The canonical framing for Writmint. Every external surface — README, docs, outreach — points here and uses this language. Last reviewed 2026-06-28.
Verifies every capability manifest your AI agents write, then hands you a clean one to approve.
An AI agent declares what a capability may do — every host it reaches, every field it touches, every action it takes. Writmint checks that declaration as the agent writes it and rejects anything underspecified, overscoped, or unsafe, returning deterministic, machine-readable errors the agent fixes on its own. By review time, you sign off on verified work instead of hunting for what the agent got wrong.
Most agent platforms surface manifest mistakes at runtime, when the capability tries to act, or downstream, when the manifest reaches a pull request — leaving your reviewer to debug the agent. Writmint moves the correction earlier and makes it deterministic, so review stays a place to approve, not repair.
One-liner, for tight contexts: Verifies AI-generated capability manifests at author time — so review is a sign-off, not a cleanup.
The difference Writmint makes is not whether someone reviews the manifest — in today's pull-request ritual they do — but what state it is in when they do: raw agent output (today) vs. a machine-verified artifact (Writmint). That repositions the reviewer's job from debugger to approver.
This is a compliance fit, not a hedge. In banking, healthcare, and insurance — Writmint's anchor scenario — segregation of duties requires that the approver is not the author. A standing approval step is mandatory there, so "you still approve" is alignment, not weakness. The agent authors and corrects its own work; you approve a manifest that arrives already verified and cryptographically bound to what you signed.
The gate — verification an agent corrects against. Two things carry it:
- Structured errors — every failure is
{code, where, expected, actual, fixHint}. Deterministic, machine-readable, the same shape for every rejection. This is the differentiated claim: the agent corrects against the contract, not against a reviewer's comments. - The capability manifest — the declarative contract the agent is making approvable: the only surface a capability has on the host system.
What makes a verified manifest trustworthy. Verification gets the manifest correct; these three make a correct manifest worth approving:
- Approval — hash-bound, lifecycle-tracked, audited. Approval binds to the SHA-256 of the manifest, so "you signed these exact bytes," not "the agent says it's clean."
- Permissions — the broker boundary. The declared scope is actually enforced at runtime; nothing outside the manifest reaches the host system.
- Replay — every execution is reproducible. Any run can be replayed against its recording to prove what happened, with strict-ordered divergence detection.
The first band is what the market leaves empty (see competitive landscape). The second is the territory runtime-governance tools already occupy — so we lead with the first and let the second land as the reason a verified manifest is also a trustworthy one.
The positioning leans on a behavioral claim that is empirically untested in shipping products: that AI agents actually self-correct manifests in a closed loop against structured rejections at scale. Evidence so far is one data point — a dogfood pass where a skill-disabled agent landed a new permission shape in two round-trips on the structured rejections alone. The reframe is a high-conviction bet that this generalizes; a measured dogfood corpus (separate work) is how we test it. If agents thrash, or converge only when the authoring skill is enabled (i.e. the value is in prose guidance, not the structured contract), this positioning must be revisited — and we will say so rather than quietly drop the claim.
Writmint's buyer values transparency and has been burned by tools that overclaim, so candor is a deliberate choice, not fine print: pre-stable software is labelled pre-stable; the "empty band" claim always carries its scope; the untested premise above is stated wherever the positioning leans on it. A vendor that tells you what is not yet proven is one you can trust about what is.