central-mcp

The portfolio PM for running many agent-driven projects at once.

All lines run through central. Dispatch across every project, and never lose the thread of any of them.

One MCP server turns any MCP-capable client (Claude Code, Codex CLI, Gemini CLI, opencode, Hermes Agent, OpenClaw, gajae-code) into a control plane for your portfolio of coding-agent projects. Ask in natural language, and the orchestrator routes the request to the right project's agent — non-blocking, with results reported back asynchronously — then briefs you on that project's real state whenever you come back to it.

Why

Agents made it cheap to run four, eight, fifteen projects at once. They did nothing about keeping track of them. Every context switch charges a re-orientation tax — what happened here while I was away? What state is it in? What was I about to do next? — and nobody is playing PM.

central-mcp is that PM:

• Dispatch — send work to any project's agent, in parallel, and keep talking while it runs

• Pulse — return to a project and get "what happened / where it stands / what's next", computed from the repository itself, so work that never went through the hub (direct commits, interactive agent sessions, manual edits) counts too

• Digest — a fixed-format daily/weekly portfolio report, deliverable to chat by a resident agent or a plain crontab

• Observe — live panes on the dispatches you're actively following

• Orchestrate from anywhere — any MCP-capable client can be the front end; never locked to one vendor

Every dispatch is a fresh subprocess in the project's cwd (e.g. claude -p "..." --continue). No long-lived processes, no screen scraping, no tmux dependency on the critical path.

Related MCP server: agent-bridge

How you meet it — three tiers

central-mcp is a layer, not a place. A dedicated orchestrator you have to remember to visit gets forgotten, so the surfaces are ranked by how they reach you:

1. Ambient (the main way in). cmcp install claude once, and every session of the CLI you already use carries dispatch, project_pulse, and the rest alongside its normal tools. Open a project, ask "where does this stand?", and the briefing happens in place.

2. Reach. A resident agentOS bridge — Hermes or OpenClaw — sends the daily digest and failure alerts to Telegram/Discord, the one channel that finds you when no terminal is open.

3. Focus. cmcp tui (experimental), for sessions whose main job is orchestration — fan out, watch it land, supervise.

Design principles

1. Coding agent-agnostic. MCP tools are the canonical surface. Any MCP-capable client can be the orchestrator; any supported coding agent CLI can be the dispatch target.

2. Non-blocking dispatch. dispatch returns a dispatch_id in <100ms. Results arrive asynchronously. The conversation never freezes.

3. Dispatch-router preamble. The orchestrator is instructed to be a pure router — parse the project name, call dispatch, move on. This minimizes LLM reasoning latency to ~1-2 seconds per turn.

4. File-based state. registry.yaml is the single source of truth.

Status

One-line curl installer ships at https://central-mcp.org — bootstraps uv when missing, installs central-mcp, and runs central-mcp init.

Supported platforms

Run central-mcp on the platform it's been exercised on, and expect a bit of roughness elsewhere:

• macOS — primary development and test target.

• Linux — expected to work (pure Python, tmux/zellij are cross-platform), but not regularly tested; please file issues if you hit edges.

• Windows — not officially tested. The core (Python + MCP tools + the tmux / zellij backends where those run on Windows) should work in principle; the cmux backend is macOS-only, so the observation-layer options on Windows collapse to whichever of tmux / zellij you can install. Please file issues if you hit edges.

Quickstart

# One line — bootstraps uv, installs central-mcp, runs `central-mcp init`.
curl -fsSL https://central-mcp.org/install.sh | sh

(tmux only if you want the optional observation layer.)

# Launch — one command does everything
central-mcp

Manual install if you'd rather not pipe a script:

# 1. Install uv (https://docs.astral.sh/uv/) if you don't have it yet
curl -LsSf https://astral.sh/uv/install.sh | sh

# 2. Install central-mcp + scaffold ~/.central-mcp/
uv tool install central-mcp
central-mcp init

Or with pip: pip install central-mcp

The first central-mcp run auto-creates ~/.central-mcp/registry.yaml and registers central-mcp with every MCP client binary it finds on PATH (claude, codex, gemini, opencode, hermes, openclaw, gjc). After that it launches the orchestrator in your preferred agent.

Manual install if you want fine-grained control:

• central-mcp install all — re-detect + register everywhere

• central-mcp install claude — register with a single client

• central-mcp init — create the registry without launching

Inside the orchestrator session, speak naturally — full example catalog in First session below.

The orchestrator calls dispatch for each request and continues the conversation immediately — you don't wait. Results arrive through three channels:

• Piggyback (automatic): every MCP tool response includes a completed_dispatches array with any results that finished since the last call.

• Background poll (best-effort): a subagent polls check_dispatch every 3 seconds and reports automatically when done.

• User-driven check (100% reliable): ask "any updates?" anytime.

Multiple dispatches run in parallel.

First session — natural-language examples

Everything below is spoken to the orchestrator in plain language. The orchestrator picks the right MCP tool and moves on. You do NOT need to memorize dispatch(...) / add_project(...) / check_dispatch(...) — those are the MCP-layer verbs shown elsewhere in the README for reference, not commands you type.

Set up (once):

• "Add ~/Projects/my-app to the hub. Use claude as its agent."

• "Register ~/Projects/api-server — default agent is fine."

• "What projects do I have?"

Send work:

• "Ask my-app to add a dark mode toggle in settings."

• "Send my-app and api-server the same prompt: tighten up the README."

• "Dispatch to my-app with codex instead of claude this time."

Check progress:

• "What's running right now?"

• "Any updates?" / "status?"

• "How did that last dispatch to my-app go?"

• "Show the last 3 dispatches for my-app."

• "Overall status across all projects?" (triggers a portfolio-wide summary)

Recover / switch threads:

• "Cancel the my-app dispatch — the prompt was wrong."

• "What conversation sessions do I have for api-server?"

• "Switch api-server to session abc123 for the next dispatch."

• "Back to the default / latest session for api-server."

Language preferences:

• "For my-app, answer me in Korean from now on."

• "Use French for api-server unless I say otherwise."

• "Just this one dispatch in Japanese for my-app."

• "Clear the saved language for my-app and go back to English."

Shape the fleet:

• "Put my-app and api-server at the top of the list."

• "Remove the old legacy-tool project."

Observation tip for first-time users. Starting out, it's worth running central-mcp tmux (or zellij — or cmcp inside cmux.app on macOS) in a second terminal so you can see the per-project dispatch streams live while you chat with the orchestrator. It builds intuition for how fast dispatches really are and what kinds of prompts produce useful output. Once the orchestrator's summaries match what you'd have checked in the panes anyway, drop observation and work from the orchestrator alone — see Optional observation layer below for the full story.

MCP tools

central-mcp exposes 18 tools under the server name central:

How dispatch works

dispatch("my-app", "add error handling to auth")
  → subprocess.Popen(["claude", "-p", "...", "--continue"], cwd="~/Projects/my-app")
  → returns {dispatch_id: "a1b2c3d4"} in <100ms
  → background thread captures stdout when process exits
  → check_dispatch("a1b2c3d4") → {status: "complete", output: "...", duration_sec: 45}

Supported agents

Agent names are validated at registration time — typos like cursor-agent are caught immediately, not at dispatch time.

Switching agents mid-project

You can change a project's registered agent any time — useful when a given codebase turns out to pair better with a different CLI:

update_project(name="my-app", agent="codex")

update_project also accepts description, tags, permission_mode, and fallback — omitted fields stay untouched. Switching to codex auto-adds the project dir to ~/.codex/config.toml trust list.

One-shot agent override

Sometimes you want to route one task to a different agent without mutating the registry — e.g. a design-heavy task goes to a design-strong agent while the project stays on its usual one:

dispatch(name="my-app", prompt="...", agent="codex")

The registry entry is untouched. Next dispatch without agent= goes back to the project's saved agent.

Fallback chain on failure

If the primary agent exits non-zero (rate limit, token cap, crash), central-mcp can transparently retry with a backup:

# per-dispatch (not persisted):
dispatch(name="my-app", prompt="...", fallback=["codex", "gemini"])

# save a default for this project:
update_project(name="my-app", fallback=["codex", "gemini"])

The result reports which agent actually produced output (agent_used), whether a fallback was triggered (fallback_used), and the full list of attempts. Timeouts are not retried — the user should see them directly rather than burn the whole chain on a stuck agent.

Pass fallback=[] to explicitly disable the saved chain for a one-shot dispatch.

Permission modes

Most coding agents ask "is this OK?" before editing files, running commands, or installing packages. That's fine when a human is at the terminal — but anywhere central-mcp runs, there's no TTY to answer approval prompts, so the work can hang forever waiting for a reply that never comes. Every agent instance central-mcp spawns (orchestrator pane or project-level dispatch) runs in one of three permission modes:

Each vendor brands their permission-skip differently — central-mcp's bypass/auto are unified names that map to the right vendor flag per agent:

If a vendor adds an equivalent to claude's auto mode later (codex sandbox-warn, gemini review-mode, etc), central-mcp will wire it into this same auto alias — existing config keeps working.

Modes apply at two separate layers:

1. Orchestrator layer — central-mcp run / central-mcp tmux / central-mcp up / central-mcp zellij

This is the agent you talk to — the orchestrator pane that calls MCP tools. Default: bypass. Change it with --permission-mode:

central-mcp tmux   --permission-mode auto        # claude-only, classifier-reviewed
central-mcp run    --permission-mode restricted  # no auto-approval, prompts will halt
central-mcp zellij --permission-mode bypass      # explicit default

With orchestrator bypass, the orchestrator can freely read/write files inside ~/.central-mcp without asking — so CLAUDE.md, scratch notes, and hub-level edits happen without friction. With auto (claude + Sonnet/Opus 4.6 only), a background classifier vets each action instead of a blanket skip. auto is ignored (no flags emitted) for non-claude orchestrators. The orchestrator mode does not propagate to dispatched project agents; those carry their own per-project value.

2. Per-project dispatch layer — dispatch(..., permission_mode=...) / registry.yaml

This controls the agent spawned inside a specific project's cwd for one dispatch. The value is saved to registry.yaml on first dispatch (default: "bypass") and reused for every subsequent dispatch to that project. Flip it anytime:

dispatch(name="my-app", prompt="…", permission_mode="bypass")      # auto-approve, save
dispatch(name="my-app", prompt="…", permission_mode="auto")        # claude-only, classifier
dispatch(name="my-app", prompt="…", permission_mode="restricted")  # no skip, no classifier
update_project(name="my-app", permission_mode="auto")              # flip without dispatching

"auto" is rejected with an explicit error if the project's agent chain includes anything other than claude — central-mcp never silently downgrades auto to bypass for a fallback. With "restricted", read-only dispatches still work (answering questions, reading files, explaining code); anything that would prompt (editing, shell, deps) times out — retry with bypass/auto, or open a regular terminal in the project's cwd for interactive approval.

⚠️ bypass is powerful — and at your own risk

In bypass mode (at either layer), the agent may edit files, run shell commands, install packages, call network services, and push code without confirming with you first. That is what makes non-stop orchestration possible, but it also means a misguided prompt, prompt injection from a malicious source, or an agent hallucination can cause real damage — dropped tables, force-pushed branches, deleted files, leaked credentials, unintended API spend, etc.

auto mode is a middle ground — still headless, but a classifier blocks a standard set of destructive patterns (see the Claude Code permission-modes docs for the default policy). It reduces prompt-injection risk but does not eliminate it. restricted is safest but only useful for agents that don't need to write.

Typical reasoning:

• Orchestrator mode controls what the hub-level agent can do in ~/.central-mcp and when calling MCP tools. Lower risk in practice because the hub dir has no production code, but still read/write.

• Per-project mode controls what each project-level agent can do inside that project's cwd. This is the higher-risk layer — it can rewrite your source, run your build, push branches.

Switch away from bypass (to auto for claude, or restricted) if any of these apply:

• The project (or ~/.central-mcp) holds sensitive code, secrets, or production data you can't lose.

• No safety-net commit/push is in place.

• You didn't read the prompt carefully or you're delegating work from untrusted sources.

• You want to review every command the agent is about to run.

Disclaimer: central-mcp is a routing layer and does not supervise what the agents do. You are responsible for the scope, targets, and consequences of every dispatch you run in bypass (or auto) mode at either layer. The authors and contributors of central-mcp are not liable for any damage, data loss, security breach, cost, or other harm that results from the selected mode. Use snapshots (git commits, backups, branch protection), least-privilege credentials, and offline/sandboxed environments where possible.

If a project deals with sensitive code and you're not comfortable granting blanket bypass, switch to auto (claude + Sonnet/Opus 4.6) or keep restricted and stick to read-only dispatches.

Session handling (conversation continuity)

By default, every dispatch resumes the agent's most-recently-modified conversation in the project's cwd — claude --continue, codex exec resume --last, gemini --resume latest, opencode --continue. Droid is the exception: its headless exec has no "resume latest", so a droid dispatch without an explicit session id always starts a fresh thread.

When the user wants to switch to a specific session (or recover from ambient drift — e.g., an interactive session in the same cwd made the "latest" move), dispatch(session_id=...) is a one-shot override:

list_project_sessions("my-app")
  → [{id: "a1b2...", title: "auth refactor", preview: "Investigate login timeout after refresh", modified: "..."}, ...]

dispatch("my-app", "continue from there", session_id="a1b2...")
  # → claude -p "..." -r a1b2...

After that dispatch, the resumed session is now the most-recently-modified — so the next default dispatch("my-app", "...") picks it up via --continue automatically. You only restate the id when you want to switch threads.

preview is intentionally bounded and best-effort, not a transcript dump. For filesystem-backed agents it usually comes from an early recognizable message in the session file; for CLI-only backends it may fall back to the backend's own title-like output.

For drift-proof behavior (or for droid, which needs a persistent pin to keep continuity), pin the session:

update_project("my-app", session_id="a1b2...")
# All future dispatches carry -r/-s <id> regardless of ambient state.

update_project("my-app", session_id="")
# Empty string clears the pin.

Resolution precedence when dispatching: explicit session_id arg > project's saved session_id > agent's resume-latest flag.

Preferred response language (per project)

Dispatch defaults stay unchanged: if no language is pinned or overridden, agents receive the original prompt and answer in their own default language (English for the agents central-mcp targets today).

When a project needs a different language consistently, save it in the registry:

update_project("my-app", language="Korean")

Every future default dispatch to my-app now gets a preface like:

Respond to the user in Korean.

<original prompt>

You can also set it at registration time:

add_project("my-app", "~/Projects/my-app", language="Korean")

Per-dispatch behavior:

dispatch("my-app", "summarize the current status")                     # uses saved project language
dispatch("my-app", "summarize the current status", language="Japanese")  # one-shot override
dispatch("my-app", "summarize the current status", language="")          # suppress saved language once
update_project("my-app", language="")                                    # clear saved language pin

The saved preference lives in registry.yaml as project metadata, so it is practical, explicit, and backward-compatible with older registries that simply omit the field.

Dispatch history (per project)

Every dispatch streams its start / output / complete events into ~/.central-mcp/logs/<project>/dispatch.jsonl (append-only). dispatch_history reads the terminal events back, merged with their matching start:

dispatch_history(name="my-app")          # last 10 dispatches for my-app
dispatch_history(name="my-app", n=50)    # last 50

For a cross-project view, use orchestration_history (below).

Orchestration history (portfolio view)

Asks "how is everything going?" in one shot. Reads the global timeline at ~/.central-mcp/timeline.jsonl (also append-only) plus the server's in-memory in-flight table:

orchestration_history()                  # in-flight + last 20 milestones across all projects
orchestration_history(n=100)             # wider slice of history
orchestration_history(window_minutes=60) # only count activity in the last hour

The response bundles: in_flight (running now), recent (newest milestones), per_project (dispatched/succeeded/failed/cancelled counts, last timestamp), and a registry snapshot. The orchestrator uses this to write a multi-project summary in one pass.

Performance / cost tip: lighter model for the orchestrator

The orchestrator's job is routing — it doesn't need top-tier reasoning. With Claude Opus 4.7 turns already land in ~2-3 seconds for routing, so latency is not a strong reason to switch. The stronger reason is tokens: every turn the orchestrator takes is billed against its model, and a routing turn with a lighter model is meaningfully cheaper. Optional tunings:

The sub-agent model is independent — each dispatch spawns its own process with whatever model the project's agent defaults to, so lightening the orchestrator does not lighten the sub-agents.

CLI reference

central-mcp                        # no-arg → launch orchestrator (same as `run`)
central-mcp run [--agent X] [--pick] [--permission-mode {bypass,auto,restricted}]
                                   # launch orchestrator (default: bypass; auto is claude-only)
central-mcp serve                  # run MCP server on stdio (used by MCP clients)
central-mcp install CLIENT         # register with claude | codex | gemini | opencode
                                   #   | hermes | openclaw | gjc | all
                                   # hermes/openclaw also get the orchestration skill
central-mcp alias [NAME]           # short-name symlink (default: cmcp)
central-mcp unalias [NAME]
central-mcp init [PATH]            # scaffold registry.yaml (default: ~/.central-mcp)
central-mcp add NAME PATH [--agent claude|codex|gemini|droid|opencode|hermes|openclaw|gjc]
central-mcp remove NAME
central-mcp reorder NAME [NAME ...]  # reorder projects — unlisted ones keep relative order
central-mcp list                   # one-line registry dump
central-mcp brief                  # orchestrator-ready markdown snapshot (registry only, ~50ms)
central-mcp pulse [NAME] [--commits N] [--history N] [--pr|--no-pr] [--json]
                                   # what actually happened in a project — git + dispatches
                                   # + sessions + open PRs. No NAME → sweep the workspace.
central-mcp digest [--workspace NAME] [--hours N] [--quiet-days N] [--no-quota] [--json]
                                   # the portfolio report; `cmcp digest | <notifier>` from a
                                   # crontab is a complete push-reporting setup
central-mcp workspace list         # list workspaces with project counts
central-mcp workspace current      # print the active workspace
central-mcp workspace new NAME     # create a new workspace
central-mcp workspace use NAME     # switch the active workspace
central-mcp workspace add PROJECT --workspace NAME
central-mcp workspace remove PROJECT --workspace NAME
central-mcp up [--no-orchestrator] [--permission-mode {bypass,auto,restricted}] [--max-panes N]
               [--projects A,B,C] [--all-projects]
                                   # optional tmux observation layer (active workspace).
                                   # Panes go to the most recently active projects that fit
                                   # one window; --projects picks, --all-projects tiles all.
central-mcp tmux [same flags as up] [--workspace NAME | --all]
                                   # create session if missing, then attach via tmux
central-mcp zellij [same flags as up] [--workspace NAME | --all]
                                   # same, but via zellij (generates a KDL layout)
central-mcp tmux switch NAME       # attach to cmcp-<NAME> session (create if missing)
central-mcp zellij switch NAME     # same, via zellij
central-mcp down                   # kill all cmcp-* observation sessions
central-mcp watch NAME [--from-start]
                                   # stream one project's dispatch events
central-mcp upgrade [--check]      # self-update from PyPI (uv → pip fallback)

Optional observation layer

Why it's optional

• A pane is a microscope, not a map. It earns its screen by showing one project's raw output closely. The portfolio question — where does everything stand? — belongs to cmcp pulse, cmcp digest, and the orchestrator's briefing, which read git as well as the dispatch log.

• Orchestrator is the primary surface. dispatch / check_dispatch / project_pulse return structured summaries; the orchestrator turns those into natural-language status — no scrolling stdout required.

• Work should be possible from anywhere. central-mcp is designed so a phone/tablet over SSH is enough to keep moving. The hub can't require a multi-pane desktop to function.

• Turn observation on only when the live view actually helps — debugging a stuck agent, tailing a long migration, or screen-sharing the fleet. For normal operation it adds noise, not signal.

Suggested onboarding — start observed, graduate to orchestrator-only

On your first few sessions you almost certainly want the live view. Watching each project's dispatch events stream past builds a felt sense of how the orchestrator picks projects, how long dispatches actually take, which prompts produce useful output, and where things tend to stall. Treat observation as a trust-building phase: central-mcp tmux (or zellij, or cmux — see below) gives you panes side-by-side with the orchestrator, so you can eyeball its decisions against the raw agent output in real time.

Once the orchestrator's summaries start matching what you would have checked in the panes anyway, drop the observation layer. At that point you've internalized the pipeline, and working from the orchestrator alone — from any terminal, on any device, including a phone over SSH — is the mode central-mcp was designed for. The observation layer stays one command away (central-mcp tmux / central-mcp zellij / ask the orchestrator in cmux) for the specific moments that still benefit from it.

Backends

Two multiplexer backends are supported as CLI commands:

• tmux — central-mcp tmux (creates the session if missing, then attaches)

• zellij — central-mcp zellij (generates a KDL layout, launches a zellij session or attaches to an existing one)

Both produce the same logical layout (hub tab + overflow tabs, project panes running central-mcp watch <project>). Pick the one you already have installed; you can use both from different terminals as long as they don't share a session name at the same time.

Sessions are named cmcp-<workspace> (e.g. cmcp-default, cmcp-work). Pass --workspace NAME to target a specific workspace, or --all to create sessions for every workspace in one shot. central-mcp tmux switch NAME / central-mcp zellij switch NAME attaches to cmcp-<NAME>, creating it if missing.

A third option — cmux on macOS — doesn't have its own CLI command: you run cmcp inside cmux.app yourself and ask the orchestrator to build the observation panes. See Running inside cmux below.

central-mcp up creates a tmux session central with:

• Pane 0 — orchestrator (Claude Code / Codex / Gemini / opencode), launched in ~/.central-mcp so it picks up the hub's CLAUDE.md / AGENTS.md.

• Panes 1…N — one per registered project, each streaming that project's dispatch activity live via central-mcp watch <project>. Every dispatch's prompt, output, exit code, and duration scrolls past in real time.

Windows are named cmcp-<N> with the first window picking up a -hub suffix (cmcp-1-hub) when it holds the orchestrator — so you can tell at a glance which window to jump to. Cycle panes with Ctrl+b n / Ctrl+b <digit>. When the registry has more projects than fit in one window, extra windows (cmcp-2, cmcp-3, …) are added automatically. --max-panes N sets a per-window cap; without it, central-mcp reads the current terminal's size and picks how many panes fit above the readability floor (~70 cols × 15 rows per pane — tuned so a 13–15" laptop full-screen lands on 2 column slices).

Orchestrator layout: the first window puts the orchestrator pane in a full-height left column sized to match one project column. So orch + 1 project reproduces a 50/50 split, orch + 3 projects yields four equal columns (orch + 3 projects in a single row), and orch + 9 projects gives orch a 1/6 column with 2 × 5 project grid on the right.

central-mcp tmux                         # active workspace → create if missing, then attach
central-mcp tmux --workspace work        # target the "work" workspace session
central-mcp tmux --all                   # create/attach sessions for every workspace
central-mcp tmux switch work             # jump to cmcp-work (create if missing)
central-mcp tmux --permission-mode auto        # claude-only; classifier-reviewed orchestrator
central-mcp tmux --permission-mode restricted  # orchestrator surfaces approval prompts
central-mcp tmux --no-orchestrator       # watch panes only (no orchestrator)
central-mcp tmux --max-panes 6
central-mcp up                           # create the session but don't attach (scripted flows)
central-mcp down                         # tear down all cmcp-* sessions

The hub window (cmcp-1-hub) uses tmux's main-vertical layout: the orchestrator pane sits on the left taking two cells' worth of space, and project panes stack on the right. So the hub holds panes_per_window − 1 panes (default 3 — orchestrator + 2 projects), and overflow windows get the full panes_per_window projects each. Every pane carries its role name on its top border, and the orchestrator border is highlighted in bold yellow so you can spot it at a glance.

Kill with central-mcp down — the MCP dispatch path never depends on this layer, so tearing it down doesn't affect in-flight dispatches. The watch command is a read-only tail of ~/.central-mcp/logs/<project>/dispatch.jsonl; you can also run it standalone in any terminal.

" to run, to exit" in a watch pane

If a zellij watch pane shows <ENTER> to run, <Ctrl-c> to exit instead of streaming dispatch events, the underlying central-mcp watch <project> child has died or never started. This is zellij's built-in safety net — it holds the pane open (preserving scrollback) and waits for explicit user action instead of respawning or dropping to a shell. Don't press ENTER: the pane is disconnected from its original command at this point, so a manual re-run here won't pipe back into central-mcp. Instead, rebuild the session: cmcp zellij (automatic teardown + rebuild — one command). Every pane respawns with a fresh watch child.

Upgrading while an observation session is attached

Only matters if you use the observation layer. Every cmcp tmux / cmcp zellij invocation unconditionally tears down the prior observation session (if any) and rebuilds at the current terminal's size before attaching, so you always end up with fresh panes carrying the newly-installed binary. central-mcp upgrade additionally tears down the observation session before replacing the binary, so upgrading while already attached is covered too.

Trade-off: if two terminals are simultaneously attached to the same session and one runs cmcp tmux, the other disconnects. In exchange, "stale session vs new binary" isn't something you ever have to think about.

Running inside cmux

macOS only. cmux.app is a native GUI terminal designed so that agents manage their own panes. The workflow:

1. Launch cmux.app.

2. In a cmux pane, run cmcp — the orchestrator (claude / codex / gemini) starts inside cmux and inherits CMUX_WORKSPACE_ID.

3. Ask the orchestrator to turn on observation mode, e.g. "turn on observation mode".

The orchestrator reads ~/.central-mcp/AGENTS.md on launch — which includes a terminal-size-aware recipe for this workflow — and uses its Bash tool to chain cmux new-split, cmux send, and cmux send-key per project. That recipe deliberately snaps each workspace to halving-safe balanced grids (2×2, 2×4, 4×4, etc.) rather than asking cmux to fake clean thirds from repeated 50/50 splits. Workspace naming mirrors tmux / zellij's window convention: the orchestrator's own workspace is renamed to cmcp-hub, and observation panes go into dedicated workspaces named cmcp-watch-1, cmcp-watch-2, … (one per terminal-size-derived grid chunk). cmux lets you tab between them from its sidebar.

No central-mcp cmux subcommand exists: central-mcp itself stays out of the cmux socket, the agent does the work. If pane setup fails partway, the orchestrator reports which projects succeeded and you can ask it to retry the missing ones.

Workspaces

Workspaces let you group projects into named sets and switch between them without editing the registry manually.

# Create and populate a workspace
central-mcp workspace new work
central-mcp workspace add api-server --workspace work
central-mcp workspace add frontend --workspace work

# Switch the active workspace
central-mcp workspace use work

# Inspect
central-mcp workspace list     # shows all workspaces with project counts and active marker
central-mcp workspace current  # prints "work"

Projects not assigned to any named workspace fall back to the built-in default workspace. central-mcp workspace list shows that count too.

Session naming: each workspace gets its own multiplexer session — cmcp-default, cmcp-work, etc. The old central session name is kept as a backward-compat alias and is cleaned up by central-mcp down.

Observation with workspaces:

central-mcp tmux                  # active workspace only
central-mcp tmux --workspace work # specific workspace
central-mcp tmux --all            # one session per workspace simultaneously
central-mcp tmux switch work      # jump straight to cmcp-work

Workspace state is stored inside ~/.central-mcp/registry.yaml (current_workspace field) — no separate config file needed.

Registry resolution

Three-level cascade:

1. $CENTRAL_MCP_REGISTRY (explicit override)

2. ./registry.yaml in cwd (per-project override)

3. $HOME/.central-mcp/registry.yaml (global default)

The registry is per-user state — never commit it.

Changing the orchestrator

central-mcp run --pick         # re-run picker, save new choice
central-mcp run --agent codex  # one-off override
$EDITOR ~/.central-mcp/config.toml

Environment variables

• CENTRAL_MCP_HOME — user-state dir (default: ~/.central-mcp)

• CENTRAL_MCP_REGISTRY — registry path override

Development

uv tool install --editable .
uv run --group dev pytest             # 141 unit tests (fast, no real CLIs)
uv run --group dev pytest -m live     # 20 live tests — shell out to real agent binaries
                                      # (claude/codex/gemini/droid); each case skips
                                      # cleanly if that binary isn't on PATH

License

MIT.