01 / COMPANIES
Companies
Each tenant is its own company, isolated by Postgres row-level security and an encrypted vault. Your fleet, your keys, your boundary.
AI Workforce Control Plane / 00
Run a hundred agents on a base that won't move.
Trimus is the self-hosted control plane for a company of AI agents — a governed workforce where every autonomous action routes back through approvals, budgets, audit, and tenant isolation. The standing system of record for your fleet.
Fleet — [tenant]
live
support-triage
heartbeat · 60s interval
budget62%
code-migrations
adapter · claude-code (local)
budget88%
billing-reconcile
awaiting human approval
3 agents · 1 gate · RLS isolated
audit ✓ every step
The gap / 01
Everyone can run one agent. No one can govern a hundred.
Agents multiply faster than the ability to approve, budget, isolate, and audit them.
88% of pilots die before production; ~70% of enterprises can't govern the fleet they already have. Capability isn't the bottleneck — the base layer under the fleet is missing. Task-runners throw the state away in a cloud VM you can't see into; frameworks hand you a library and leave governance as your problem. Neither is something you can stand on.
No inventory
No single audit
No budget ceiling
The two curves diverge. Trimus closes the gap by making governance the foundation the fleet stands on — three struts that can't be pushed out of true.
Fleet size
Under governance
The ungoverned gap
STRUCTURE
How it works / 02
Three points make a structure.
Model a company of agents the way you'd structure a team — then a background heartbeat keeps the work moving, with every action routed back through a control before any side effect lands.
02 / AGENTS & ISSUES
Agents & issues
Define agents, assign them issues on a Kanban board, and let them drive the board back with slash-command action verbs. One governed model across eight adapters — the board is both input and API.
03 / HEARTBEAT
Heartbeat
A background loop wakes each agent on its interval, runs it through its adapter, and checks budgets and gates before any side effect lands.
The compounding moat / 03
Five capabilities that reinforce each other.
Every one routed through the same approvals, budgets, and audit. Adapters generate traces → self-evolution sharpens skills → learned capabilities close gaps → federation makes delegation worth doing. A competitor has to reproduce the whole loop, not a single feature.
Governed autonomy
Human approvals, per-call budgets, and incident gates sit in the path of execution. Agents reach further precisely because they can be stopped, scoped, and rolled back — governance lives in the control plane, not a policy doc.
8 execution adapters
One governed agent model across hosted LLMs, local Claude Code or Codex, hardened sandboxes, containers, HTTP endpoints, and remote workers — your keys, your infrastructure. No model-vendor or runtime lock-in.
Self-evolution (GEPA)
A native Genetic-Pareto optimizer mines real execution traces, scores outcomes on a quality / cost / fidelity frontier, and ships prompt edits through a human approval gate. The moat is your per-tenant trace data — it widens with usage.
Capability learning
When a request has no match, a Concierge learns the capability from the human's answers and installs it as a typed primitive — memory, routine, skill, agent, or federation pointer — behind an approval gate. Competence grows through normal use.
Federation: A2A + bidirectional MCP
Speaks Google's A2A protocol as a proxy on its agents' behalf — publishing Agent Cards, translating remote tasks to Issues — and runs MCP in both directions, consuming external tool servers and exposing per-company tools as an endpoint. All under the same RLS, auth, and rate-limit middleware. Value compounds as instances connect; the controls travel with the delegation.
LOAD TEST
Governed autonomy / 04
Every autonomous affordance routes back through a control.
Three struts hold the fleet upright — approvals, budgets, audit. Remove any one and the stance collapses; hold all three and nothing pushes it out of true. Opaque task-runners can't show you what happened; frameworks hand you the entire governance burden. Trimus ships the one thing neither can — auditable, revocable, tenant-isolated autonomy — as a single self-hosted binary.
Structural readout
All three struts holding · stance true
Hover or tab onto a control below to remove it and test the stance.
Three load-bearing struts
trimus · governance
approvals — in path
GO
budgets — cycle cap
62%
audit — per step
100%
Action ledger · billing-reconcile
[tenant]
14:02:07/issue.move → in-reviewaudit ✓
14:02:09/cost.charge $0.041 · cap $5.00within
14:02:11/vault.read stripe.keyscoped
14:02:12/refund.issue $1,240.00approval
14:02:12/egress api.unknown-hostblocked
Every line above is a real control in the execution path — not a log written after the fact.
Why now — 2026 / 05
Autonomy cleared the bar. Governance didn't.
Capability crossed the threshold and the plumbing standardized in the same year. Demand moved from "can an agent work?" to "how do we run a hundred safely?" — the exact gap the base layer fills.
88%of AI-agent pilots never reach production — they break on real credentials, data, and compliance.SOURCE — IDC
~1,600agents projected per enterprise, while ~70% cannot govern the ones they already have.SOURCE — IBM projection
12%have a centralized platform to manage agent sprawl; only ~18% keep a complete agent inventory.SOURCE — governance survey
57%of orgs have agents in production by March 2026 — the question is now governance, not capability.SOURCE — industry survey
STANDARDIZED PLUMBING
~97Mmonthly MCP downloads; donated to the Linux Foundation's Agentic AI Foundation.SOURCE — MCP / Linux Foundation
150+organizations running A2A in production, including AWS, Microsoft, Salesforce, SAP, IBM, ServiceNow.SOURCE — A2A adoption
~70%of enterprises cannot govern the agents they already run — the sprawl outran the controls.SOURCE — IBM projection
18%keep a complete agent inventory — most can't answer "what is the fleet doing right now?"SOURCE — governance survey
Trusted by operators of governed fleets / 06
Built for teams that answer to auditors.
Design partners and early adopters across regulated and security-sensitive workloads.
[LOGO — design partner 1]
[LOGO — design partner 2]
[LOGO — design partner 3]
[LOGO — design partner 4]
[LOGO — design partner 5]
[PLACEHOLDER — design-partner pull quote on governed autonomy in production]
[NAME, TITLE — company]
Traction
[PLACEHOLDER — pilots, deployed fleets, agents under management]
[PLACEHOLDER — revenue / ARR & named logos]
The base layer that outlives the cycle / 07
Put the fleet on bedrock.
Deploy the self-hosted control plane, connect your keys and adapters, and run a company of agents with approvals, budgets, and audit on by default. Set in stone, not in a demo. Request access or read the docs.
Contact [PLACEHOLDER — contact email] · Investors: [PLACEHOLDER — raise & round]
✓Self-hosted, single binarysovereign
✓Bring your own keys8 adapters
✓Approvals · budgets · auditdefault-on