SPYN3 field notes / MCP + agent infrastructure
The context layer,explained
MCP adoption is accelerating, but connecting an agent is only the visible edge. These field guides explain the retrieval, permissions, provenance and tool design that make connected business data genuinely useful.
Resource library
From signal to system
Read by job, then follow the contextual links. Every page answers one search intent and connects it to the evidence, guide or worksheet that makes it actionable.
News & shifts
News & shifts
Time-sensitive protocol and platform changes, verified against primary sources.
MCP 2026-07-28 stateless specification
What the July MCP specification changes for remote servers, authorization and scaling.
News & shifts / MCP adoption02MCP authorization in 2026
A practical guide to MCP authorization in 2026: issuer validation, issuer-bound credentials, client registration changes and migration priorities.
News & shifts / MCP adoption03ChatGPT custom MCP connectors
What builders need to know about ChatGPT custom MCP connectors, developer mode, tool safety, administration and production-ready data access.
News & shifts / MCP adoption04The official MCP Registry
Understand the official MCP Registry, what server metadata can solve, what discovery cannot guarantee and how builders should prepare for distribution.
News & shifts / MCP adoption05A2A vs MCP
A2A and MCP are complementary agent protocols. Compare agent-to-agent coordination with tool and context access, plus an architecture using both.
News & shifts / MCP adoption06Remote MCP client support in 2026
A source-backed snapshot of remote MCP support claims for Claude, Cursor, ChatGPT, VS Code and related clients.
News & shifts / MCP adoption07Stateless MCP core and proxying
How the 2026 stateless MCP core, routing headers and cacheable lists help gateways and proxies operate MCP traffic.
News & shifts / Security08PHP2Shell lessons for agents
The October 2026 PHP engine RCEs hit every server that ran PHP. The same pattern threatens agent integrations scattered across laptops and VMs.
How-to
How-to
Practical implementation guides for dependable agent context.
What is Model Context Protocol?
A plain-language explanation of MCP clients, servers, tools and resources.
How-to / Agent data10Connect business data to an AI agent
A step-by-step architecture for connecting files, websites and databases to an AI agent with retrieval, permissions, citations and MCP or REST.
How-to / Retrieval quality11Evaluate RAG retrieval
Evaluate RAG retrieval with representative questions, relevance judgments, recall, precision, ranking checks and failure analysis before generation.
How-to / Tool design12Design reliable MCP tools
Design MCP tools agents can choose and use reliably with clear names, bounded inputs, compact outputs, errors, risk hints and task-based evaluations.
How-to / Tool design13Reduce agent tool-response tokens
Reduce agent tool-response tokens with projection, pagination, snippets, progressive disclosure, structured formats and response budgets.
How-to / Retrieval quality14Build source-aware AI answers
Build source-aware AI answers by preserving provenance through ingestion, retrieval and generation, then validating citations against returned evidence.
How-to / Agent data15Read-only Postgres for agents
How to expose Postgres to an AI agent safely with schema inspection, SELECT-only queries, limits and source-scoped credentials.
How-to / Security16Audit logging agent tool calls
What to log for MCP tool calls: actor, source, tool, permission decision, timing, outcome and failure details.
How-to / MCP adoption17Claude Code with one endpoint
How to use one remote MCP endpoint for many connected systems when working with Claude Code or Claude MCP connectors.
How-to / MCP adoption18Cursor MCP single endpoint
A practical setup pattern for using Cursor with one MCP endpoint instead of many per-project server definitions.
How-to / Agent data19Google Drive picked files for agents
How to give agents Google Drive access through user-selected files instead of broad Drive-wide permissions.
How-to / Tool design20Lazy tool loading
Use lazy tool loading and source-level toolsets to reduce MCP context size while preserving access to many systems.
How-to / MCP adoption21MCP upstream failures
How to report unreachable MCP servers and connector failures to agents without hiding sources or fabricating empty results.
How-to / Tool design22MCP tool descriptions
How to write MCP tool names and descriptions that help agents pick the right tool and avoid unsafe or irrelevant calls.
How-to / MCP adoption23Migrate to one MCP endpoint
A step-by-step migration from scattered MCP server configs to a single workspace endpoint with scoped keys.
How-to / Tool design24Namespacing MCP tools
How to namespace MCP tools so agents, logs and humans can tell which upstream server owns each capability.
How-to / Security25Read-only vs write MCP tools
How to separate read-only and write MCP tools using annotations, runtime policy and source-scoped permissions.
How-to / Security26Rotate credentials without agent changes
How to rotate upstream tokens and agent keys when integrations are centralized behind one MCP endpoint.
How-to / Security27Scoped API keys for agents
How to issue least-privilege API keys for agents by source, read/write class and rotation lifecycle.
How-to / MCP adoption28Switch agents without rebuilding
How to keep integrations stable when moving between Claude Code, Cursor, ChatGPT and other MCP-capable clients.
How-to / Tool design29Too many MCP tools
How large MCP tool lists waste context, confuse selection and can be reduced with grouping, lazy loading and source-level boundaries.
How-to / Agent data30REST API to read-only tools
How to generate read-only MCP tools from JSON REST API GET endpoints without exposing unsafe methods.
How-to / Security31Upstream MCP OAuth once
How centralized OAuth for upstream MCP servers reduces repeated sign-ins while preserving per-source permission boundaries.
How-to / Agent sandboxes32Scoped keys for agent fleets
Give each agent in a fleet its own key, scoped to the sources and permissions its job needs, so you can audit and revoke one at a time.
How-to / Agent memory33Memory retention and deletion
Decide how long raw transcripts and extracted memories live, who can delete them and what happens when a person leaves.
How-to / Agent memory34Syncing agent instruction files
Each agent reads its own instruction file. Keep one source of truth and stop copy-paste drift.
How-to / Agent sandboxes35Agents in CI pipelines
Agents in CI pipelines need tool access per run. Scope keys narrowly, store them as CI secrets, and keep upstream credentials out of the pipeline.
How-to / Agent memory36Claude Code hooks for memory
How SessionStart, UserPromptSubmit, Stop, PreCompact and SessionEnd hooks can recall and capture memory in Claude Code without manual tool calls.
How-to / Agent sandboxes37Cloud coding agents and tools
Cloud-hosted coding agents run outside your network. Give them scoped, read-mostly access to internal systems through one remote MCP endpoint.
How-to / Agent memory38Codex hooks for memory
Use Codex lifecycle hooks to load memory at session start, recall on each prompt and capture finished turns.
How-to / Agent sandboxes39Agents in VMs, one endpoint
Agents in VMs and containers each need tool access. Point them at one MCP endpoint with a per-VM key instead of configuring every service per machine.
How-to / Agent memory40Session-start context budget
Loading everything at session start wastes tokens and attention. Load a profile, an index and a catalog, and fetch the rest on demand.
How-to / Agent sandboxes41Credentials in agent sandboxes
Agents running in VMs and sandboxes need tool access, but copying upstream secrets into each sandbox spreads them. Give each sandbox one scoped, revocable key instead.
How-to / Agent sandboxes42Context for ephemeral sandboxes
Sandboxed agents lose everything between runs. Load memory and skills from a shared endpoint at start instead of baking them into images.
How-to / Agent memory43Evaluating agent memory
Measure whether memory improves agent outcomes: task sets across sessions, recall precision, harmful recalls and token cost.
How-to / Security44Agent incident response
A runbook for a compromised agent or leaked agent key: contain, revoke, review the audit log, clean memory, restore.
How-to / Security45Secrets out of agent configs
Agent config files like .mcp.json and mcp.json often hold live tokens. Move secrets behind one scoped credential and keep configs safe to share.
How-to / MCP adoption46Local models with MCP tools
Run the model locally and keep tools and memory behind a remote MCP endpoint. What works, what to watch.
How-to / Agent memory47Memory extraction from transcripts
Turn raw conversation turns into durable memory entries with a model: when to extract, what types to keep, how to avoid noise.
How-to / Agent memory48Merging agent memories
Appending every extracted fact produces contradictions. Merge new memories into existing ones and resolve conflicts explicitly.
How-to / MCP adoption49Onboarding a teammate’s agent
A new teammate’s agent should get the team’s tools, conventions and skills on day one, without a credentials hand-off.
How-to / Agent memory50Recall on every prompt
Per-prompt memory recall adds latency and tokens. Cap it, time it out, and fail open so a slow store never blocks a turn.
How-to / Security51Revoke agent access in one place
When a laptop is lost or an agent misbehaves, revocation should be one action, not an inventory of tokens across services.
How-to / Security52Rate limits for runaway agents
Agents loop. Put per-agent rate limits and monthly quotas in front of upstream systems so a loop costs little.
How-to / Security53Scrub secrets from transcripts
Transcripts captured for memory contain pasted tokens and tool output. Strip secrets and noise on the client, before upload.
How-to / Agent memory54Shared skills repository
Store procedures as skills once and let every agent on the team load the same version on demand.
How-to / Agent memory55Shared memory across agents
Each coding agent keeps its own memory files. Put memory behind one MCP endpoint so every agent reads and writes the same context.
How-to / Agent memory56Portable SKILL.md skills
How to write a SKILL.md with a clear name, description and steps so different agents pick and follow it correctly.
How-to / Agent memory57Switch agents, keep context
Move from one coding agent to another and keep tools, memory and skills, by keeping them outside both.
How-to / Security58Reviewing agent skills
A skill is instructions plus files an agent will follow. Review shared skills like code before a whole team’s agents run them.
How-to / Security59Where agent OAuth tokens live
Access and refresh tokens for agent integrations need encrypted storage, refresh handling and revocation. Decide where they live before they spread.
Research
Research
Evidence-led reviews that turn retrieval findings into testable decisions.
Hybrid search vs vector search
Compare hybrid search and vector search for RAG: semantic similarity, exact-match retrieval, reciprocal-rank fusion and evaluation criteria.
Research / Retrieval quality61Does chunk size matter in RAG?
RAG chunk size affects retrieval precision, context completeness and cost. Review current evidence and run a corpus-specific chunking experiment.
Research / Retrieval quality62Metadata filtering in enterprise RAG
Metadata filtering narrows enterprise RAG by workspace, permission, source, time and document type before ranking. Learn the architecture and pitfalls.
Research / Security63Prompt injection in connected data
Understand indirect prompt injection in documents, websites and connected data, plus layered defenses for retrieval systems and tool-using agents.
Strategy
Strategy
Architecture choices, trade-offs and informed build decisions.
MCP vs RAG vs API
A decision guide to three layers that solve different architecture problems.
Strategy / Tool design65One MCP endpoint per workspace
Why a small governed surface is more useful than exposing every connector separately.
Strategy / MCP adoption66MCP vs function calling
Compare MCP and function calling: protocol portability, local tool definitions, discovery, authentication and when a product should support both.
Strategy / Tool design67Build vs buy an MCP server
Decide whether to build or buy an MCP server by comparing differentiation, source breadth, authorization, operations, evaluation and total cost.
Strategy / MCP adoption68Remote vs local MCP servers
Compare remote and local MCP servers for deployment, authentication, updates, latency, data access and multi-user governance.
Strategy / Agent data69Agent memory vs RAG
Agent memory and RAG solve different context problems. Compare conversation state, durable user facts, knowledge retrieval and safe storage design.
Strategy / MCP adoption70Agent-agnostic architecture
Design an agent integration layer that keeps clients replaceable while preserving permissions, observability and source identity.
Strategy / MCP adoption71Avoid AI vendor lock-in
How MCP and a stable integration layer help teams switch agents without rebuilding source connections and permissions.
Strategy / MCP adoption72Hosted vs self-hosted MCP aggregation
A decision guide for hosted and self-hosted MCP gateways across control, operations, compliance, cost and client compatibility.
Strategy / MCP adoption73Aggregator vs gateway vs registry
Compare MCP aggregators, gateways and registries so teams choose the right layer for discovery, runtime routing and governance.
Strategy / MCP adoption74What is an MCP gateway?
A practical definition of MCP gateways and aggregators, when they help, and when a direct server connection is simpler.
Strategy / Security75MCP GDPR and EU residency
How to think about GDPR, controller and processor roles, transfers and EU hosting for MCP-based agent integrations.
Strategy / MCP adoption76Proxied MCP latency and cost
How to reason about latency, retries, token cost and source fanout when MCP calls pass through a gateway.
Strategy / MCP adoption77Shared MCP workspaces
How teams can use workspace-level MCP endpoints to share connections while keeping agent access scoped.
Strategy / MCP adoption78Useful MCP servers by job
How to choose MCP servers by developer, support, revenue, design and operations jobs rather than connector logos.
Strategy / MCP adoption79Agent-agnostic context layer
Why tools, memory and skills belong in one layer that any agent can reach, and what that layer must guarantee.
Strategy / Security80Leaked agent key blast radius
Estimate what a leaked agent credential can reach, and shrink it with per-source scoping, read-only defaults and fast revocation.
Strategy / Agent memory81Agent memory and GDPR
Agent memory is personal data more often than not. Plan where it is stored, how it is erased and who can read it.
Strategy / Agent memory82BYO model key for memory
Memory extraction and semantic search need a model. Using your own provider key keeps cost, data flow and provider choice with you.
Strategy / Agent memory83Context databases vs MCP layers
Context databases store memory, resources and skills. MCP layers route tool calls. Compare the approaches and where they combine.
Strategy / Agent memory84File-system vs vector memory
Compare path-based, browsable memory with embedding-only memory stores: debuggability, cost, recall and when each fits.
Strategy / Security85Read-only by default
Most agent tasks only need to read. Make writes a separate grant, scope keys to sources, and treat unmarked tools as writes.
Strategy / Security86Local stdio MCP server risk
stdio MCP servers run as local processes with the developer’s permissions. Weigh that against remote servers behind an authenticated layer.
Strategy / Agent memory87MCP memory server design
Design choices for serving agent memory over MCP: tool surface, path layout, personal versus shared scope, permissions and budgets.
Strategy / MCP adoption88Resources vs prompts vs tools
MCP offers resources, prompts and tools. Compare them for serving memory and skills to agents that support each unevenly.
Strategy / Security89MCP server supply chain
Adding an MCP server installs code, trusts tool descriptions and hands over credentials. Review each link of that chain.
Strategy / Security90Memory poisoning
Persistent agent memory lets one injected instruction outlive the conversation that carried it. Label, scope and audit memory to contain it.
Strategy / Agent sandboxes91Multi-agent shared context
Orchestrators and sub-agents need common context and different permissions. Separate what they share from what they may do.
Strategy / MCP adoption92One endpoint for your AI stack
Desktop chat apps, coding agents, local models and scripts all need the same tools and context. Give them one endpoint.
Strategy / Security93Patch MCP servers once
Locally installed MCP servers multiply patch work and credential copies. Compare per-machine servers with one hosted layer you patch once.
Strategy / Agent memory94Portable agent memory
Memory kept in one harness’s files is lost when you switch tools. Keep it in a store any agent can reach.
Strategy / Security95Build vs buy agent security
What it takes to secure agent integrations in-house: credential storage, scoping, audit, revocation, patching. Compare it with a managed layer honestly.
Strategy / Agent memory96Team vs personal memory
Some context belongs to one person, some to the whole team. Scope memory so sharing is deliberate and private stays private.
Strategy / Agent memory97What agents should remember
Preferences, decisions and procedures are worth keeping. Secrets, transient state and other people’s data usually are not.
Strategy / Security98Zero trust for agents
Apply zero-trust principles to agents: authenticate every call, re-check roles, scope per source, log everything.
Templates & tools
Templates & tools
Reusable worksheets, scorecards and checklists for production work.
MCP security checklist
A practical review for permissions, credentials, tool risk and observability.
Templates & tools / Tool design100MCP tool design worksheet
A reusable MCP tool design worksheet for purpose, naming, input schema, output budget, permissions, risks, errors and evaluations.
Templates & tools / Retrieval quality101RAG evaluation scorecard
A reusable RAG evaluation scorecard for retrieval relevance, evidence coverage, faithfulness, answer quality, permissions, latency and cost.
Templates & tools / Agent data102AI data connector requirements
Evaluate an AI data connector for ingestion, permissions, freshness, retrieval, citations, MCP, operations, security and portability.
Templates & tools / MCP adoption103Test MCP before rollout
A rollout checklist for MCP integrations covering auth, tool selection, permissions, failures, latency and client-specific behavior.
Templates & tools / Security104Tool poisoning defenses
How MCP proxies can reduce risk from malicious tool descriptions, changed behavior, prompt injection and unexpected upstream updates.
Templates & tools / Security105Agent rollout security checklist
A checklist for security review before agents reach company systems: credentials, scopes, audit, memory, sandboxes, incident response.
Connect once. Swap agents freely.
SPYN3 puts one MCP endpoint in front of every system your workspace connects, so the agent you use next quarter is a one-line change.
Explore the developer guides →
Start free