Governed knowledge for AI coding agentsYour team's lore,
agent · claude-codegoverned in console
Your team's lore,
at agent velocity.
Lorvel stores your team's conventions, decisions, and gotchas as atomic, governed knowledge — and serves them to every AI coding agent over MCP. Agents write back what they learn. You stay in control.
Works with Claude Code · Codex · any MCP client
tool call
search_knowledge("refund policy")
→ 1 unit · no LLM in the read path
tool call
propose_knowledge({ type: "gotcha", title: "Stripe test mode skips webhook signatures" })
→ active immediately · embedded async
business-ruleactive
Refund window is 90 days
Refunds are accepted within 90 days of purchase — storefront and backoffice alike.
agent · claude-codegoverned in consoleThe problem
Why agents still ship the wrong code
Your agents are fast. They just don't know what your team knows.
The 80% problem
AI code looks right and still misses the invisible 20% — your conventions, your business rules.
Context is the top failure mode
Missing project context is the most-measured cause of agent mistakes.
Hidden memory
What your agent "learned" lives in opaque files no one reviews, owns, or fixes.
Review pays the bill
Missing context resurfaces as longer PR reviews and the same corrections, repeated.
How it works
Capture → Govern → Retrieve
One loop. Feedforward guides what agents do next; feedback captures what they learned.
01
Capture
Agents propose knowledge as they work — one atomic, typed unit at a time. Live instantly, guided by an authoring guide gate.
propose_knowledge(…)02
Govern
Everything is visible: typed units, lifecycle, links, history. Edit, archive, or correct anything — nothing hides in an agent's private memory.
web console03
Retrieve
Any MCP agent pulls the right knowledge at the right moment. Hybrid semantic + keyword search with no LLM in the read path — answers in milliseconds.
search_knowledge(…)The loop closes — what agents learn while executing flows back as proposed knowledge.
Features
Everything below is live today
No roadmap padding — this is what your first agent connects to.
Atomic, typed knowledge
Conventions, decisions, gotchas, API contracts, business rules — one fact per unit, typed and retrievable.
MCP-native
A per-project MCP endpoint with 27 tools agents already know how to call.
Hybrid retrieval
Semantic + full-text search fused into one ranking. Reads are LLM-free and fast.
Trust-by-default capture
Proposals go live instantly — an authoring guide gate instead of approval queues.
Agent-first tasks
Tasks and subtasks with append-only progress logs and human-readable refs like LO-42.
Knowledge graph
Inline links between knowledge and tasks, with automatic backlinks.
Keys & OAuth 2.1
Scoped read / read-write API keys, plus OAuth 2.1 for MCP clients.
Share links
Agents publish markdown to expiring public links (up to 30 days).
Web console
Browse, govern, and connect agents from a clean web UI.
Project isolation
Every project fully isolated — keys and endpoints are scoped to one project.
all 10 shipping — nothing here is "coming soon"
Tasks
Work agents can pick up, progress humans can read
Tasks are first-class in Lorvel — written, searched, and closed by agents over MCP, governed in the same console as your knowledge.
Spec stays stable, progress appends. The body is the spec; every update lands in an append-only progress log — one log_progress call, no ceremony.
Subtasks with rollup. One level deep, by design — parents report “3/5 completed”, never auto-close.
Agents find open work. search_tasks is hybrid search over open work by default — closed history only when asked.
Linked to knowledge. Task bodies cite the KUs they depend on; statuses are yours to define — up to 7, agent-legible.
LO-42in progresslast activity 4m ago
Migrate webhooks to Stripe API v2
subtasks 3/5 completed
progress log · append-only
09:12agent · claude-code — picked up; reading KU-12 Refund window is 90 days before touching handlers.
10:47agent · claude-code — signature verification ported; test mode gotcha applied. → in_progress
11:03minh · console — hold rollout until finance signs off.
log_progress(…) · spec body untouched
In your workflow
From spec to merged, context flows both ways
Spec
Read before planning
The agent pulls your constraints and prior decisions before it writes a plan.
Execute
Context while coding
Knowledge is pushed and pulled at the moment the code is written.
Reconcile
Learning flows back
What the agent figured out returns to the knowledge base — governed, not hidden.
Governed vs naive RAG
The same call. Very different answers.
both systems askedsearch_knowledge("refund policy")
Naive RAG
3 chunks · conflicting…customers may request a refund within 30 days of purchase, except where…
conflictswiki/billing-faq.md · last touched 2023
refund window — 30 or 90 days? TBD, check with finance before shipping
conflictsnotes/2024-06-planning.md
## Getting started — clone the repo and run npm install to…
irrelevantpayments/README.md · similarity 0.41
Agent picks one and hopes.
Lorvel
1 unit · cited · millisecondsbusiness-ruleactive
Refund window is 90 days
Refunds are accepted within 90 days of purchase — storefront and backoffice alike.
Agent cites KU-12 and codes it correctly.
| Naive RAG | Lorvel | |
|---|---|---|
| Unit | 512-token chunks, cut mid-sentence | Atomic knowledge units |
| Capture | Crawl & chunk whatever exists | Agents propose; humans govern |
| Freshness | Re-crawl and hope | Lifecycle: active → archived — stale knowledge retires |
| Provenance | Lost at chunk boundaries | Every answer cites its unit |
| Search | Vector-only | Hybrid vector + keyword, fused ranking |
| Trust | Whatever was crawled | Typed, governed, auditable |
Naive RAG drifts. Lorvel knows what's still true.
Integrations
Works with Claude Code, Codex, and any MCP client
Paste your project's MCP URL. Done.
Claude CodeOpenAI CodexMCP
Give your agents a memory you can trust.
Set up in minutes. Connect your first agent today.
Get started