Show HN: Aaf-gate – policy gate and kill switch for existing agent runtimes A developer released aaf-gate 1.0.0rc1, a policy gate and kill switch that sits in front of tool calls for existing AI agent runtimes, with its primary target being Hermes Agent pinned at v0.23.1 for a Mac mini appliance track. The tool applies deterministic per-tool-class decisions (read, draft, internal_write), adds a gate-stop/gate-resume/gate-status kill switch, per-window call limits (3600 s; read 160, draft 40, internal_write 80), and an append-only hash-chained SQLite audit log, using only the Python standard library in core/ with 719 local unit tests passing. The author states aaf-gate is not production-ready, not yet device-accepted, makes no compliance claims, and has not been granted Google Gmail Developer Preview access; Pilot 1 scope is reading Gmail, categorizing, and drafting a reply, with no send capability. Deterministic policy gate, kill switch, call limits and tamper-evident audit for existing AI agent runtimes. Not an agent of its own. Models, providers and the agent loop are not reimplemented. The primary runtime is Hermes Agent https://github.com/NousResearch/hermes-agent pinned at v0.23.1 for the Mac mini appliance track . Adapters can target other runtimes later. Core version 1.0.0rc1 v1 candidate . See CHANGELOG.md https://github.com/buberlo/aaf-gate/blob/main/CHANGELOG.md , docs/v1/STATUS.md https://github.com/buberlo/aaf-gate/blob/main/docs/v1/STATUS.md and docs/v1/INSTALLATION-AND-OPERATIONS.md https://github.com/buberlo/aaf-gate/blob/main/docs/v1/INSTALLATION-AND-OPERATIONS.md . aaf-gate sits in front of tool calls and applies a deterministic policy per tool class: - read — allow or deny reads for example Gmail search - draft — allow or deny drafts for example a reply draft; sending is out of scope - internal write — allow or deny internal writes for example labels ; unattended internal writes stay blocked It also provides: - a kill switch gate-stop / gate-resume / gate-status that stops further work for a principal - call limits per time window and tool class tool-call counts, not model-token cost - an append-only, hash-chained audit log in SQLite - backup and restore with a chain anchor gate-backup / gate-restore - versioned, locked industry templates and example customer profiles - a device acceptance pack for a Mac mini appliance track A Pilot 1 scope: read Gmail, categorize, and draft a reply. No send. tool call → gate class decision → GateKeeper policy, limits, kill switch, audit → hermes adapter Hermes hook mapping → gatesocket Unix socket → Hermes plugin aaf-gate 0.5.0 pre tool call / post tool call Python standard library only in core/ . No third-party Python dependencies. This is a v1 candidate 1.0.0rc1 . It is implemented and unit-tested locally. It is not production-ready , not yet device-accepted , and it makes no compliance claims . | Implemented and unit-tested locally | yes Ran 719 tests , OK , Python stdlib | | Tested on the target Mac mini | no | | Google Gmail Developer Preview access | not granted yet | | Production | no | | Send mail, computer use, own agent loop | not included | Start values for the call-limit window are assumptions in the Pilot 1 policy 3600 s; read 160, draft 40, internal write 80 and will be reset after a pilot week from recorded gate calls . cd core python3 -m unittest discover -s tests -q python3 -m framework.cli version version prints core version , schema version and ok as JSON. Further commands python3 -m framework.cli