W3villa
Support teams (50–500 agents)
Compile refund and escalation policies once. Every agent — human or AI — executes the same approved version instead of improvising.
Turn company knowledge intoTurn company knowledgeinto
AI
executable skills
ModusBrain compiles your policies, runbooks, and wikis into approved, versioned operational skills that AI agents execute consistently, with human oversight and a full audit trail.
$npm install -g @genthropic/modusbrain
ModusBrainExample workspace — not live data
Skills
FilterSortOptions
Skill
Confidence
Compiled By
Risk Tier
Status
Runs/Mo
Docs
Category
Lead Reviewer
Runs/Day
Runbooks
Added
R
Refund Handler95%
JC
Jane CooperHigh Stakes
Approved
12,400
refund-handler
Finance Ops
JC
Jane Cooper412
R
Refund EscalationJul 1, 2026
U
User Provisioner88%
AK
Alex KimHigh Stakes
Pending Approval
8,200
user-provisioner
Identity & Access
MT
Maria Torres274
A
Access ReviewJul 3, 2026
L
Lead Scorer78%
DP
Devon ParkLow Stakes
Approved
31,800
lead-scorer
Revenue Ops
PN
Priya Nair1,060
Q
Qualification RunJul 5, 2026
C
Compliance Auditor92%
API
API - Key nameInformational
Pending Approval
5,600
compliance-auditor
Compliance
SR
Sam Reyes187
S
SOC2 AuditJul 8, 2026
I
Invoice Processor94%
WF
Workflow nameHigh Stakes
Approved
19,100
invoice-processor
Finance Ops
LC
Leo Chen637
AP
AP ApprovalP
PO MatchJul 12, 2026
S
Support Triage91%
NW
Nora WhitfieldLow Stakes
Approved
9,300
support-triage
Customer Support
EB
Ethan Brooks310
E
Escalation PathJul 14, 2026
D
Data Enricher89%
RM
Ravi MalhotraInformational
Approved
47,800
data-enricher
Data Ops
CL
Casey Lin1,593
E
Enrich RunJul 15, 2026
C
Contract Reviewer86%
JL
Jordan LeeLow Stakes
Approved
18,000
contract-reviewer
Legal Ops
JL
Jordan Lee600
M
MSA ReviewJul 17, 2026
D
Deployment Monitor97%
TV
Taylor VanceHigh Stakes
Pending Approval
64,000
deployment-monitor
DevOps
TV
Taylor Vance2,133
A
Alert TriageJul 20, 2026
E
Email Drafter93%
MW
Morgan WestLow Stakes
Approved
22,400
email-drafter
Comms Ops
MW
Morgan West747
O
Outbound CampaignJul 21, 2026
R
Report Builder85%
CL
Casey LopezInformational
Pending Approval
11,250
report-builder
Analytics
RW
Robin Wu375
W
Weekly DigestJul 23, 2026
R
Risk Classifier84%
RM
Reese MillerHigh Stakes
Approved
33,800
risk-classifier
Security
PL
Parker Lopez1,127
T
Threat ScanJul 24, 2026
P
Policy Enforcer96%
SYS
SystemLow Stakes
Approved
7,900
policy-enforcer
Governance
PW
Peyton Ward263
P
Policy SyncJul 25, 2026
O
Onboarding Agent90%
SP
Sunny PatelInformational
Pending Approval
55,200
onboarding-agent
Platform Ops
SP
Sunny Patel1,840
U
User JourneyS
Seat ActivationT
Trial ConversionJul 1, 2026 2:25 pm
Calculate
Compile a refund policy into an operational skill with confidence gating, human-in-the-loop approval for high-risk transactions, and a full audit trail.
…my-modusbrain-app/src/constants/schema-identifiers.ts
+84-01export const SCHEMA_IDS = {
2 rocket: { object: '733956fd-…', fields: { launches: '5b877c2a-…' } },
3 launch: { object: 'e7f1e750-…', fields: { rocket: '42c9106f-…' } },
4 payload: { object: '16ffcc45-…', fields: { customer: 'd84468aa-…' } },
5 launchSite: { object: '2f18d525-…', fields: { launches: 'b94b7f00-…' } },
79 unmodified lines
…my-modusbrain-app/src/objects/launch.object.ts
+237-01import { defineObject, FieldType, RelationType } from 'modusbrain-sdk';
3export default defineObject({
4 nameSingular: 'launch',
5 labelSingular: 'Launch', labelPlural: 'Launches',
6 icon: 'IconRocket',
7 fields: [
8 { name: 'missionCode', type: FieldType.TEXT, isUnique: true },
9 { name: 'status', type: FieldType.SELECT, options: [ … ] },
10 { name: 'plannedLaunchAt', type: FieldType.DATE_TIME },
11 { name: 'rocket', type: FieldType.RELATION, relationType: RelationType.MANY_TO_ONE },
226 unmodified lines
…my-modusbrain-app/src/objects/payload.object.ts
+198-01import { STANDARD_OBJECT } from 'modusbrain-sdk';
4 { name: 'payloadType', type: FieldType.SELECT },
5 { name: 'customer', type: FieldType.RELATION,
6 relationTargetObjectMetadataUniversalIdentifier: STANDARD_OBJECT.company.universalIdentifier,
7 universalSettings: { relationType: RelationType.MANY_TO_ONE, joinColumnName: 'companyId' },
8 },
190 unmodified lines
…my-modusbrain-app/src/objects/rocket.object.ts
+28-32140 unmodified lines
141
141 {
142 name: 'launches',
143 type: FieldType.RELATION,
144 relationType: RelationType.ONE_TO_MANY,
145 relationTargetObjectMetadataUniversalIdentifier: SCHEMA_IDS.launch.object,
146 },
…my-modusbrain-app/src/views/upcoming-launches.view.ts
+82-01import { defineView, ViewFilterOperand, ViewType } from 'modusbrain-sdk';
3export default defineView({
4 name: 'Upcoming launches',
5 type: ViewType.TABLE,
6 filters: [
7 { fieldMetadataUniversalIdentifier: SCHEMA_IDS.launch.fields.plannedLaunchAt,
8 operand: ViewFilterOperand.IS_IN_FUTURE },
9 ],
73 unmodified lines
…my-modusbrain-app/src/__tests__/schema.integration-test.ts
+412-40150 unmodified lines
151 expect(application.objects).toHaveLength(1);
151 expect(application.objects).toHaveLength(4);
152 expectObject(application.objects, SCHEMA_IDS.launch.object, 'launch', 'Launch');
153 expectObject(application.objects, SCHEMA_IDS.payload.object, 'payload', 'Payload');
154 expectObject(application.objects, SCHEMA_IDS.launchSite.object, 'launchSite', 'Launch site');
155 expectRelationPair(payloadCustomerField, [payload.object, STANDARD_OBJECT.company.universalIdentifier], [...]);
325 unmodified lines
used by
open source
Le problème.
Your AI agents hallucinate, because your knowledge is scattered and unapproved
Agents improvise
When AI executes directly against raw docs and chat responses, it applies policy inconsistently and misuses outdated information.
No paper trail
Without versioned, approved skills and execution logs, you can't audit what your agents did or prove which policy was active at the time.
From scattered docs to executable skills.
Compile once, execute consistently, audit everything
ModusBrain gives your agents a standing operational brain — not one-off chat answers.
Compile from real knowledge
ModusBrain reads your existing policies, runbooks, and wiki pages and compiles them into structured operational skills with version history and provenance tracking.
Gate before agents act
Every skill gets a risk tier. High-stakes actions are blocked for agents until a human approves them — so agents act consistently, not creatively.
Audit everything, always
Every agent execution is logged — which skill, which version, what confidence score, what happened. Exportable for compliance and dispute resolution.
01
02
03
Begin with production-grade agent governance
Compose your operational AI agents, runbook policies, and safety gates with a single unified toolkit: data model, audit logs, and confidence rules.
Audit executions in real-time
Track every agent run, monitor confidence gating scores, and review trigger events as they happen. Tune your thresholds dynamically without interrupting execution.
Stay in control of your AI infrastructure
Don't lock your critical business logic into proprietary model provider APIs. Deploy ModusBrain locally, run verification tests, and version control your rules like standard software.
Policies· 840
Domain
Compétences
22 champs
Compétences· 120
Source Policy
Executions
39 champs
Skill Versions· 48
Parent Skill
Policy Snapshot
8 champs
Executions· 64
Skill Version
Agent Context
11 champs
Approvals· 45
Skill
Execution
6 champs
Déclencheur
Skill is Triggered
Action
Evaluate Risk Tier
Action
Match Confidence
Action
Agent IA
Action
Log Telemetry
Action
Request Override Token
Action
Commit Audit Trail
Général
URLrefund.ops.internal
Skill Owner
Dario Amodei
Dario AmodeiSource Policyrefund_policy.md
Gated✓ True
Executions500
Espace de travail
Compétences
Agents
Executions
Approvals
Audit Notes
Telemetry
Rule Chains▾
Rule Chains
Execution Runs
Rule Versions
+ New skill✧ Evaluate✎ Edit rules
Champs
Mise en page
Général
GénéralTerminé
Skill Owner·Relation
Executions·Nombre
Supplémentaire
Gated·Booléen
Ajouter une section
Autre
Version·Nombre
Source Policy·Text
Activation date·Date et heure
Nouveaux champs·Position/visibilité par défaut pour le champ…
Ajouter une section
The architecture that makes agents trustworthy.
Skills your agents execute reliably, every time
Versioned operational skills
Every skill has a version number, edit history, and a record of which source material it was compiled from.
Confidence-gated execution
Agents execute against approved skills, not fresh chat responses. Confidence scores and risk tiers control what runs automatically and what needs a human.
Full action audit trail
Every execution logs the skill name, version, confidence score, and outcome. Export for SOC2-style compliance reviews.
Built for every team.
Operational knowledge that scales with your team
AC&T
Fintech and insurance
High-stakes tier and approval tokens for any action involving money or regulated decisions. Conflict queue catches contradictory policy updates before they cause incorrect actions.
NetZero
Legal and compliance
Version history proves exactly which policy was active when an agent took an action — critical for audits and disputes.
1/3
Eux, ce sont les vraies ventes
C’est vraiment la flexibilité qui a fait la différence. Nos besoins évoluent très vite. Je découvre un nouveau besoin et en deux clics j’y réponds. C’est un vrai avantage quand on avance vite.
Nous ne voulions pas rafistoler le problème. Nous voulions construire quelque chose sur lequel des institutions pourraient compter à grande échelle, et cela exigeait une fondation assez solide pour porter toute la complexité de notre vision.
It is just such a nicer experience than dealing with rigid proprietary governance systems. My mission has been to get every operational process API-accessible via compiled skills, so everything talks to each other. ModusBrain made that possible in a way older systems simply do not.
Shubham Chavan
founder & CEO at Genthropic
Yes. ModusBrain is a self-hosted, source-available tool. A developer can install it, initialize a brain, and run the full compile → approve → execute → audit workflow locally or on their own infrastructure.
Install in minutes with Bun. Run `npm install -g @genthropic/modusbrain`, then `modusbrain init` to get a zero-config embedded database up and running. Your first operational skill can be compiled in under an hour.
ModusBrain reads wiki pages, policy documents, support tickets, runbooks, and any text-based knowledge source. You point it at your existing material and it drafts the skill from there.
When an agent goes to execute a skill, ModusBrain scores how confident it is that the skill applies to the current context. Low-confidence or high-stakes executions are held for human approval before any action is taken.
Yes. ModusBrain connects to your agents via MCP (Model Context Protocol) with OAuth-based access. Your agents call a skill by name and ModusBrain handles the versioning, gating, and logging.
The conflict is routed to a named owner instead of silently picking one answer. Teams can configure whether to serve the last-approved version or block execution entirely until the conflict is resolved.
The self-hosted open-source tier is free — full CLI, embedded database, and the complete operational-skills workflow. A hosted team tier with shared approval workflows and OAuth agent access is on the roadmap.