One useful outcome. A clear path to get there.
Understand the work, agree the boundaries, and prove the result with the people who will use it.
From scattered context to a working system
Understand. Repetitive work, the people involved, existing sources, current effort.
Scope. An outcome, access boundaries, deliverables, what success looks like.
Build. Configure knowledge and tools. Keep consequential actions reviewable.
Prove. Real examples, access controls, failure handling, everyday use.
Run or hand over. Ongoing support, or a documented handoff to your team.
Illustrative example: new client onboarding
Review approved intake → prepare workspace and context → draft brief and checklist → a person reviews → keep decisions and sources current. External sharing, invitations and contractual commitments always require explicit approval.
You own the brain. We build the walls around it.
Every client gets their own accounts on every service that holds their data, their own vault, and their own company memory. We can read down into your node to do the work; nothing of yours can be seen by another client, and nothing of ours flows into yours.
- 01
Moss Logic
Templates, playbooks, sandbox. Never your data.
- 02
Your accounts
Database org, hosting, AI workspace, vault.
- 03
Your company memory
Search, one API door, a token per computer.
- 04
Your agents
Any agent that supports MCP; the same context on every machine.
What every client gets
| Part | What it is |
|---|---|
| Client repo | Working rules for people and agents: never commit to main, no secrets in files, a decision log |
| Company memory | Full-text and semantic search, soft delete, an author on every row |
| One API door | The only way into the data; admin keys never leave the database |
| AI agent connection | Installed once per computer, token in the OS keychain, revocable, read-only option |
| Context pack | Nine files about your business, loaded by your agents before they answer |
| Session capture | Each agent session writes a summary back to your memory |
| Autonomy gates | Manual → assisted → autonomous; production writes stay human-approved |
| AI governance layer | Use policy, audit trail, review-on-change, kept current with AiGovOps practices |
Before any agent touches your systems
- Signed agreement, written scan scope and stop-line, deposit cleared.
- Password manager with MFA on every admin account; a shared vault with us. Credentials never travel by email, text or chat.
- Read-only security baseline scan. Findings ranked P1 / P2 / P3.
- Every P1 fixed, by you or as a paid fix. Adding agents to an exposure widens it.
- Backups confirmed (restore tested in the last 90 days) or declined in writing. We snapshot anything before we change it.
- Each agent framework in its own isolated environment with its own scoped token. Read-only by default.
- Least-privilege access with a named owner on your side who can revoke it.
Every project gets the same six documents
discovery · research · architecture · plan · progress · decisions. progress is the first file any new person or agent reads.
Autonomy is earned, one layer at a time
Manual: a human does it; we watch it succeed and capture what we learned. Assisted: an agent does it on demand; a human reviews before anything lands. Autonomous: runs on a schedule; only exceptions reach a human. What stays human by design: anything touching customers, money or production; adopting a new model or tool; destructive changes.
Two ways every build ends
Managed: we keep holding and running your accounts under a Service Agreement that lists every offering and cost on its own line. You still own the work and can ask for a transfer any time. Handoff: once invoices are settled, your accounts move to you whole; we rotate everything we ever saw, remove our access, walk your admin through it, and stay on email for 30 days. Offboarding fails closed: if a step can’t be verified, our access stays revoked.