Every task gets a branch, worktree, terminal, and conversation
Proliferate addresses a problem that appears after a developer runs more than 1 coding agent. Concurrent agents collide in one working directory, share credentials in unclear ways, and leave the human matching conversations to branches and reviews. Proliferate makes the task the unit of isolation. Each task receives a Git worktree, its own branch, terminal, conversation, and review state, so Claude Code can change one area while Codex works elsewhere.
Native harnesses matter here. Proliferate does not reduce every agent to a generic chat adapter; the README lists Claude Code, Codex, OpenCode, Cursor, and Grok as supported agent kinds. MCP integrations, skills, Computer Use, Browser Use, and custom tools can be configured once and shared. Subagents allow a running agent to delegate bounded work, while recurring or event-driven workflows handle jobs such as nightly review and dependency updates.
Parallel work helps only when the merge queue stays understandable
A worktree prevents file collisions, but it does not settle architecture disagreements or make 5 simultaneous patches easy to review. Someone still chooses task boundaries, watches agent output, compares overlapping changes, resolves merge conflicts, and verifies behavior. Proliferate gives that person a common view and durable state. It cannot decide whether parallelism is improving delivery or simply producing a larger queue of plausible code.
For one agent and one branch, the added layer may be needless. Claude Code, Codex, and OpenCode already preserve their own conversations and can work in a manually created worktree. Proliferate becomes more persuasive when several tasks overlap, credentials must be scoped, or a team wants repeatable workflows and review gates. The adoption question is not whether parallel agents look impressive; it is whether branch ownership and verification are currently painful.
What happened when we ran it
Our sandbox installed 1,108 pnpm packages in 50 seconds and consumed 1,128 MB on disk. We measured commit 8f2dcd3 in an unprivileged Debian container with 3 CPUs and 8 GB of RAM. The 61.8 MB checkout contained 8,133 files, roughly 1,380,405 source lines, 28 CI workflows, monorepo workspaces, and a tests directory. It did not have a root Dockerfile.
The root package exposed no generic build script or target to our harness, so the build step was skipped. It also exposed no generic test script or target, so tests were skipped. That does not mean Proliferate has no builds or tests. The repository has scoped commands for its web client, SDK, runtime, mobile app, and release suites, while the Makefile drives the full development stack. It means there is no single root command whose result we can report.
Full control-plane self-hosting needs more than Docker Compose
The production guide puts Caddy in front of a FastAPI server and compiled Vite web app, with Postgres 16 as the bundled database. The installer checks Docker, Compose, disk, ports, and release assets, then writes generated secrets and boots the stack. A public domain must point to the host so Caddy can obtain TLS. This is a real self-hosted control plane, not a thin client redirected to Proliferate Cloud.
Operating it still means owning data and identity. Admin email settings can reinstate access on sign-in, invitations can be domain-gated, and GitHub OAuth is separate from the GitHub App used for repository access. Optional cloud workspaces add E2B and Redis; a central model gateway adds LiteLLM, another database, provider keys, and budget settings. The security policy calls both Proliferate Cloud and self-hosted control planes beta and tells self-hosters to track the latest release.
Anonymous telemetry is enabled by default in the example production environment unless the operator sets the disable flag. The same file contains JWT, cloud, database, OAuth, webhook, agent-provider, and GitHub App secrets. Generated values help the first boot, but backups, secret rotation, restore tests, image pinning, and release updates remain operator duties. Air-gapped support is documented, yet every enabled agent and integration still needs a deliberate source for binaries, models, and credentials.
Credential readiness can disagree with session launch
Open issue 2164 reports an agent shown as ready because one check sees a credential in the host environment, followed by session creation failing with LoginRequired because the workspace-scoped path does not see it. Issue 2107 describes the model picker remaining in a probing state after server restart even though logs show a successful probe. Both reports concern the layer Proliferate exists to simplify: translating several native agent harnesses into one dependable workspace.
Linux ARM64 has another current boundary. Issue 2141 says the agent catalog declared 6 platforms but generated download pins for only 3, leaving Claude and Codex installation unavailable on ARM64 self-hosts. The report gives a server-v0.4.20 reproduction, while the latest release is already server-v0.4.28. Confirm the issue against the exact release you plan to deploy rather than assuming either that it persists or that a newer tag fixed it.
Workflow recovery is still being designed around real failures
An agent workflow can complete useful nodes and still fail later on an unsupported model. Issue 1898 reports a model ID accepted when a workflow definition was saved, then rejected only when that node launched, making the whole run terminal. A related issue says Fail and redo inherits the failing model configuration without offering an override. Those reports make approval and retry semantics worth testing before automating nightly or alert-triggered work.
GitHub showed 428 stars and 82 combined issues and pull requests. The repository and server-v0.4.28 release were both updated on August 27, 2026. That pace supports active maintenance but also means self-hosters must pin releases and read changes before updating. Proliferate is a strong idea with a serious implementation: adopt it when coordinating agents is already a real systems problem, and prove recovery, credentials, and merge discipline before turning on unattended workflows.