mcp-first multi-agent orchestration for git repos: global task store, shared http mcp endpoint, daemon, admin spa, and tui in one tool.

docs → kasmos.kasthe.co

---

---

features

• global task store — tasks live at ~/.config/kasmos/taskstore.db, shared across all managed repos via the daemon
• shared http mcp endpoint — kas serve (or the kasmosdb systemd/launchd service) hosts a single mcp server at http://127.0.0.1:7434/mcp; all scaffolded harness configs point agents here — no per-agent stdio subprocess needed
• wave-based orchestration — planning → architect → implement → review → readiness review lifecycle with per-wave agent concurrency
• multi-harness support — works with claude, opencode, codex, and other mcp-aware agents
• tui + admin spa + daemon — terminal and browser control surfaces backed by the same service pair
• git worktree isolation — each task runs in its own branch and worktree; merges are handled at review time
• cross-platform services — systemd (linux) and launchd (macos) for always-on operation; see service management docs

---

install

homebrew

brew install kastheco/tap/kasmos

go install

go install github.com/kastheco/kasmos/cmd/kas@latest

install script

curl -fsSL https://raw.githubusercontent.com/kastheco/kasmos/main/install.sh | bash

release asset

prebuilt binaries for macos and linux are on the releases page.

the compiled binary is kas. kasmos remains the project name, module path (github.com/kastheco/kasmos), config namespace (.kasmos/, ~/.config/kasmos/), and release archive prefix — only the installed binary artifact is named kas.

---

quick start

from inside a git repo:

kas serve        # local/dev: start rest api + admin ui + shared http mcp endpoint
kas setup        # scaffold harness configs (.mcp.json, .codex/config.toml, agent prompts) pointing at the shared endpoint
kas              # open the tui

open http://127.0.0.1:7433/admin/ for the admin spa. for always-on operation, run the kasmosdb + kasmos user services instead of foreground kas serve / kas daemon commands.

kas setup (and kas reset) write per-harness mcp config so mcp-aware agents connect to the shared http mcp endpoint (http://127.0.0.1:7434/mcp): claude and opencode use .mcp.json, codex uses a project-local .codex/config.toml with an [mcp_servers.kasmos] block (codex cli reads this natively for trusted projects). the endpoint must be running before starting managed agent sessions. kas check will warn if the shared endpoint is unreachable or if stale kas mcp stdio processes are still running.

note: kas mcp (stdio mode) is still available as a fallback for manual or harness-specific use, but it is not the default transport and should not appear in scaffolded configs.

see the getting started guide for a full walkthrough including harness setup and your first task.

---

mcp server

kasmos exposes its tools via a shared http mcp endpoint hosted by kas serve (or the kasmosdb background service) at http://127.0.0.1:7434/mcp. all scaffolded harness configs (written by kas setup / kas reset) point agents at this single endpoint — no per-agent stdio subprocess is spawned, eliminating the memory overhead of one kas mcp process per agent session.

kas mcp (stdio) is still available for manual or custom harness use, but it is the fallback transport, not the default. managed agent sessions require the shared http endpoint to be running; if it is down, kas check will report the missing endpoint as an error and list any lingering kas mcp stdio processes as a warning.

scaffolded agents are expected to use the shared http mcp endpoint for task, signal, and instance operations. kas task / kas signal are operator and fallback tools, not the default agent path.

tool groups exposed:

group	tools
filesystem	find_files, list_dir, read_file, grep
git	git_status, git_diff, git_log
tasks	task_list, task_show, task_create, task_update_content, task_transition
signals	signal_create
instances	instance_list, instance_pause, instance_resume, instance_send
daemon	daemon_status

filesystem and git tools are sandboxed to the repo root. see the mcp server docs for details.

---

useful commands

kas setup                              # scaffold project config and harness files
kas serve                              # start rest api + admin web ui
kas task list                          # list all tasks
kas task create <name>                 # create a new task
kas task show <task-file>              # show task details
kas task implement <task-file>         # start orchestration for a task
kas task transition <task-file> <event># advance task lifecycle
kas reset                              # refresh scaffold state (with backup)
kas check -v                           # verify scaffold health
kas status                             # show daemon and server status

see the cli reference for the full command surface.

---

links

• documentation: kasmos.kasthe.co
• architecture: docs/architecture/README.md — runtime topology: signals, state, and data-flow diagrams
• contributing: CONTRIBUTING.md
• license: BSL 1.1 — converts to Apache 2.0 on the change date