PathruleProjectevery plan
How does this project work?
- holds
- Decisions, constraints, procedures
- lives
- Your machine, or Cloud when signed in
- decides
- You and your team
One intelligence.
Every agent.
A persistent intelligence layer for your project, for you, for your team and for every coding agent you use. It keeps what was learned and gives each task only what applies.
works with
Claude Code
Codex
Cursor
Copilot
Windsurf
apps/web/checkout"Add Apple Pay to checkout"
Projectthis codebase
- Payments go through billingClient
- Amounts are integer cents
- Mobile uses Expo Router
Youhow you work
- Small, reviewable diffs
- Ask before touching the schema
- Two options, not five
Teamverified patterns
- New providers ship behind a flag
- Release notes go in CHANGELOG
- Dates are stored in UTC
- Everything
- all 9 notes, every prompt
- What applies
- 4 sent, 5 left out, within budget
Every agent starts over.
- "We decided against an ORM for billing."a closed pull request
- "That migration has to run in reverse."someone's head
- "I told you, use the billing client."your last four sessions
The knowledge is not missing. It sits where a session cannot reach it. Writing it all into one instruction file only moves the problem.
Three scopes. Three jobs.
Youevery plan
How do you like to work?
- holds
- Working preferences
- lives
- Your machine, across repositories
- decides
- You: candidate, active, retired
Teamteam plans
What does the team keep doing?
- holds
- Patterns across the team's work
- lives
- Pathrule Cloud
- decides
- Only what an owner verified
When a personal preference and a project rule disagree, the project rule wins. A preference shapes how the work is done, not what counts as correct.
Scoped to where you work.
workspaceediting services/billing
- /applies
- apps/webout of scope
- apps/mobileout of scope
- services/billingapplies
- webhooks/out of scope
- packages/uiout of scope
pathrule3 sent, within budget
- Payments go through billingClientrule
- Refunds are async since the incidentmemory
- add-payment-providerskill
listed by title, ask over MCPWebhook retry policyCurrency rounding
A rule on a folder applies under it, a rule at the root applies everywhere. Within that scope, what applies is ranked for the task, and what did not fit is named so the agent can ask for it.
A person decides what counts.
- draftNew providers ship behind a flagseen across the team's work
- suggestionNew providers ship behind a flagnot trusted, not delivered
- verified by an ownerNew providers ship behind a flagagents start from it
Team patterns start as drafts and count only once a person verifies them. Findings an agent records stay advisory and never become rules on their own.
Some knowledge is not advice.
services/billing/charge.tsClaude Code, example
export async function charge(order) {+ const res = await stripe.charges.create({ amount })}
Edit denied Never call Stripe directly. Use billingClient. strict rule, /services/billing
can block on
- Claude Code
- Codex
A strict rule with an approved block pattern is checked by the Pathrule hook before a matching edit runs. Cursor, Copilot and Windsurf get no Pathrule hook and Remote MCP installs none, so there it arrives as guidance.
Same agent. Same knowledge. A third of the reading.
2,000 items of team knowledge, Claude Code with claude-sonnet-5-5, Python
CLAUDE.md files
- input tokens a chain
- 11.8M
- cost a chain
- $3.89
- memory probes passed
- 22/26
With Pathrule
- input tokens a chain
- 4.3M
- cost a chain
- $1.76
- memory probes passed
- 25/26
PathruleBench: eight sessions of four turns, 26 probes checked by tests and structural checks, three chains an arm. One measured setup, and the gap only opens from 1,000 items. How we measure
Questions.
What is Pathrule Intelligence?
The part of Pathrule that keeps what a project, a developer and a team have learned as memories, rules and skills, scopes it to paths in the code, and decides what reaches a coding agent for a specific task. It is not a coding model and it does not write your code.
Is Pathrule another coding model?
No. Pathrule does not generate your application code. Claude Code, Codex, Cursor, GitHub Copilot and Windsurf remain the execution and reasoning engines, and each keeps its own model provider. Pathrule is the knowledge they start from.
What is the difference between Project, User and Team Intelligence?
Project Intelligence is what one codebase has learned, as path-scoped memories, rules and skills. User Intelligence is how you prefer to work; it is stored on your machine and follows you between repositories. Team Intelligence runs in Pathrule Cloud on team plans, finds patterns across a team's work, and delivers them only after a person verifies them.
Does Pathrule trust what it learns automatically?
No. Findings agents record are advisory and are never promoted to rules automatically. Team patterns go from draft to suggested to verified, and only verified ones reach agents. Rules are written or approved by people.
What happens when my preferences conflict with a project rule?
The project wins where correctness is at stake. A rule is a requirement of the project and is not dropped because a personal preference would be more convenient. User Intelligence shapes how the work is done, not what counts as correct.
Does User Intelligence work across repositories?
Yes. It is scoped to you rather than to a project, so it comes with you when you move to another repository. Project Intelligence is what changes when the repository does.
Can I correct what Pathrule has learned about me?
Yes. Every item User Intelligence holds has a state (candidate, active, retired or rejected), so a preference that was wrong can be rejected and one that stopped being true can be retired. It is stored on your machine, and without an account it stays there.
Does this reduce my token usage?
In the published benchmark it did, once the knowledge base was large enough to matter: at 2,000 items the same agent read 2.7 times fewer input tokens and cost about 55% less a chain than with CLAUDE.md files, because knowledge is scoped to the path, ranked for the task and delivered within a per-prompt budget instead of one file loaded whole on every step. At 500 items the saving is 10 to 19%, and a small knowledge base in a plain file is hard to beat. Your result depends on how much your team has written down. See how we measure.