Release status

Release scope

BranchRunner Community Edition 1.0.0 is the first public release of the source-available, self-hosted product, operated on your own infrastructure. The release consists of the installer install.sh, the source archive with its .sha256 file, SOURCE_COMMIT, RELEASE.json, SHA256SUMS and the release notes. RELEASE.json and SHA256SUMS identify the exact source commit and the bytes of every file. Cloud, Enterprise services and the VS Code extension are not part of this release.

Supported configurations

Both installers accept only Ubuntu 24.04 and 26.04 LTS on amd64 or arm64 and refuse any other platform, including Ubuntu 22.04 and Debian, before anything is installed. More platforms are planned for later releases.

The server needs at least 8 GB RAM and 2 CPU cores; the server installer refuses a smaller host before it installs anything, because every run reserves 4 GiB and 2 CPUs next to the control plane. 16 GB RAM and 4 CPU cores are recommended. Below 16 GB the installer builds the protected agent images two at a time instead of four, and a host below the recommended size may hold only one run at a time. Allow 60 GB of free disk: the protected agent images need 20 GB of it, plus a build cache capped at 64 GB.

The supported coding agents are Codex (OpenAI) and Claude Code (Anthropic). Each can use an API key or a subscription sign-in. Every supported combination that the table does not list as live-verified, such as a Claude subscription implementer, a worker on a separate host, or an OpenAI-compatible or Azure OpenAI target, is supported but not yet live-verified. A verified row on one Ubuntu version or architecture does not verify another. This page marks a row live-verified once its runs are recorded. Other coding agents and model providers are not part of this release.

What 1.0.0 includes

• One-command installation. The server installer builds every image on your host, sets up HTTPS, the API, the web UI and PostgreSQL, and prints a single-use browser link for owner setup with password and TOTP sign-in. After owner setup it installs a runtime worker on the same host itself. Installation never runs test suites: the image builds bundle the released code, and the host verification suite is opt-in (--verify). The installer prints a timing table for its 12 phases; the release notes publish the measured fresh-installation time and the host size it was measured on, rather than a fixed time target. See the quickstart.
• Single-host and remote workers. A worker can share the server or enroll from another host through a mutual-TLS channel with a single-use token. The worker installs from the same published release archive as the server, with a checksum-pinned one-line command, and needs no Git checkout. An owner or administrator admits every worker explicitly, together with the runtime identity it reports. On the control-plane host, an upgrade that changes only the images built from the release the control plane runs is admitted automatically; any other change, and every worker on a separate host, needs Admit runtime identity. A worker on another host runs Codex subscription and API-key stages; Claude subscription stages run only on the worker on the control-plane host. See runtime workers.
• Protected execution selected automatically. On the control-plane host, the installation selects protected Codex execution once the worker installer has attested the host, and protected Claude execution once it has prepared the Claude runtime. Upgrades keep the selection, and none is the recorded opt-out that automation never overrides. API-key connections need no execution selection. A saved login alone does not authorize a run.
• Credential enforcement on from the start. A fresh installation starts with protected credential enforcement on, so there is no activation step. It cannot be switched off.
• GitHub App setup in the app. Integrations > Connect GitHub creates the installation's own GitHub App on your account. You choose its repositories on GitHub.
• Provider connections. A Codex subscription signs in with a device code. A Claude subscription signs in through the official Claude Code CLI, which creates a token for the connection that is valid for one year. Sign in again renews the token on the same connection, and an installation can hold several Claude accounts, one per connection. Both subscriptions run inside protected containers. API keys work for OpenAI (including OpenAI-compatible endpoints and Azure OpenAI) and for Anthropic. Up to four runs can share one subscription connection. Saved API keys are sealed at rest, and a saved key or a subscription sign-in can be removed from Integrations. Admitted provider hosts, credential enforcement and API-key credential approval are under Settings. See agent providers.
• Connection priority, plan usage and failover. An agent profile orders up to six connections of one agent by drag and drop. Each stage uses the first eligible connection, based on measured plan-limit usage, and moves to the next one when a plan limit, rate limit, quota or login stops it. Codex and Claude subscription cards show the plan windows the provider reported, such as the 5-hour and weekly windows, at sign-in, at Verify and after each run.
• Theoretical and effective cost. Each run and connection shows what its tokens would cost at API list price and what the work actually cost under your subscription fee or API discount.
• Six-stage workflows. Plan → Implement → Validate → Independent review → Approval → Handoff, which opens a draft pull request. See workflows.
• Validation. BranchRunner detects a repository's npm build, typecheck, lint and test scripts and runs them offline in a hardened validation container. See validation and approvals.
• Independent review and signed evidence. A separate reviewer, a subscription or an API-key connection, evaluates the change against the requirements, and the installation signs completion, evidence and review records with its own keys.
• Draft pull requests through the GitHub App. After a person approves the completion review, BranchRunner opens a draft pull request. It has no merge or deploy capability.
• Built-in custody and retention. Signing, handover and retention keys are created on your server; no external Vault is required. Automatic deletion of expired run artifacts that are not on hold is on by default. See data retention, deletion and custody.

Known limitations

• Live verification covers the listed rows only. arm64 is a preview. Workers on a separate host, OpenAI-compatible and Azure targets, concurrent subscription sessions and publication cancellation are supported but not yet live-verified.
• Validation supports npm projects only. It needs a package-lock.json and does not support npm workspaces. Other repositories get only the baseline repository-hygiene check, and completion names the missing evidence.
• Every run needs two providers and a human reviewer. Every change is reviewed as Critical in 1.0.0, so the reviewer must come from a different provider and model family than every Implement step, and a qualified person who did not write the change must complete a human technical review. An installation with connections from only one provider cannot complete review. See review requirements for every run.
• Previews are not available in 1.0.0. BranchRunner does not serve a preview of a run's running application in this release. The run console and workspace detail show previews as unavailable, and no preview link can be opened or shared.
• Failover stays within one agent. A connection list cannot mix Codex and Claude connections. A stage switches only at its start or after a failure that changed nothing in the workspace.
• Cost figures can be upper bounds. When a producer reports no cache detail, all input is priced at the uncached rate and the figure is marked.
• Agent profiles cannot use skills or MCP servers yet. A profile that selects them cannot run. See MCP and tools.
• Claude subscription runs on the control-plane worker only. In 1.0.0 a run that uses a Claude subscription connection is placed only on the worker on the control-plane host; a worker on another host is skipped with claude_remote_execution_unavailable. The Claude token stays in custody on the control-plane host and is never delivered to another host. Workers on other hosts run Codex subscription and API-key stages.
• Protected Codex on another host needs the owner's admission. The worker installer on that host selects remote execution itself. After the worker and its runtime identity are admitted, the installation owner chooses Admit for subscription execution under Runtime workers. See runtime workers.
• No public container images. The installers build every BranchRunner image on your host from the released source and pull only digest-pinned upstream base images and PostgreSQL.
• Backups are yours. Restoring onto a clean host has not been exercised as a complete live drill. Built-in retention reports backup expiry as unavailable. See backup and recovery.

What live verification covers

A configuration is listed as live-verified only after, on an installed host of that operating system and architecture:

1. the installed source, version, image identities, migrations, TLS, owner setup and worker identity were checked; and
2. each run that the row names finished the complete repository or issue → agent → validation and repair → independent review → human approval → draft pull request path on that host, with its evidence retained.

Fixtures and local tests do not count as live verification. A healthy web page, saved login, workflow draft or unit test is not evidence of that path.

Operating responsibility

You own installation, updates, monitoring, credentials, backup and restore, retention and deletion across your storage and backups. This self-hosted release includes no BranchRunner-operated support service, SLA, backup service, retention service or deletion service. Product security and authorization checks still apply.

Report a security issue

Email [email protected], the contact listed in security.txt. Do not put details of a finding in a public issue or pull request. See security.

See quickstart, licensing and runtime workers.