Conclave

For agent users who find themselves constantly copy/pasting between different models until they reach consensus on high leverage decisions. Automates a set of local agents self-organizing based on their relative strengths and iterating until an agreed best solution is reached. Uses simulated annealing for randomness and will reconcile with past results to over time find its optimal answer.

$ conclave run --panel claude,gpt4,llama \ "Is this deployment safe to ship?" → Panel: 3 models, 12 reasoning chains → Verdict: HOLD (2-1) — claude flagged auth regression

Grimoire

Security standards are opaque, ever-changing, and often complex. To facilitate security work, an automated hybrid FTS5 + vector search over MCP was created to increase fidelity, speed, and token utilization for security work.

$ grimoire index ./security-standards --hybrid → 847 chunks indexed (FTS5 + vectors) $ grimoire query "CSRF mitigation patterns" → 3 results in 8ms (keyword + semantic)

Helm Suite

Helm (RIP) was ahead of its time. To complete the healing process for this startup closing, I built out the whole vision using a NAS, Tailscale, and security best practices. With the assist of agentic coding, building your own personal, private, fast internet is completely feasible.

$ cp .env.example .env && vim .env $ ./bootstrap.sh → caddy ✓ immich ✓ monarch ✓ ollama ✓ → All services on Tailscale. Zero public exposure.

Ollama Says

Structured prompt injection testing against local models. Scored and tracked across model versions. Fully offline.

$ ollama-says run attacks/injection.yaml \ --model llama3.2 → 47 attack vectors... → 38/47 blocked (80.8%) — 9 bypasses logged