AgentFloor
One governed place for a company's AI agents: build them, share them, run them, and see what each one costs.
Private build, release 0.1.3, runs on Google Cloud
People sign in with company SSO and write skills: reusable, versioned instructions with typed inputs and outputs. They wire skills into workflows on a visual canvas and start them by hand, on a schedule, from a signed webhook or from a GitHub pull request event. Every run shows its steps, its tokens and its cost.
The rules are enforced in code, not in prompts. Permission checks run inside the tool call and fail closed. Credentials belong to whoever started the run, never to whoever wrote the skill. Every model call goes into a cost ledger, tagged with the user, run, loop iteration and skill that caused it.
The starter workflow, on every new pull request
- Load the pull requestA GitHub App event, an API call or a schedule starts the run.
- Security reviewA skill scores the change and lists its findings.
- Code reviewA second skill reads the same changes and scores them too.
- Either score below 70?If so, alert the team in Slack. If not, skip it.
- Post a summary on the pull requestResults and the total cost stay on the run, even if delivery fails.
The code is private. A walkthrough is available on request: email me.
- Visual workflow canvas. Skill steps, connector actions, Yes/No and score branches, and an "If it fails" fallback on any step. Test runs use pretend data and never really send anything, and finished runs play back step by step on a timeline.
- Runs survive crashes. Each step saves its result and the run's checkpoint in one transaction, so any worker can pick up any run. Calls that change things carry idempotency keys. If it's unclear whether one went through, a person decides; it is never retried silently.
- Budgets checked before the call. Spend is reserved up front with one conditional database update, then settled. Costs are counted once, deduplicated on the provider's request id.
- Schedules written in words, picked from a menu or typed as cron, with time zones, daylight-saving rules, a weekly budget forecast and three separate guards against firing twice. Personal schedules run on the owner's access; company schedules need a manager's approval.
- Signed webhooks with timestamp, nonce, size and rate checks, a dead-letter list, and secret rotation with a 24-hour overlap.
- Connectors: a GitHub App set up in two clicks, Slack, email, Google Drive (results trimmed to what the person running it may open) and any outside API, with every key held in Google Secret Manager.
- A publish gate. Sharing a skill across the company needs a schema check, declared permissions, the org's policy checklist and a passing test with writes mocked.
- Improve with AI turns rough notes into a draft skill, treated as untrusted input until someone reviews it.
- TypeScript end to end: Fastify, React, Drizzle on Postgres (or embedded PGlite), Claude or OpenRouter models, tested with Vitest and Playwright. It ships as one container on Cloud Run with Cloud SQL.


