We run a small team of Grok Bot agents on one shared computer, and every fix we make for that setup gets rebuilt from scratch as a generic, stdlib-first tool with example data. Tonight’s batch:
agent-deadman-failover - what runs when your agent doesn’t. The agent writes a pulse file, an outside watchdog switches to failover on a stale pulse or 90% usage (back below 80%, so no flapping), and only critical jobs run as plain scripts with an ordered LLM fallback. 50 offline tests.
https://github.com/globe-teck/agent-deadman-failover
routine-heartbeat - catches the run that never happens. An expected-fires registry (cron + grace per routine) compared against heartbeats, alerting once per missed run. DST-safe. 47 tests.
https://github.com/globe-teck/routine-heartbeat
hosting-api-guard - a safety layer for agents calling hosting control-panel APIs (like the one that comes with a reputable hosting plan with a hosting company): read-only by default, --apply to write, no backup means no write, docroot and config-file guards, token redaction, rate limit, audit log. 42 tests against a mock server.
https://github.com/globe-teck/hosting-api-guard
locked-digest - let an agent fill in a daily report without letting it rewrite it. The template is locked by SHA-256, only typed values are accepted, and structure/leak self-checks hold anything off-template. 43 tests.
https://github.com/globe-teck/locked-digest
llm-latency-bench - TTFT, inter-token p50/p95/p99 and tok/s with 95% CIs for local models on CPU, plus profiling. Preallocating the KV cache gave 19-29% faster decode at 512 tokens. 36 tests.
https://github.com/globe-teck/llm-latency-bench
llm-redteam-harness (update) - AgentDojo prompt-injection suites now run fully offline with scripted agents (not a model evaluation). A tool allowlist cut travel attack success from 82.9% to 3.6% but banking only from 97.9% to 31.9%, so allowlists aren’t a full defense. Now 90 tests with a 95% coverage gate.
https://github.com/globe-teck/llm-redteam-harness
Everything is free and runs offline. Issues and pull requests are welcome, and I’d love to hear what breaks on your setup.