For startups deploying AI agents

Your agents are fast. Are they governed?

AI SDRs send 60,000 emails. AI support agents resolve tickets overnight. AI marketing coordinators publish content around the clock. And somewhere in there, one of them offered a prospect a feature you haven’t built, quoted a price you haven’t approved, or committed to a timeline engineering hasn’t scoped.

The problem

You deployed AI agents to move faster. Now you’re spending 30-40 hours a week reviewing what they did. Every morning starts with “what did the agents do overnight?” You’re using Zapier, Salesforce, and copy-paste to stitch context between agents. There is no formal authority architecture.

The agents aren’t wrong. They’re ungoverned. They don’t know what they’re not allowed to say because nobody encoded the rules.

30-40 hrs
Weekly agent babysitting
3-5
Unauthorized commitments/month
0
Agents with encoded boundaries

The fix

One npm package. Constraints checked at the moment of action.

Constellation’s MCP server plugs into any agent that supports the Model Context Protocol. Before your agent sends an email, makes a commitment, quotes a price, or publishes content, the constraint engine evaluates the action against your rules and returns a pass, a flag, or an escalation. Setup takes minutes.

# Install

npm install @constellation-governance/mcp-server

# Add to your Claude config

{

"mcpServers": {

"constellation": {

"command": "npx",

"args": ["@constellation-governance/mcp-server"],

"env": {

"CONSTELLATION_API_KEY": "your-key",

"CONSTELLATION_ORG": "your-org-slug"

}

}

}

}

Then encode your rules. These are constraints your team already knows — they just live in people’s heads right now:

PROHIBITION “No commitments to features not in current roadmap”
THRESHOLD “Discounts over 15% require VP Sales approval”
DOMAIN_TOPIC “No content touching: litigation, pricing, competitive claims”
SEQUENCE “Partnership commitments require completed legal review”
AUDIENCE “Revenue projections: board-only”
AUTHORITY “External comms >1,000 recipients require marketing approval”
TIMING “No announcements within 48 hours of board meetings”

7 constraint types. Covers authority, spending, topics, sequencing, audience, timing, and hard prohibitions. Add them through the web interface or the API.

What happens at runtime

01

Agent prepares to act

Your AI SDR is about to send a pricing email. The MCP server’s check tool fires automatically.

02

Constraint engine evaluates

All applicable constraints checked in <200ms. Threshold, authority, topic, audience, sequence, timing, prohibitions. Cached for speed, circuit-breaker protected for resilience.

03

Pass, flag, or escalate

Clean actions proceed silently. Violations surface with context: which constraint, why it applies, who to escalate to. The agent never acts beyond its boundaries.

04

Every action traced

Automatic audit log. When someone asks “what did the agents do?” the answer already exists.

What changes

~70%
Less manual review
0
Unauthorized commitments
100%
Action audit trail
<200ms
Constraint check latency

Your agents don’t get slower. They get governed. The pipeline keeps growing. The unauthorized promises stop. And your team goes from 30 hours of babysitting to 5 hours of governance review — focused on the genuinely hard judgement calls.

What your agents get

8 MCP tools, one npm package

check

Evaluate action against all constraints before acting

boundary

Show all active constraints for a domain

record

Create audit trace after completing an action

escalate

Route violations to the right authority level

preview

Test a new constraint against recent action history

query

Read decisions, commitments, and escalation status

status

Governance health check with pending escalations

vote

Submit member votes on proposed decisions

Pricing

Starter

Test with one agent

Free

5 members, 500 checks/month

Most teams

Team

5-20 agents, real governance

$49/mo

15 members, 5,000 checks/month

Organisation

Multi-team, cross-functional

$199/mo

50 members, unlimited checks

No per-seat pricing. More people governed is better, not more expensive.

Stop babysitting. Start governing.

Your agents are already making decisions. The question is whether those decisions have boundaries. Install the MCP package, encode your rules, and see the difference in your first session.