+ nAIvigateA SHEET FOR PEOPLE WITHOUT 2 HOURS
▣ fiche-005.md ×
□ index.md☰ Contents
The path at a glanceSkill or MCP? (the starting question)Plug in an existing server or code your own?Is a tool-equipped agent safe? The attack surfaceDefense in depth (4 layers)Adopting a third-party block: the 7-point checklistWhere to find skills & servers (verified Aug 2026)Skills & MCP: the fiche that maps the whole path
Skill or MCP? Plug in or code? Is an agent actually safe? This sheet condenses the key decisions from the full course, and opens onto the 9 detailed lessons.
The path at a glance
9 steps, one thread, two audiences3 rowsTHE MODELThe brain that decides and orchestrates. The single decision point: the arrow always starts from it.CORE
THE SKILLThe manual: know-how loaded on demand. Says how to do a task well.expertise
MCPThe hands: a server exposing tools. Supplies the what and does: DBs, APIs, files, email.access
Skill or MCP? (the starting question)
the right tool for the need3 rowsNEED KNOW-HOW"Write per our method", "follow this house format" → a skill. No external access required.SKILL
NEED ACCESS"Fetch from the CRM", "send this email" → an MCP server. Know-how is not enough, you need hands.MCP
NEED KNOWLEDGE"What do our documents say?" → rather RAG. The skill carries the procedure, not the knowledge base.≠ neither
Plug in an existing server or code your own?
the trade-off that saves days3 rowsPLUG IN THE EXISTINGStandard need (files, DB, Git, calendar) → the ecosystem already built it. 10 min to plug in vs days of code.DEFAULT
CODE YOUR OWNThe hand must fit your system (internal API, business DB, house rules) → minimal server, two tools often suffice.SPECIFIC CASE
LOCAL VS REMOTEWhat touches your workstation stays local; what touches the team is shared remotely (secrets service-side).pocket rule
Is a tool-equipped agent safe? The attack surface
what the CISO must know4 rowsINDIRECT INJECTIONA booby-trapped text read by a tool enters the context and gives the agent orders. You don't prevent it: you secure the action, not the conviction.BLIND SPOT
CONFUSED DEPUTYThe agent acts with its rights, not the requester's. It is made to cross a door forbidden to the user.BLIND SPOT
EXFILTRATION BY CHAININGA tool that reads sensitive data + a tool that writes outside = a leak in two calls, neither suspicious alone.BLIND SPOT
SHADOW AIServers plugged in without inventory. The first question is not "are they well coded?" but "which are plugged in, with which rights?"GOVERNANCE
Defense in depth (4 layers)
no single countermeasure is enough4 rows① LEAST PRIVILEGEMinimal-scope token, at the source. What the agent lacks, no injection can make it do. The layer that holds when others fall.bedrock
② SEPARATIONOne server per domain. Sensitive reading and free egress do not cohabit → breaks the chaining.
③ HUMAN IN THE LOOPConfirmation before the irreversible: sending, deleting, paying. The friction is a feature.
④ SUPERVISIONA log of every call. Detect, trace, prove after the fact, to an auditor, a client, a regulator.
Adopting a third-party block: the 7-point checklist
you don't adopt on reputation, you verify7 rowsPROVENANCEIdentifiable, accountable origin, not an anonymous fork.
FULL READINGSKILL.md and scripts read in full, including the "boring" parts where traps hide.
SCOPEFor a server: tools = permissions. Every tool earns its place or disqualifies.
COMBINATIONSWhich read + write capabilities are brought together? (the chaining risk).
LEAST PRIVILEGEToken restricted at the source, never "admin for convenience".
PINNED VERSIONA precise commit, never a moving branch.
REGISTRYName, version, owner, next review, recorded.
Where to find skills & servers (verified Aug 2026)
the official bedrock + how to search4 rowsSKILLS OFFICIALgithub.com/anthropics/skills: reference skills in source + the open standard agentskills.io.OFFICIAL
MCP OFFICIALmodelcontextprotocol.io (spec), modelcontextprotocol/servers (reference servers), registry.modelcontextprotocol.io (authenticated registry).OFFICIAL
COMMUNITY"awesome-mcp-servers" lists (wong2, appcypher, abordage). To be filtered: always trace back to the source repo.to filter
HOW TO SEARCHOn GitHub: sort by activity (pushed:>date, recent stars), read the Insights tab. The star total tells the past, not the present.method
Go deeper
fundamentals, in Learn9 links→1 · La carte avant le territoire12 MIN→2 · Anatomie d'une skill12 MIN→3 · Créer sa première skill15 MIN→4 · Skills en entreprise : gouvernance13 MIN→5 · MCP : le protocole expliqué13 MIN→6 · Utiliser un serveur MCP existant13 MIN→7 · Créer son serveur MCP minimal16 MIN→8 · Sécuriser ses MCP15 MIN→9 · L'écosystème : où trouver, où ça bouge13 MIN
Moving from theory to governance This course condenses what our firm implements for its clients. To turn these reflexes into team processes (inventory, security, AI Act compliance), let's talk.
Discover nAIvigate StudioFollow nAIvigate | naivigate.com/en/fiches