Skip to content
Guardi
NewHand GitHub & GitLab issues straight to an agent

Your AI engineering team.
Governed end to end.

Guardi turns the coding agents you already use into a team that plans, builds, reviews and ships through your workflow — and stops for you at every gate that matters. Quality gates, budgets and security scans included.

Get startedSee how it works

Works withClaude Code·Codex·OpenCode·Gemini CLI·Kimi·Zcode

Guardi Workflows page: nine pipelines, two running, one waiting for approval, with a live activity panel

Runs the agents you already use

  • Claude Code
  • Codex
  • OpenCode
  • Gemini CLI
  • Kimi
  • Zcode

Works the backlog you already have

  • GitHub
  • GitLab
  • Self-hosted instances

Reaches you where you already are

  • Slack
  • Telegram
  • Webhooks
  • REST API

Why Guardi

Agents that write code are easy. Agents you can trust with your codebase are not.

A single agent in a terminal can write a function. Shipping features takes a plan, a review, a budget, and someone accountable for what merges. Guardi gives your agents that structure — so you get the output of a team without losing the judgement of one.

Ship more, without hiring for it

A team of specialised agents takes a one-line request to a reviewed pull request — PRD, tech plan, implementation, review and deploy — while your engineers work on what only they can do.

Stay in control of every change

Agents stop at the stages you mark. You approve, send back with a note, or abort — from the web or your phone — and every decision is recorded with who made it and why.

Code that gets better, not worse

Quality gates check tests, the spec and mutation score before a stage can pass. The quality report names what broke a limit — and which agents and models write the code that lasts.

Know what every run costs

Cost per run, per stage and per model, against a budget you set. When a run goes over, Guardi says by how much, which stage crossed the line, and why.

How it works

From a sentence to a reviewed pull request.

A workflow is your SDLC written down once: which agent does each stage, and where a person has to say yes. Run it as many times as you like.

01

Describe it, or hand over an issue

Type a sentence, or pick an issue from GitHub or GitLab. Guardi suggests the workflow that fits and attaches the issue as context.

02

Agents plan, build and review

Each stage runs on the agent best suited to it, inside an isolated session on your fleet, with project memory carried from stage to stage.

03

You decide at the gates

Gated stages wait in your Inbox with the brief, the diff, the findings and the cost. One tap approves; a note sends it back.

04

It ships, and the tracker knows

The branch is pushed and a pull request opened. Status, progress comments and the PR link are written back to the issue it came from.

Add idempotency keys to payment retries

payments-api · Feature delivery · run #214

Runs on

technical-product

Produces

PRD.md — goals, requirements R1–R6, acceptance criteria

Then it waits for you

Approval gate · you sign off on scope

A typical feature pipeline. Click any stage — dots are agent stages, diamonds wait for a person.

Inside the product

Every screen answers the question you opened it with.

Not a wall of tiles. Each page leads with a sentence — where the run stopped, what is over budget, which measure broke a limit — and keeps the detail one click below.

Every pipeline, and what it is doing right now

Each workflow draws its pipeline — dots for agent stages, diamonds for gates — and pulses at the stage a run is on. Live work leads: the approval waiting on you, then what is running.

  • Build from a gallery or import JSON with a live preview
  • Agent, approval, security-scan and HTTP stages
  • Search by name, stage or project
Guardi Workflows page listing pipelines with live stage progress and an activity panel

Human at the gate

Decide in seconds, with everything you need on one card.

A gated stage waits in your Inbox with the brief, the findings, the diff and the cost. The buttons say what will happen — and every decision is kept with who made it and their note.

Approve

The next stage starts on its own. No re-triggering, no context to paste again.

Send back

The stage runs again with your note in the agent's prompt.

Abort

The run stops. It isn't a dead end: retry later from the same stage, with your reason carried in.

NEEDS YOUR DECISIONwaiting 1h · run #214

Add idempotency keys to payment retries

payments-api · Feature delivery · Atlas

  1. Plan
  2. Tech Plan
  3. Implement
  4. Review
  5. Ship

THE BRIEF · WHAT YOU ARE APPROVING

  • Retries carry an Idempotency-Key, stored for 24h.
  • 11 files changed, all tests pass, no downtime.
  • Approving sends it to Ship: push the branch and open a PR.
  • HIGHKey is logged in plain textinternal/retry/handler.go:88
  • MEDNo TTL index on idempotency_keysmigrations/0042_idem.sql

Try it — write a note, then approve, send back or abort.

Guardrails

Autonomy, with limits that hold.

Two things go wrong when agents run unattended: the code quietly gets worse, and the bill quietly gets bigger. Guardi checks both on every run — and tells you in a sentence when something crossed a line.

Quality gates

A stage passes only when its tests pass, the plan's requirements hold, and the tests catch mutants in the changed code. The agent gets an automatic fix; if that makes no progress, the gate asks you.

Implement didn't pass the quality gateattempt 1
  • TEST

    Tests

    1 test fails · TestRetry_ExpiredKeyIsRejected

  • SPEC

    Spec adherence

    R2 contradicted · loop allows 4 attempts, not 3

  • MUT

    Mutation score

    Mutation score 53% · min 60%

Budgets that explain themselves

Set a budget per workflow and watch a live estimate against it. Go over, and the breakdown says by how much, which stage crossed the line, and why.

Over budget by $0.61Feature delivery · example run
$1.61 spent$1.00 budget · 161%

Retries, not the work. 3 retries on Development task burned $0.74 — 46% of the run. Without them the run would have finished inside its budget.

Guardi names one of five causes:

Issues

Your backlog lives in GitHub. Now your agents can work it.

No more copying a title into a prompt and pasting a link back. Connect a tracker once; any issue can be handed to an agent in a sentence, and everything the agent does is written back to where the issue came from.

1

Connect a repository

GitHub or GitLab, cloud or your own instance. Least-privilege token, encrypted, never shown again.

2

Issues arrive in one stream

Mirrored next to the ones written in Guardi, labels in colour, provenance on every row.

3

Hand it to an agent

The likely workflow is suggested from the labels; the issue fills the run's inputs.

4

Progress is written back

Status changes, the agent's progress comments and the PR link land on the original issue.

GitHubGitLabJirasoonLinearsoon
Guardi Issues: GitHub #142 and GitLab !87 side by side, each row showing what the agent is doing
Guardi Sources page: one card per tracker connection with webhook status and write-back toggle

Security scans

Find what matters in a scan, then make it someone's job.

Scans run as a stage in any workflow or on their own. The result opens on what changed and what is causing it — and a click turns findings into tracked issues an agent can fix.

Guardi security scan: 873 findings, 96% from one rule, grouped with the fix stated once

Grouped, not repeated

One rule firing 837 times is one group with one fix — and the files and lines beneath it.

What changed since last time

Each scan compares with the previous one, so a new critical never hides in the total.

Findings become tracked work

Sync a rule's findings to a GitHub or GitLab issue, kept up to date — then hand it to an agent.

Gate the pipeline on it

Put a scan in any workflow. Findings travel with the approval and with the pull request.

Autonomous team

A team of specialists, set up in about three minutes.

Describe what you're building and Guardi suggests a team: named co-agents with a role, skills, standing instructions and the CLI agent they run on. Keep them, rename them, or bring your own — then decide how much they can do without asking.

Guardi setup wizard suggesting a five-agent team from a short brief, each with its reasoning and skills

Team autonomy

Low-risk stages advance on their own; planning and shipping still wait for you.

  1. Product Planningwaits for you
  2. Tech Planwaits for you
  3. Developmentauto
  4. Security Scanauto
  5. Code Reviewwaits for you
  6. Deployauto

Set per project, and change it as trust builds. Quality gates and budgets apply at every level.

Project memory

The tenth run knows what the first one learned.

Every stage of a project shares one memory. The tech plan reads the PRD's decisions; the reviewer knows the patterns this codebase follows. Nobody pastes context between stages — and nothing is lost when a chat window closes.

Episodic log

What every agent read, decided and changed — searchable by task.

Decisions

Architecture and product calls, captured where they were made.

Issues

Bugs and follow-ups an agent surfaced mid-task.

Patterns

How this project tends to work, learned across episodes.

Guardi knowledge graph: the project at the centre, decisions, issues and patterns around it, episodes on the outside

Your fleet

Agents run on your machines. You can see every one.

Guardi orchestrates; the work happens on nodes you control. Bring your own servers and laptops, or spin up a sandbox for a project — and always know whether anything online can actually run the next job.

Guardi Nodes: 3 of 4 online, 8 parallel jobs, which CLI agents the fleet can run, and a warning that no online node can run opencode

Bring your own machines

Install the worker on any Linux or macOS box. It registers, heartbeats and reports which CLI agents it has.

Or spawn a sandbox

No node for a project? Start an isolated sandbox from the project page, sized to the job, reclaimed when idle.

Health you can trust

Heartbeat age drives status — a machine that has gone quiet reads as quiet, not as active.

Isolated sessions

Each job runs in its own session and workspace, so one job's blast radius never reaches another.

Git Ops built in

Branches named from the run title, commits and pull requests opened for you, protected branches refused.

Capacity at a glance

Parallel job slots, agents installed per node, and a warning before a job is queued with nowhere to run.

Guardi Mobilenow on Android

Your team keeps working. You don't have to be at a desk.

The decisions that unblock your agents are small — so make them anywhere. See what is running, approve a stage, or send it back with a note, from your phone.

Download for AndroidiOS coming soon

v1.3.0 · APK · 26.7 MB

Install help & checksum
  1. Open the downloaded file on your Android phone.
  2. If asked, allow your browser to install apps — Android asks once.
  3. Sign in with your Guardi account.

SHA-256

dea78c9dbbb5049fc39054c4c0b980d1726a2881e5e12ff1bec48c86b8322c68

A pulse when you open it

What's running, who's waiting, and the one decision that needs you.

Decide in a tap

The brief, the findings and the cost — approve and ship, or send it back with a note.

Told when it matters

A push when a gate needs you; a Live Activity while a run is going.

FAQ

Questions teams ask before they hand over the keys.

Put your agents to work. Keep the final say.

Start with one workflow and one gate. Loosen the reins as the quality report earns your trust.

  • Connect a node in minutes
  • A suggested team from a one-line brief
  • Your first workflow from the gallery
Get startedBook a demoJoin the community