DROPGoogle launches Gemini 3.8 Flash TTS and Flash-Lite TTS///DROPGemini Omni brings 1080p generation to Google Vids///BREAKClaude Opus 5.5 is here: 1M context, $4 input / $20 output///BREAKOpenAI adds GPT-6 Sol and Luna to the GPT-6 family///PAPEROpus 5.5's system card documents gains—and evaluation limits///BREAKGrok 4.7 launches at $2 input / $6 output per million tokens///DROPGrok Voice Transcribe 2.0 targets difficult real-world audio///TRENDGoogle steers new Gemini API projects to 3.8 Flash or 3.5 Flash-Lite///DROPGoogle launches Gemini 3.8 Flash TTS and Flash-Lite TTS///DROPGemini Omni brings 1080p generation to Google Vids///BREAKClaude Opus 5.5 is here: 1M context, $4 input / $20 output///BREAKOpenAI adds GPT-6 Sol and Luna to the GPT-6 family///PAPEROpus 5.5's system card documents gains—and evaluation limits///BREAKGrok 4.7 launches at $2 input / $6 output per million tokens///DROPGrok Voice Transcribe 2.0 targets difficult real-world audio///TRENDGoogle steers new Gemini API projects to 3.8 Flash or 3.5 Flash-Lite///
DARC.DEV
All guides
GuideLevel · Deep#MCP#Security#Agents

MCP in 2026: connect tools without giving away the keys

The July 28 specification, HTTP authorization, token audience checks and the practical difference between a tool description and permission to execute.

DCDariusz Czechowicz23 Sept 20263 min readPractical guide
58
Guide
Source-checked Sources & corrections
4 practical sections · checklist · sources

Practical guide. Provider claims are attributed to the sources below. This page does not claim a new independent DARC benchmark.

For: Developers connecting agents to internal or third-party tools

Before you start: HTTP and OAuth fundamentals · A read-only test tool

⚡ TL;DR — the 30-second version

  • Pin the MCP revision and SDK versions.
  • HTTP authorization and STDIO credentials are different paths.
  • Tool annotations do not grant permissions.
  • Token passthrough and unverified state handles are security hazards.

The 2026-07-28 MCP release describes a stateless protocol core, multi-round-trip requests, routing headers, cacheable lists and authorization hardening. That means older tutorials may encode assumptions that no longer match the current revision. Pin the protocol and SDK versions used by both ends of your integration.

Start with the trust map: the human user, the agent host, the MCP client, the MCP server, the authorization server and any downstream service. Each transition is a place where credentials, instructions or data can cross a boundary.

“MCP standardizes a connection. Your authorization policy still decides whether an action is allowed.”
— Editorial interpretation

Field guide / 01

01 / Separate local transport from remote authorization

The HTTP authorization specification is optional for MCP implementations but normative when that authorization mechanism is used. It describes OAuth-based access to a protected MCP server. The same document explicitly says STDIO implementations should retrieve credentials from the environment rather than follow the HTTP flow.

Do not copy a remote server's OAuth design into a local process without examining the actual transport. Conversely, do not assume a remote server is safe merely because a local STDIO example worked.

Field guide / 02

02 / Validate resource and issuer

An access token must be intended for the MCP server receiving it. Token passthrough is forbidden by the reviewed specification: a token for an unrelated downstream API must not be accepted as the MCP credential. Downstream access is a separate authorization relationship.

The current revision also hardens issuer validation during the authorization response. Clients record the chosen issuer and apply the specified comparison before redeeming the authorization code. Use a maintained OAuth implementation; a hand-written flow that omits state or issuer checks can turn a useful integration into an account-linking vulnerability.

  • → Validate the intended resource and token audience.
  • → Keep bearer tokens out of query strings and logs.
  • → Bind authorization callbacks to the initiating user/session.
  • → Never treat possession of a workflow state handle as authorization.

Field guide / 03

03 / Read annotations as claims

A server can describe a tool as read-only, but the client must treat untrusted annotations as untrusted. Validate the arguments, constrain accessible records and place approval before a sensitive operation. Display the proposed tool input so the reviewer can see what data would leave the system.

Keep tool catalogs deterministic when their underlying contents have not changed. The current tools specification recommends stable ordering, which also helps avoid accidental prompt-cache churn.

Field guide / 04

04 / A safe first integration

Expose one read-only lookup over synthetic data. Test a missing scope, an expired token, a resource owned by another test user, an oversized argument and an unexpected tool result. Confirm the server rejects the unauthorized operation independently of what the model requests.

Only after that should you add a write tool. Make its idempotency, approval and rollback behavior explicit. The test deliverable is a permission matrix and a trace of both a permitted and a denied call—not a screenshot of a successful chat.

Before you ship

0/5

Use this as a review aid. Checks reset when you leave this page.

Questions that come up

Does installing an MCP server make an agent trustworthy?

No. The server expands the available tool surface. Trust depends on its code, authentication, authorization, inputs, outputs and execution environment.

Put it into practiceCreate an agent permission brief

Sources & verification

Last checked 23 Sept 2026. Publication dates and access conditions belong to the linked originals.

  1. 01MCP 2026-07-28 — specification releaseModel Context Protocol↗
  2. 02MCP 2026-07-28 — authorizationModel Context Protocol↗
  3. 03MCP — security best practicesModel Context Protocol↗
  4. 04MCP — tools, permissions and validationModel Context Protocol↗
Editorial policy and correction log →
#MCP#Security#Agents