A task-graph engine fans audits out across the project portfolio, a deterministic verifier rejects incomplete work before it is accepted, and every file write is a replay of a verified report through code-enforced gates. Four executable workflows run on the engine today, operated from Studio, a hub inside the VS Code dashboard that prices a run before launching it and observes it live. Underneath: a context portfolio served over MCP, with shared skills and memory across Claude, Codex and Cursor. Context becomes infrastructure, not a paste job.
The system that builds this website can now audit its own output, identify where it undersells the project, and propose concrete improvements. The content-audit loop compares git evidence against the live site, generates a synthesis layer with repositioning recommendations, and ranks proposals by impact, effort, and risk, so 'what to do first' is a code decision, not a guess. A grounding split separates repo-backed claims from speculative ideas.
The audit generates fields like whatFeelsUnderstated and recommendedHeroShift, each backed by citations to commits and changelog entries. The model proposes; the evidence decides.
Proposals are ranked by a deterministic nextBestActionScore: impact weighted by confidence, minus penalties for effort and risk, plus a bonus for grounding. Quick-wins clustering surfaces changes worth doing together.
The latest project to join the showcase, onboarded through the formal process with multiple audit reports already generated. The loop is production-ready.
Every time a new Claude session opens, or a new agent spins up, the same question hits: does it know who I am, how I work, what the project context is? Without a system, the answer is always no, and the first ten minutes of every session are re-establishing ground that was covered yesterday.
AgentOS is the answer to that. A personal operating system for the agentic era: a persistent context layer that travels across tools, sessions, and projects.
The foundation is a Context Portfolio: ten structured markdown files covering identity, role, projects, team, communication style, decisions, and more. Deployed as an MCP server so any connected agent can read any file on demand. Versioned in git. Synced automatically at the end of every session.
Active build. Context portfolio complete and served under pm2. Claude, Codex and Cursor wired in via the shared HTTP server. Task-graph runtime live with four executable workflows, deterministic verification, and two real gated applies committed. Dashboard at v0.5.79 with the Studio hub.
The portfolio is the source of truth. The MCP gateway is the access layer. The clients are the consumers. The task-graph runtime lives inside the gateway and does the actual agent work. Git is the backup for everything.
Ten markdown files covering identity, role, tools, projects, communication style, goals, and domain knowledge. Machine-readable. Versioned in git. Changes are live immediately.
Node.js server exposing portfolio files as typed MCP resources over 127.0.0.1:3000/mcp. Runs under pm2, auto-starts on login. Tracks every read, skill call and tool invocation with token counts.
Claude Code, Codex and Cursor all connect to the same MCP server. Shared memory and 16 shared commands are deployed to all three via per-client sync scripts. One canonical source; three runtime instances.
Context Portfolio, MCP server, canonical vs runtime model, shared memory and skills layer, and how it all bootstraps on a new machine.
Fan-out audits with deterministic verification, an advisory challenger, model failover routes, and a replay-only apply engine. Every write behind code-enforced gates.
A live multi-client control plane inside VSCode. Studio drives the runtime end to end: discover a workflow, price it, launch it behind a spend confirmation, and watch it run live.
The plan/execute/review pipeline every graph node wraps, with bounded self-correction.
Three-layer config, auto-sync hooks, sub-agent delegation, and 16 shared slash commands.
Twelve MCP tools; OpenRouter role profiles route planning, tooling, and agentic work pay-per-token.
Streamable HTTP to the shared server; command logging via hooks; shared jn-* skills deployed.
Built session by session since April 2026: context portfolio, local MCP server, multi-client wiring, multi-provider models, cost tracking, a VSCode extension, and then a task-graph runtime with deterministic verification and gated applies on top. The build log records every phase.
read the build log →The conceptual and practical foundations draw from the AI Daily Brief's work on agentic systems and context management.
Deep dive into the Personal Context Portfolio concept and how to deploy it as an MCP server.
The episode that started this. Introduces the AgentOS framework and the free AIDB training programme.
The official AIDB AgentOS project hub: a structured programme for building a personal agentic OS.