The author of the One Man Enterprise channel has assembled their daily open-source stack for agentic development from five projects: T3 Code, herdr, NetBird, Lific, and Kandev. Instead of a single monolithic tool, the stack covers four layers — agent management, terminal runtime, networking, and persistent task memory. All components are self-hosted, can be installed with a single command, and work with already purchased Claude Code, Codex, and Cursor subscriptions.

image
image
image

What Happened

The post describes five components of the daily stack. T3 Code is an agent harness by Theo Browne (pingdotgg) with 23,000 stars on GitHub under the MIT license: mobile apps for iOS and Android, the web app app.t3.codes, and an Electron desktop app are available for managing Claude Code, Codex, Cursor, Grok Build, OpenCode, and Antigravity on remote machines. herdr is a Rust-based terminal multiplexer with 39,500 stars under the Apache-2.0 license: agents via CLI and socket API create their own panes and "prompt" each other, and each pane is marked as working, blocked, or idle. NetBird is an open-source alternative to Tailscale with a self-hosted control plane and a WireGuard-based network; the project has 29,400 stars and BSD-3/AGPLv3 licenses. Lific is a Rust-based issue tracker in a single binary of about 25 MB using embedded SQLite: it is MCP-native, contains 30 tools in 6,300 tokens, and the lific connect command sets up 11 AI clients at once. Kandev is a self-hosted AI Kanban in Go under the AGPL-3.0 license, which orchestrates 22+ agents via the ACP protocol with isolation through git worktree and runtimes local, Docker, Kubernetes, and SSH. All projects are open source and can be installed with a single command — via brew, a curl script, or cargo install.

Context

The stack emerged against the backdrop of a shift in development practices: work has moved from single sessions with one model to dozens of parallel agent runs, and a dedicated operational layer has grown for them. Demand for it is visible on GitHub — communities of 39,500 stars for herdr and 23,000 for T3 Code show that orchestrating multiple agent sessions has become a standalone category rather than remaining a feature inside an IDE. Standardization is happening through protocols: MCP has already become an integration point for tools, and ACP, through which Kandev connects two dozen agents, is vying for the role of a common standard for multi-agent environments. Another prerequisite is the well-known problem of long sessions: a work plan dies with the context window, so Lific moves tasks to embedded SQLite outside the context. The decomposition into layers — control panel, runtime, network, memory — reads as an architectural pattern of an "agent OS."

Why This Matters for the Industry

For the industry, this is a platform shift, not a single feature: a stack of layers has formed around the orchestration of agent sessions — management (T3 Code, Kandev), terminal runtime (herdr), networking (NetBird), and task memory (Lific). At the same time, open-source code and zero licensing costs devalue thin wrappers around Claude Code, Codex, and Cursor as a business: what remains defensible is what cannot be copied overnight. The protocol layer here is more important than any individual tool — standardizing interfaces through MCP and ACP determines the reproducibility of multi-agent environments, and consolidation is likely in the coming months: major IDEs and platforms will start natively embedding control panels and orchestration, MCP and ACP integrations will become a hygienic minimum, and the first observability overlays will appear on top of runtimes.

Why This Matters for Users

For readers, this is a ready-made free set for building their own "director's desk" over AI agents. All five projects are self-hosted with MIT, Apache-2.0, BSD-3, and AGPL licenses, and they work with already purchased Claude Code, Codex, and Cursor subscriptions — you only need to pay separately for the models. Remote agent sessions can be managed from a phone via T3 Code apps or the web app app.t3.codes, a stuck agent is immediately visible in the herdr pane status, and task plans do not die with the context window because Lific stores them in its own database. As a result, stuck sessions are diagnosed faster, and a scenario with multiple parallel agents can be assembled without a budget and without vendor lock-in.

What Is Still Unknown / Limitations

GitHub stars are not installations: 39,500 for herdr and 23,000 for T3 Code are a weak proxy metric for engagement, sensitive to media waves and inflation, and without data on actual installations and user retention, the conclusion that the category has been validated by demand is methodologically not closed. From a production standpoint, this is orchestration without observability: the stack lacks latency metrics, token accounting, and budget limits. The claimed token budget savings of Lific (30 tools in 6,300 tokens) is presented without a measurement methodology, and for the quality of inter-agent coordination in herdr, there is no tracing, conflict metrics, or comparisons with a baseline configuration. The ability of ACP to actually hold 22+ agents in a working scenario is also not confirmed by independent measurements.

Sources

Author

Look at AI, editorial team