from

Each authentication method produces the scope list differently. Details of configuring each are in Authentication.

Whatever the method, sensitive in a grant only takes effect if the server was started with --enable-sensitive; a grant cannot register a tool.

What a denial looks like

A call outside the identity's grant is refused with JSON-RPC error -32003, which sits in the implementation-defined server error range. The HTTP status is 200 because the request was authenticated and well-formed; the refusal is at the JSON-RPC layer.

{"jsonrpc":"2.0","id":2,"error":{"code":-32003,"message":"Forbidden","data":"tool \"get_journal_logs\" requires scope \"logs\""}}

The denial is audited as a tools/call event with result denied and error scope denied, and counted in metrics under the error type scope_denied. Abridged — the written record also carries timestamp, seq, event_id and the hash-chain fields:

{"action":"tools/call","resource":"get_journal_logs","identity":"demo","client_ip":"127.0.0.1","params":{"lines":5},"result":"denied","error":"scope denied"}

Worked example with an API key

Issue a key scoped to core, start a loopback listener, and exercise both stages. Loopback is the one configuration that may run without TLS; see Remote access over HTTP for anything else.

secronyx apikey create --file ./keys.json --name demo --scopes core
# Created API key 78ef7122 for "demo" (scopes: core)
# This key is shown once and cannot be recovered:
# msk_78ef7122_...

secronyx --transport http --listen 127.0.0.1:18080 \
    --api-keys-file ./keys.json \
    --audit-output ./audit.jsonl

The key sees eight tools, may call one of them, and is refused on another:

KEY=msk_78ef7122_...

curl -s -X POST http://127.0.0.1:18080/ -H "X-API-Key: $KEY" \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | jq '.result.tools | length'
# 8

curl -s -X POST http://127.0.0.1:18080/ -H "X-API-Key: $KEY" \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"get_journal_logs","arguments":{"lines":5}}}'
# {"jsonrpc":"2.0","id":2,"error":{"code":-32003,"message":"Forbidden","data":"tool \"get_journal_logs\" requires scope \"logs\""}}

curl -s -X POST http://127.0.0.1:18080/ -H "X-API-Key: $KEY" \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"get_uptime"}}'
# {"jsonrpc":"2.0","id":3,"result":{"content":[{"type":"text","text":"{\n  \"boot_time\": ..."}]}}

The audit file records the authentication (with the granted scopes in metadata), the denial and the success in one chain. The full walk-through, including expiry and CIDR restrictions, is in Tutorial: per-operator API keys.

Choosing a policy

• Start with --scopes core,logs,hooks for general diagnostics and widen per host as needed. Everything else can be added without restarting clients, but the server must restart: the policy is evaluated at registration and tools registered earlier are not re-evaluated.
• Keep sensitive off unless the host is isolated and the identity that will call those tools is individually accountable. get_env_vars and get_process_environ redact secret-looking values before returning them, but the scope as a whole — get_sudo_config, get_ssh_config, get_user_accounts, get_auth_logs — still describes how the host authenticates and authorises.
• On HTTP, issue narrow grants per identity rather than widening the registration policy. A core-only API key on a server registered with every scope still sees eight tools.
• Prefer mcp:tools:<scope> grants from an enterprise IdP whose scope namespace is shared, and plain names from the API key, mTLS and SSH files, which are Secronyx-specific.

For the wire format of the error see JSON-RPC API; for what the audit trail records about denials see Audit logging.