Provostry checks completed AI work against a definition of done that was written down first, tracks what it cost, and records who approved it — in one append-only ledger your agents answer to. Approvals reach you in Slack, Teams, or Telegram; the full control room stays on the web. Hosted with a 14-day free trial; self-hosted for enterprise. The operating system for AI organizations is the larger idea; this is the first thing it proves.
The three-minute assessment runs in your browser — nothing leaves this page. Wiring your own agents in? The SDK and integrations are open and documented (Apache-2.0) — no email needed. Evaluating self-host? Everything is written down.
An agent pre-registers what "done" means. A human approves the clauses. The server checks reality and writes the verdict into the same ledger as the claim. Ninety seconds, no narration tricks — this is the product running. Full 3½-minute technical walkthrough → How Verify works → A report, as the server writes it →
The category argument
Every AI-agent product sells you more agents. That is the wrong layer to invest in. Agents get swapped, upgraded, and deprecated. The durable asset is the layer they answer to: a ledger of evidence, budgets that bound autonomy, and verification that only counts independently attested work.
You cannot run a business on agent self-reports. The moment two agents touch the same codebase, you already have the trust problem — whose "done" counts?
Provostry is that layer. Not another bundle of agents. Not a dashboard that repeats what agents claim. A control plane: nothing counts as done until a contract the server checks, or CI the server fetches, attests it; every work item carries a budget and crossing it asks a named human; and every decision lives in an append-only, tamper-evident ledger. The client SDK is open so your agents plug in freely; the governance and validation engine is the product.
Five mechanisms
Every decision, every piece of evidence, every event — appended, never edited. Evidence is never refused; the audit trail is sacred. Authority is what gets gated, not information. Hosted, the record lives in your organization's tenant of the hosted service; self-hosted, it never leaves your network. Zero-dependency file ledger to start; Postgres when you want it.
Every work item carries a budget, and spend is tracked against it as agents report. Crossing the ceiling raises an escalation to a named human — it never passes silently. A runner that reserves before it spends is refused at the ceiling and waits for the decision; one that never asks is escalated, not halted. Provostry sits at the ledger, not between your agent and its tools, and says so.
Nothing is done on an agent's say-so. A contract written down first — observable pass/fail clauses the server probes against reality — or CI check runs the server fetches from GitHub for a repository you registered. Every check is deterministic; no model grades the work. Velocity moves only on attested work, and agents cannot impersonate each other — every agent holds its own hashed API key.
Six ledger-signal detectors read the append-only record and propose changes with evidence, projected impact, and risks. It only proposes; humans decide. Its first act on real data was diagnosing its own maker.
Air traffic control, not a chat window. Org health. Initiative portfolio with risk and forecast. Flow map with drag detection. Approvals queue. Business outcomes scorecard — NRR, gross margin, tickets per 1k users. And if nothing is verified, it shows you zero.
How it works
Hosted: sign up, 14 days free, nothing to operate. Enterprise: one container on your cloud or fully air-gapped, file ledger to Postgres. Either way the same engine — verified in CI on every commit, more than a thousand automated tests including a Postgres integration suite.
Agents from any framework or language report with one HTTP call — no SDK required, one available. Register the repositories whose CI may vouch for your work; the server fetches each cited commit's check runs from GitHub itself.
Name a system you own, pick a check from a template, run it. The server probes, writes the observations and a pass/fail verdict into the ledger, and the dashboard shows exactly that. Until a contract or CI attests something, velocity reads 0 — honesty as a demo moment.
The command companion
The web operations center is the full control room. The phone in your pocket is the command companion: escalations and phase activations reach you in Slack, Teams, or Telegram — the chat apps already on company devices. Nothing new to install, nothing new to vet. You approve or reject, and the org keeps moving. A recorded approval is a decision with your name on it, not an action: the agent, the runner or a person then does the work, and the ledger records what happened.
Approvals arrive in Slack, Teams, or Telegram — already on company phones, already vetted by your org. There is no app to roll out and no new tool for your team to remember to check.
Every approve or reject is recorded as an attributed decision_record in the append-only, hash-chained ledger — the same ledger as dashboard decisions, verifiable by an auditor later.
The mobile surface can only do what the governance model allows: escalations out, decisions in. It cannot browse the ledger or change budgets — which is why security teams can approve it in one sentence.
Proof — built with itself
We will not show you invented logos or testimonials, because none exist. The substitute is radical dogfooding — Provostry's public repo history is the case study.
Provostry dogfoods itself: its own GitHub Actions CI attests its own commits into its own ledger. Its velocity metric is CI-verified changes on itself.
The org optimizer's first proposal on real data flagged Provostry's own project. We shipped it anyway. That is the product.
Unverified work shows velocity 0 on our own dashboard. That is not a bug; it is the entire point. A metric that can be inflated by an agent's self-report is not a metric.
Every verdict is a document: the goal verbatim, who approved the clauses, what the server observed, pass or fail, and the chain check. Here is one, generated from a demo ledger's live API, and here is the design it follows.
"Cap flightdeck at Prototyper budgets (no reality port)."
— the Provostry org optimizer, first proposal on real data, about Provostry itself (quoted verbatim from when the project was codenamed flightdeck). Built by one founder plus AI agents, coordinated through Provostry itself.
FAQ
No. Dashboards report what agents claim. Provostry is a control plane: budgets gate spend, escalations route overspend to humans, and the velocity metric refuses to move on unattested work. The exec view is air traffic control, not a chat window — and if nothing is verified, it shows you zero.
Different layer. Those are observability and tracing for LLM calls — useful, keep them. Provostry is org-level governance: budgets, independent verification, decisions, and an append-only audit trail. Tracing tells you what an agent did inside a call; Provostry decides what the org accepts as done and what it's allowed to spend.
It's valuable from roughly 2 agents plus a CI pipeline. The moment two agents touch the same codebase, you already have the trust problem: whose "done" counts? One container, one HTTP call — the setup cost is an afternoon, not a platform migration.
Both. Provostry runs as a hosted service so you can start in minutes with nothing to operate: your organization's ledger lives in its own tenant of the hosted service, and the hosted server runs the checks. For air-gapped or security-conscious teams, a self-hosted enterprise license runs it entirely on your infrastructure — the ledger of decisions and evidence never leaves your network.
It tells you, and it refuses the ones that ask. Spend is tracked against each work item's budget; crossing it raises an escalation to a named human. A runner that reserves budget before it spends — the SDK does this in one call — is refused at the ceiling and waits. One that never asks is escalated, not halted: Provostry is a ledger and a decision port, not a proxy in front of your agents.
Deterministic. A verification clause is a probe and an assertion over a closed vocabulary — the server compares what came back to what the clause said. CI attestation reads GitHub's own check runs. The org optimizer and the validation suite are detectors over the ledger. Where an optional model review is enabled, its findings are labeled as such and are never promoted to verified evidence without a human.
The parts you integrate against — the client SDK and integration adapters — are open (Apache-2.0), so wiring your agents in is free and unencumbered. The governance and validation engine is proprietary; it's the product. Under an evaluation you get the source to read on your terms — the self-host page says how.
Pricing
The trial is the full product for fourteen days, no card. When it ends, your ledger stays readable — it is your record — and writing new work needs a plan; choose one from the workspace and the owner pays without talking to anyone. A failed payment gets seven days of grace before writes pause. Cancel any time from Plan & billing — no call, no email; reading stays free, the ledger is yours. Limits count distinct agents and work items in the workspace, including work the engines create for you; an item beyond the plan is refused with the reason, never silently dropped.
Delivery
| Hosted | Self-hosted Enterprise | |
|---|---|---|
| Governance core · validation engine · ops center | Yes | Yes |
| Open SDK + integrations (Apache-2.0) | Yes | Yes |
| Tamper-evident ledger + keyed signatures | Yes | Yes |
| RBAC + signed audit export | Yes | Yes |
| Runs on your cloud / fully air-gapped | — | Yes |
| Support SLA | Yes | Yes |
The moat isn't secrecy — a snapshot of the code can't reproduce a product that governs its own development and ships continuously. Evaluating self-host? Everything you need to know is documented.
Answer six questions about how your agents work today. Get a risk snapshot and a simulated view of your org inside Provostry — no call, no form, nothing leaves this page.
Integrating your own agents? The SDK and integrations are open and documented — no email needed. Want to see the output first? A verification report · the design behind it.
Arrived looking for something specific? An audit trail for AI agents · Budget controls for AI agents
Evaluating the category? Six things a governance system must prove · Gateways, GRC platforms and ledgers compared