Blog
Engineering notes from building Preloop — the open-source AI agent control plane. Mostly real bugs with real diagnoses, plus the occasional reference piece on how agent governance actually fits together.
-
· 6 min read
Preloop 1.0.0: a compatibility promise, your laptop as a runner, and reviews that must land
1.0.0 is the first Preloop release under a published compatibility policy. It adds MCP protocol 2026-07-28, enforced API key scopes, credentials encrypted at rest and tested upgrades, puts users and teams in open source, turns your own machine into a runner on your own agent seat, fails reviews the code host never accepted, and brings single sign-on with SCIM to Preloop Cloud and Enterprise.
-
· 8 min read
Preloop 0.17.0: governing the devices and people behind your agents, plus local models and self-hosting
0.17.0 takes a first step past agents and models: a live Fleet Map and a device fleet console that links devices to the agents and models they run, staged MDM recipes, owner consent before anyone runs work on your machine, and personal runners. It also brings runner-proxied local models, native host execution for Codex, OpenCode and DeepSeek Harness, a self-hosted path with GitLab, plain-language governance policies and a redesigned weekly review.
-
· 5 min read
We let three flows run our backlog. Here is one issue, start to merge, for 76 cents
A real run on the Preloop repository: a one-line bug report went through triage, implementation, review, repair and re-review by three open source flows, with zero human actions before the merge. About $0.76 in model cost.
-
· 7 min read
Preloop 0.16.0: flows for software factories, Alibaba Cloud Model Studio and AWS Bedrock
0.16.0 turns flows into the parts of a software factory: flows that start other flows under one cost ceiling, notes that let an agent steer the subagents it started instead of killing them (and let humans steer both), runners that hold several jobs and a one-shot mode behind a GitHub Action, a guard that stops runs going nowhere, and a searchable record under retention and legal holds. Alibaba Cloud Model Studio (Qwen) and AWS Bedrock join as model providers with live discovery, plus a repo review preset family and OTLP export.
-
· 6 min read
The Claude Code audit trail has three layers. Anthropic operates two of them.
Securing Claude Code takes more than enabling the Compliance API. The audit story has three layers: the org control plane, OpenTelemetry, and a self-hosted proxy and tool layer where enforcement actually lives. What that third layer looks like when it is purpose-built.
-
· 3 min read
Preloop 0.15.0: flow schedules, self-hosted runners, security audit presets
0.15.0 runs flows on native schedules, leases jobs onto your own machines with preloop runner, ships a CRA-oriented security audit preset pack, makes flows eval-grade with matrix fan-out and result.json, and stops showing unpriced usage as $0.00.
-
· 1 min read
Preloop 0.14.0: Cursor usage import, rate-limit intelligence, agent identity v2
0.14.0 lands Cursor bundled-model usage in Cost analytics, turns upstream 429s into a headroom report, and gives managed agents a durable id across renames.