Your codebase has standards. Your agents don't know them yet.
Engraph turns your decisions into boundaries agents follow every session - no drift, no rework, no policing.
Use the team's shared utilities. Don't reinvent helpers that already exist in the codebase.
The problem
Undoing your agent's work still counts as work.
Your agents are fast. But fast in the wrong direction is rework, not progress.
Rules files don't scope. They decay.
CLAUDE.md, .cursorrules, system prompts - they dump everything into one flat file. No subsystem scoping, no lifecycle, no enforcement. The first context dropped under pressure is the one you wrote by hand.
Your agent doesn't know your architecture. It just writes code.
Agents don't know your boundaries, your naming conventions, or why that module is isolated. They make decisions that work in isolation and break in context.
Linters catch syntax. Tests catch bugs. Nothing catches drift.
Nothing checks whether your agent followed the decisions you already made. That gap is where drift compounds - silently, session by session, until you're doing a cleanup sprint nobody budgeted for.
Why Engraph
Standards that stick. From your first project to your whole org.
Engraph works whether you're one person shipping a product or a full engineering org - regardless of how you structure your projects.
Your agent stays in its lane. Every session.
You've made decisions about how your project works - what patterns matter and what not to touch. Engraph turns those into scoped boundaries your agent respects automatically.
- 01.1Define boundaries once, enforced every session
- 01.2Scoped to the parts of your codebase that need them
- 01.3No setup beyond one enable command in your project
One person's fix becomes every agent's guardrail.
When someone on your team defines a boundary - don't touch billing, always use this pattern for auth - every agent on the project inherits it. No meetings. No docs nobody reads.
- 02.1Constraints scoped to subsystems so each team sees relevant context
- 02.2New contributors' agents start with the team's boundaries on day one
- 02.3No manual coordination of what agents should or should not do
Same standards. Every agent. Full audit trail.
Standards that hold across dozens of agents and contributors - scoped to the subsystems that need them, visible in an audit trail, enforced at session start before a single line is written.
- 03.1Full audit trail on every constraint state change
- 03.2Subsystem-scoped governance that does not slow down individual contributors
- 03.3Consistency across every agent operating on the same codebase
How it works
Nothing new to learn. Just your terminal.
Engraph runs alongside your agent. One CLI command to set up, then it stays out of the way - capturing corrections and turning them into constraints as you work.
One correction, permanent memory
Correct an agent once and Engraph captures it as a constraint. No commit, no PR - it's live immediately. The next time any agent works in that part of the codebase, it already knows.
"postgres-only" proposed from correction
"hono-framework", "no-plaintext-secrets", "postgres-only"
localStorage vs "httponly-sessions"
Self-evident from .ts files and strict tsconfig
"validate-public-input" - 4 agents, 3 sessions
Every file is .go, Makefile builds binary
"no-queue-system" - 6 agents, 5 sessions
Express corrected to Hono. "hono-framework" confidence +1
OAuth2 flow guided by team decisions
Every signal is tracked
Corrections, tensions, promotions, and engagements - all visible.
Agent support
Works with Claude Code today via Model Context Protocol. More agent integrations coming soon.
[ok] installed hooks and configured MCP[ok] updated CLAUDE.mdDetected 7 candidate rules in memory files "Stripe calls must go through the billing service" "API responses always use the envelope format" "Feature flags required for user-facing changes" + 4 more rulesMigrate to engraph? (Y/n)[ok] migrated 7 rules as accepted constraintsSubsystem Discovery Scanned 487 files across 3 languages api packages/api/** 86 files dashboard apps/dashboard/** 142 files billing packages/billing/** 34 files auth packages/auth/** 28 files ui packages/ui/** 67 files[ok] created 5 subsystems with map rulesEngraph enabled.Start a new session to activate.acme-saas %
Enable Engraph
One command sets up everything. Installs hooks, configures MCP, migrates existing rules from your memory files, and discovers subsystem boundaries from your code structure.
The platform
Guardrails from day one. Sharper every session.
Engraph reads your project structure and proposes boundaries on first run. As you work, it refines them - so your agents get more disciplined, not less.
Boundaries from your codebase. Automatically.
Engraph scans your project structure, dependencies, and patterns on first run. It proposes an initial set of constraints - the boundaries your agents follow before anyone corrects anything.
- 1.1Maps your project structure, dependencies, and boundaries
- 1.2Detects patterns across different parts of your project
- 1.3Pre-built skills for common frameworks and patterns
- 1.4Proposes constraints for your review
Your corrections become constraints.
When you redirect your agent, Engraph captures the gap between what it did and what you wanted. Recurring redirections surface as candidates - you approve, they become rules. Every agent on the project inherits them.
- 2.1Captures the gap between what the agent did and what you wanted
- 2.2Recurring patterns surface as candidates you approve
- 2.3Constraints apply when agents touch the relevant code
Constraints earn their place. Or get flagged.
When agents follow a constraint and the work lands, it earns confidence. When a constraint goes stale, the system flags it. Your guardrails evolve with your codebase - not against it.
- 3.1Gains confidence when agents follow it and the work lands
- 3.2Flags stale constraints for review
- 3.3Rules start as suggestions and graduate to requirements based on real usage
Pricing
Simple pricing. No per-constraint charges.
Pick the plan that fits when we launch.
Join the waitlist now. When we open access, every plan starts with a free trial — no credit card required.
Pro
Analytics and 3 workspaces for small teams.
- Up to 3 members
- 3 workspaces
- 10K agent sessions / month
- 90-day audit trail
- Analytics dashboard
Team
Consistent agent behavior across your team with org-wide constraints and full audit history.
- Everything in Pro, plus:
- Up to 15 members
- 10 workspaces
- 100K agent sessions / month
- 1-year audit trail
- Org-level constraints
Business
For regulated teams. Compliance templates, audit export, and room to scale.
- Everything in Team, plus:
- Up to 50 members
- 25 workspaces
- 500K agent sessions / month
- 2-year audit trail
- Compliance constraint templates
- Audit log export
Enterprise
For teams over 50. Unlimited everything with custom session volume. EU AI Act impact reporting and dedicated compliance onboarding. Coming soon.
All plans include a free trial at launch. No per-constraint or per-subsystem charges.
FAQ
Common questions.
From the blog
Ideas on agents, constraints, and engineering intelligence
Your context window is not free
MCP servers, CLI tools, skills files, constraints - they all cost tokens. The question isn't how to minimize that cost. It's what your tokens are buying you.
8 min readBut do it better
Your agents aren't broken. They're amnesiac. And you've gotten used to it.
6 min readYour agents don't read the middle
Two papers measured what happens when agents drift from their instructions. Both proposed better enforcement. Neither asked whether the rules were worth enforcing.
10 min readYour codebase has standards. Now your agents will too.
One CLI command. Scoped boundaries. Every session.