marcmarti9 MCP Tooling Fit
Choose tools for the reviewed task, not because they are popular or globally installed. The primary AI owns the semantic choice; deterministic code may resolve an explicit server/stack ID and enforce risk gates.
When to use
Load this skill JIT when the task materially depends on MCP/tool selection, when the current tool inventory is noisy, or when an important capability appears missing.
Do not load it for ordinary work whose existing tools are already sufficient.
Protocol
1. Inspect current state
Agent-facing helpers include:
agentit mcp status --project .
agentit mcp available --project .
agentit mcp active --project .
The agentit-manager MCP can expose the equivalent inventory/enable/disable operations.
Record only what matters for the current task:
- explicit current availability;
- required permissions/capabilities;
- write vs read-only behavior;
- secrets/env requirements;
- risk classification.
Catalog presence is not runtime availability.
2. Infer needs from task and repository facts
Inspect manifests, frameworks, DB/config, browser/UI surfaces, design sources, CI/deployment and external integrations. Do not ask the user for facts already visible in the project.
Example curated stacks may include:
| Need | Candidate stack |
|---|---|
| repository/docs/browser baseline | developer_core |
| frontend/browser verification | frontend |
| design inspection | design_studio |
| mobile / Expo / RN UI | mobile_design (appllama + context7; paid credits) |
| backend/data work | backend_data |
| current-source research | research |
| product operations | product_ops |
These names are discovery conveniences, not semantic routing rules. The primary AI decides whether a stack actually fits.
3. Prefer the minimum useful set
Keep or enable only tools that earn their permission/context/coordination cost. Disable unrelated always-on noise when doing so is safe and useful.
Plan mutations before applying them. Higher-risk write-capable tools require the applicable human/review gate.
4. Discover gaps when necessary
If Agentit's catalog does not cover a required capability:
- prefer official/maintained providers;
- check source, maintenance, scopes and secret handling;
- compare a small number of realistic candidates;
- do not silently install or authorize external software.
Current-source discovery belongs in the task's reference/tool plan, not in a permanent global preload.
5. Persist only operationally useful state
When a substantial task needs continuity, record the selected MCP/tool set and pending approvals in local .agentit/STATE.md. Do not publish transient tool enablement decisions as tracked project documentation unless they are a durable architectural/operational contract.
Safety
- least privilege by default;
- secrets through environment/provider secret stores, never repository files;
- verify current auth/availability before depending on a mutable external service;
- RISK_3/RISK_4 or destructive external actions require the applicable human and independent-review gate;
- disabling/re-enabling tools is allowed when reversible and authorized.
Anti-patterns
- enabling every MCP "just in case";
- treating a stack name as a natural-language router;
- assuming catalog presence means authenticated availability;
- silently installing third-party servers;
- asking for stack facts that the repository exposes;
- writing secrets or mutable tool state into public docs;
- using powerwords to activate tooling.
Verification
- current inventory was inspected when material;
- selected tools map to explicit task capabilities;
- unnecessary permission/context noise was avoided;
- external installs/actions respected risk and consent gates;
- substantial continuity notes, if needed, remain in
.agentit/STATE.mdby default.