Get started

Connect a codebase, set your rules, install one harness, and let every coding agent work from the same process.

Teams building with AI agents have run into the same wall: the code is moving fast, but rules, decisions, skills, and project memory are scattered across tools. Every new agent session starts from scratch.

gjalla gives you one harness for all of it. Humans manage it in the portal; Claude Code, Codex, Cursor, Gemini, and other agents inherit it through the CLI and MCP. The steps below take about five minutes.

Add a codebase

In the web app, open Master specs in the side nav and hit the + to connect the repository you want to track. gjalla starts a background analysis to discover your architecture, capabilities, rules, and tech stack. It takes a few minutes.

Connecting GitHub here is what enables two key features of gjalla:

reviews on your GitHub pull requests
the generation of your full master spec

Set your rules

Open Rules in the side nav and add the things you expect to be true in your code. Teams capture rules like “customer data can never be sent to external APIs” here. Anything you find yourself describing to new teammates is a good candidate.

Write rules in plain language. gjalla turns them into real gate behavior for reviews and agent workflows, not just advisory text.

Install the CLI

pipx install gjalla
cd your-repo
gjalla setup

The wizard installs everything your agent needs in one pass: it links the repo to your project, installs hooks, configures MCP where available, and writes each agent’s native guidance and skill files. Install once; every supported agent in the repo gets the same harness.

Reviews on every PR

With GitHub connected, gjalla runs an independent production-readiness review on each pull request. It’s grounded in your master spec, verifies your rules, and checks against industry best practices.

The result is a review that tells you what actually changed about your system and whether anything needs your attention.

Results are available directly in GitHub, and are surfaced in your gjalla dashboard.

If you’d prefer daily, weekly, or on-demand reviews instead of on PR, you can adjust the cadence from Automations.

One review uses one credit. Starter includes 5 credits to get started; Startup includes 100 a month, with top-ups whenever you need them.

Orient the agent

Open your coding agent in the repo and ask it something deliberately architectural, a question that requires it to know the system before answering:

What are our core architecture elements, and which rules apply when changing the auth flow?

A grounded agent will call gjalla state show and gjalla rules show (or the MCP equivalents) before answering. If it just guesses from filenames, re-run gjalla setup and restart the agent session.

Why teams find this useful

One of the most common patterns in early sessions: an agent that would’ve invented a new auth pattern instead pulls up the existing decision and proposes work that fits the system. The architecture isn’t in its training data. It’s in your gjalla state.

Use your coding agent like normal

That’s it. From here on, you have:

Steering: agents read the same rules, decisions, and project context before generating code.
Native process: skills and rules install into the machinery each agent already uses, so the harness is automatic.
Visibility: reviews show what changed, what rules were checked, and what needs attention.

When the harness changes, sync the repo:

gjalla sync

Every supported agent picks up the update on its next session.

On the Startup tier, you also get:

A living master spec, shared rules across your team, and cross-repo context. The two steps below cover the extra setup.

Your living master spec (Startup)

When you connect GitHub on Startup, gjalla builds a full, always current map of your system: architecture, data flows, dependencies, and more. Review it from the codebase’s Master Spec tab. Your agents build against the whole picture, not just your rules, and it stays current as you ship.

Bring your team in (Startup)

Invite the people whose agents touch your codebase under Team Settings. One consistent set of rules, context, and steering is then available to every engineer and every agent. Group multiple repos into a System for cross-repo context so breaking changes get caught before they merge.

If something didn’t work

Hook didn’t run on commit: run gjalla setup hooks, or check whether you’re using husky/simple-git-hooks (see the CLI guide)
Agent ignored the guidance file: confirm it’s in the repo root and your agent is configured to read it
Empty Overview after analysis: Troubleshooting covers the common causes
Anything else: email support@gjalla.io with a screenshot. We read every one.