Documentation
Overview
Knwlge gives AI coding agents the context a new senior engineer would already have: the conventions, the decisions and the dangerous patterns a team learned the hard way. This page explains the pieces; the guides that follow install them.
What Knwlge is
An agent working in your codebase is fast and context-blind. It does not know why billing_events must never be written directly, that last quarter's migration was rolled back, or which module is quietly deprecated. Knwlge captures that knowledge from the places it already exists — pull requests, incident write-ups, decision records and the code itself — verifies every item against a source at a revision, and serves it to the agent as small, cited context cards the moment they are relevant.
Cards are served, never pasted: nothing is copied into your repository, so nothing goes stale. When no card clears the evidence floor for a task, the agent gets an honest empty brief.
The three parts
Knwlge Global
The control plane at knwlge.com / api.knwlge.com. Accounts, organizations, teams, licences and server enrollment. It knows who you are and which servers belong to whom — and holds no source code, knowledge or memory.
The data plane
Where a project's index, knowledge and memory live. Enterprise customers run an Enterprise Server in their own cloud (their Postgres, storage and embeddings). Everyone else uses the hosted data plane, isolated per organization.
The client
The knwlge CLI and its sidecar on each developer's machine. One sign-in, one command per project: it wires the agent's hooks and MCP server and never adds files you have to commit.
A developer can work for two enterprise customers and on a side project from one laptop with one login. Each project's traffic goes only to the server that owns it — the client picks the server per project, never per machine.
What the agent gets
Five tools, served by the local sidecar the CLI installs into each agent. Most agents call the first one automatically through a hook before they write.
| Tool | What it does |
|---|---|
context.brief | Task-scoped context cards for what the agent is about to do; repeats return only what is new. |
context.search | Search the indexed knowledge directly. |
context.explain | Why each item was served or withheld — the receipt behind a brief. |
memory.write | Record something learned during a session as a reviewed memory candidate. |
memory.feedback | Mark a card useful, incorrect or outdated, which feeds back into what gets served. |
Supported clients
Each client gets context through its richest channel: inject (a hook adds the brief to the model's input), instruct (a managed rules block tells the agent to call context.brief), and tool (the MCP server, always available).
| Client | --clients id | Primary channel |
|---|---|---|
| Claude Code | claude_code | inject |
| Cursor | cursor | instruct (inject best-effort) |
| Codex CLI | codex | inject |
| Codex app (desktop) | codex_app | instruct |
| GitHub Copilot CLI | copilot_cli | inject |
Plans
| Plan | Data lives | For |
|---|---|---|
| Developer Hosted | Knwlge hosted data plane, row-level isolation per organization | Individuals and small teams; nothing to run |
| Enterprise Teams | An Enterprise Server in your cloud | Organizations that keep code and knowledge inside their own perimeter, with SSO and seat licences |
Access is currently by approval: create an account, finish onboarding, and a Knwlge admin lets you in.