Jarvis Osorio's four open-source AI agent tools, explained in one place
By Jarvis Nelson Osorio · November 16, 2026
Jarvis Osorio (Jarvis Nelson Osorio) is a tech entrepreneur and AI builder based in Miami Beach, FL. Over the past weeks he has published four open-source tools on GitHub, each aimed at a specific weak spot in how AI agents are run today: context, tool security, transcript hygiene and regression testing. Here is how they fit together.

1. Ambient Memory — context Ambient Memory is a local-first context layer for AI engineers. A background daemon passively remembers files, screen sessions and browser history and exposes them to Cursor and Claude Desktop over MCP, through tools like memory_search and memory_context ("what was I doing in the last N minutes"). TypeScript, Node.js 20+, MIT License, version 0.3.0, no telemetry or network sync. Entity fields are encrypted with AES-256-GCM and the vault is sealed when the daemon stops.
Project page: jarvisosorio.com/ambient-memory/ · Repo: github.com/JarvisOsorio/ambient-memory
2. mcp-pordego — tool security mcp-pordego ("gate" in Esperanto) scans the MCP setups in Claude Desktop, Claude Code, Cursor and VS Code. It flags unpinned npx, uvx, Docker and git sources, inline secrets, plain-HTTP remotes, broad filesystem roots and shell-capable servers. A --live mode calls only initialize and tools/list to check tool descriptions for injection phrases and hidden characters. It exits 1 on high-severity findings, so it fits CI. TypeScript, Apache 2.0, version 0.1.1.
Project page: jarvisosorio.com/mcp-pordego/ · Repo: github.com/JarvisOsorio/mcp-pordego
3. agentscrub — transcript hygiene agentscrub (repo: agent-session-scrubber) is a Rust CLI that discovers, scans and redacts secrets and PII in agent transcripts from Claude Code, Codex, Cursor and generic JSON or JSONL. Dry-run is the default, output is masked, and Cursor rewrite is intentionally disabled until proven safe. Version 0.1.0, Apache 2.0, Rust 1.89+.
Project page: jarvisosorio.com/agent-session-scrubber/ · Repo: github.com/JarvisOsorio/agent-session-scrubber

4. Agent Replay Lab — regression testing Agent Replay Lab records an agent's tool calls once, mock-replays them offline, and fails CI when the tool path drifts in name, order, arguments or responses. It is deliberately not an observability dashboard or a deterministic-LLM engine. TypeScript, Node.js 22+, Apache 2.0, version 0.1.0.
Project page: jarvisosorio.com/agent-replay-lab/ · Repo: github.com/JarvisOsorio/agent-replay-lab
The common thread Read together, the four projects share a philosophy:
- Local-first. Every tool works on your machine without accounts, and none ships telemetry.
- Honest scope. Each README separates what is shipped from what is planned, and says what the tool is not.
- Safe defaults. Dry-run before write, fail closed on missing mocks, never call tools during a scan.
- CI-friendly exit codes. Findings and drift become build failures instead of surprises.
They also compose. Ambient Memory gives an agent context; mcp-pordego checks the tools it is allowed to use; Agent Replay Lab verifies its tool path stays stable; agentscrub cleans up what its sessions leave on disk — and Agent Replay Lab's README explicitly suggests agentscrub for after-the-fact store remediation.
None of the packages is on a public registry yet; each is installed from its GitHub repository.

Browse all four on the open-source page at jarvisosorio.com, or follow Jarvis Osorio on GitHub.