Internal Safety Override
- Status: active for local design and scoped local implementation.
- Tools must default to least privilege.
- Network-capable tools must be disabled until explicitly approved.
- Destructive tools require dry-run and confirmation.
Tool Orchestration MCP
Current Model Baseline
All agents and subagents use the same single local Ollama qwen2.5:3b model through http://127.0.0.1:11434. No cloud LLMs, remote model APIs, separate providers, or per-agent model pools are used by tool orchestration.
Tool Registry Fields
{
"name": "string",
"description": "string",
"permission": "read_local|write_local|execute_local|automate_app|network|high_impact",
"side_effects": false,
"dry_run": true,
"input_schema": {},
"output_schema": {},
"redaction": "string",
"approval_required": false
}
Tool Classes
Read local:
- File read, search, metadata, local app state.
Write local:
- Patch file, create note, update memory.
Execute local:
- Tests, formatters, safe scripts.
Automate app:
- Browser, Office, desktop UI.
Network:
- External docs, APIs, cloud.
High impact:
- Send, publish, deploy, delete, buy, account/security changes.
MCP Server Guidelines
- Bind to loopback only.
- Use explicit schemas.
- Bound output length.
- Redact secrets at source.
- Separate read tools from write tools.
- Add dry-run to risky tools.
- Log tool summaries, not raw sensitive payloads.
1---2name: tool-orchestration-mcp3description: Use to design Jarvis tool registry and local MCP-style tool orchestration with typed schemas, permissions, dry-run modes, audit logs, loopback-only services, and safe tool execution.4category: local-agent5metadata:6 source: https://skillmd.com/skills/brathod1432/tool-orchestration-mcp7---89## Internal Safety Override1011- Status: active for local design and scoped local implementation.12- Tools must default to least privilege.13- Network-capable tools must be disabled until explicitly approved.14- Destructive tools require dry-run and confirmation.1516# Tool Orchestration MCP1718## Current Model Baseline1920All agents and subagents use the same single local Ollama `qwen2.5:3b` model through `http://127.0.0.1:11434`. No cloud LLMs, remote model APIs, separate providers, or per-agent model pools are used by tool orchestration.2122## Tool Registry Fields2324```json25{26 "name": "string",27 "description": "string",28 "permission": "read_local|write_local|execute_local|automate_app|network|high_impact",29 "side_effects": false,30 "dry_run": true,31 "input_schema": {},32 "output_schema": {},33 "redaction": "string",34 "approval_required": false35}36```3738## Tool Classes3940Read local:41- File read, search, metadata, local app state.4243Write local:44- Patch file, create note, update memory.4546Execute local:47- Tests, formatters, safe scripts.4849Automate app:50- Browser, Office, desktop UI.5152Network:53- External docs, APIs, cloud.5455High impact:56- Send, publish, deploy, delete, buy, account/security changes.5758## MCP Server Guidelines5960- Bind to loopback only.61- Use explicit schemas.62- Bound output length.63- Redact secrets at source.64- Separate read tools from write tools.65- Add dry-run to risky tools.66- Log tool summaries, not raw sensitive payloads.