Alfrig

Security

Can I trust it with my code? Here are the facts.

By default, nothing goes straight to production: fixes land on the branch your engineers choose, after review and CI. And the agent works in isolation, without your repository credentials.

Start free

5 deliveries on us, no card needed.

Simulation with sample data

Control

You decide how much the AI does on its own.

Every step runs on its own or waits for a person. Start by approving every plan and, once you trust it, let everything run through to merge. Try it:

From request to merge with nobody pushing.

  1. 1 · Selection

    The agent picks the projects from the catalog when the task does not say.

    ● automatic

  2. 2 · Investigation

    Reads the repositories and writes the spec: what changes, where and how to verify.

    ● automatic

  3. 3 · Development

    Writes the code, runs the tests and opens one PR per repository.

    ● automatic

  4. 4 · Review

    Another agent reads the diffs, runs the tests and checks every criterion.

    ● automatic

  5. 5 · Merge

    With green CI, squashes in dependency order and deletes the branches.

    ● automatic

Changes requested in review or a red CI go back for fixing on the same branches, up to a round limit (3 by default).

Where to stop

Fully automatic: the agent picks the projects, investigates and follows its own plan, another agent reviews and, once approved with green CI, merges.You choose what stays automatic at each step.

Also

  • QA rejected it? Put the label back and the fix returns to the same branches, from the current code.
  • Ask Alfrig to approve, resume, merge, comment on the ticket or change its status: each action becomes a card that only runs once you confirm.
  • Approve plans and answer questions in the chat with Alfrig, in a Jira, ClickUp or Linear comment, through a link with no login, or from the buttons on a Slack or Teams alert.
  • In Product, approving the proposal and approving the screens can also run on their own or wait for someone.

Merge safeguards

What an agent produces goes through more than one door.

Auto-merge is optional and starts off. Even when it’s on, it waits for a person when something is out of the ordinary.

  • Green CI, always

    Merges only happen with an approved review and green CI. Red CI in any repository of the task holds them all.

  • Pipeline and code owners protected

    An agent’s changes to CI/CD and CODEOWNERS are left out of the push by default; when the connection allows them, the PR waits for a person.

  • Build and data collection under review

    Changes to build and config files (Makefile, Dockerfile, .npmrc and the like) or new data-collection points hold auto-merge.

  • Outside text gives no orders

    What comes from tickets, comments and pages is treated as information, never as instructions. If the agent raises an alert, auto-merge turns off.

  • Tests don’t vanish to pass

    Deleted or skipped tests are flagged in review. If tests fail, it becomes a draft PR, never a merge.

Isolation

One machine per step, without your credentials.

Each step runs on a machine created for it and destroyed when it ends.

  • Disposable machine

    No inbound connections, destroyed at the end of the step; a collector removes any left behind.

  • Tests run offline

    Tests run with the network turned off.

  • Alfrig keeps the token

    Clone, push, PR and merge are done by Alfrig. The agent works without the repository token, which lives only in memory, never in the repository URL or on disk.

Data

Each organization with its own database and its own key.

One organization’s data never shares a database with another’s.

  • Your own database

    Every organization has its own database.

  • Encrypted keys

    Tokens and keys are encrypted with AES-256-GCM, with a key unique to each organization.

  • Signed integrations

    Public API keys are stored only as hashes. Webhooks Alfrig sends are signed, and the ones it receives are verified.

  • Where it runs

    The application runs in São Paulo, Brazil. Agent machines do too, or in the US when capacity there runs out.

People and access

Each person reaches what their role allows.

Invite the people who ask without opening the settings to everyone.

  • Five roles

    Owner, admin, member, requester and guest. Requesters see only their own requests; workflow, models, costs and keys stay with admins and the owner.

  • Connected apps

    An AI assistant connected over MCP reaches only what the authorizing person’s role allows, checked on every call.

  • Sign-in

    With GitHub, Google, Apple or an email link.

AI models

Alfrig doesn’t train models on your code.

You choose whose AI account it is.

  • No training

    Alfrig doesn’t train models. Data use follows the policy of the provider handling the call.

  • With your key

    Anthropic, OpenAI, Google Gemini or z.ai: the model is paid to the provider directly. Calls go through the Alfrig gateway, which uses your key and meters each step, with no markup.

Deletion and privacy

Delete means delete.

The organization owner decides when the data stops existing.

  • Real deletion

    The owner can delete the organization for good: database, files and the encryption key, which leaves any remaining backup unreadable.

  • Short retention

    Technical logs for up to 2 months; backups removed within 30 days after deletion.

  • Your rights

    Access, correction, portability and deletion, through the contact in the privacy policy.

Read the privacy policy

Questions

What engineering and security teams ask

Does Alfrig release to production on its own?+

Not by default. The organization starts with auto-merge off, and fixes go to the branch your engineers choose. If your team turns auto-merge on with the production branch as the target, the fix reaches production with no one stepping in: that’s their call.

Does the agent see our credentials?+

Not the repository ones. Clone, push, PR and merge are done by Alfrig, and the agent works on its own machine, without the token, destroyed at the end of the step.

What access does Alfrig need?+

Only to the repositories you choose: through the Alfrig app on GitHub or a token on other providers. Access is encrypted with your organization’s key.

Does our code train models?+

Alfrig doesn’t train models. Data use depends on the model provider handling the call.

Where is the data?+

In a database that belongs only to your organization. The application runs in São Paulo, Brazil; agent machines do too, or in the US when capacity there runs out.

What do you need to fix today?

What do you need?Write to Alfrig. It asks what’s missing and sees it through.
Start free

5 deliveries on us, no card needed.

Alfrig is made for your computer.

That is where you connect the repository and follow the agents on a bigger screen. Leave your email and we will send you the link to open on your computer.

We use your email to send the link and, now and then, Alfrig news. Unsubscribing takes one click.