harbormaster-mcp 26.0.1

Multi-project MCP router + operator dashboard for Claude Code, Codex, Cursor. Ask any local or SSH-reachable project a question without `cd` — the project's subagent loads its own CLAUDE.md and Serena memories and answers. Ships with a live web console: KPI strip, inter-project network graph, trace waterfall, memory editor, multi-axis budgets. Part of the FleetQ ecosystem.

Harbormaster

Stop cd-ing between projects. One MCP server routes any question to any project's subagent — locally or over SSH — with a live operator dashboard for the whole fleet.

Harbormaster is a Model Context Protocol (MCP) server + operator console for developers who work across many projects at once. Point your Claude Code, Claude Desktop, Codex, Cursor, or any MCP client at it, and you can ask any project a question — without cd — and the target project's subagent loads its own CLAUDE.md and Serena memories, answers, and returns a summary.

The bundled web dashboard turns the same server into a glanceable mission control: a project grid with KPIs, an inter-project network graph, a live dispatcher trace waterfall, a markdown memory editor, multi-axis budgets (per-host / per-tool / per-project), and a light/dark theme.

Part of the FleetQ ecosystem. Standalone OSS works fully without FleetQ; FleetQ Bridge integration is purely additive.

---

Why Harbormaster?

You probably already have:

• 5–80 projects in ~/htdocs/ (or ~/work/, or ~/code/)
• Each with its own CLAUDE.md, Serena memories, framework, Git state
• Possibly 1–5 SSH-reachable VPS hosts running production copies
• One LLM-coding session per project (Claude Code, Codex, Cursor) — which means context-switching costs every time you cd somewhere new

Without Harbormaster: You change directory, wait for the agent to re-bootstrap context, ask the question, copy the answer back. Repeat 12 times a day.

With Harbormaster:

> ask_project "accounting-fleetq" "did the cron job run today?"
> fan_out_ask "any flaky tests since last release?" --hosts=local,prod-1
> delegate_task "hr-fleetq" "summarize last 10 commits" --deliverable=changelog

The calling session never leaves its own cwd. The target project's subagent loads its own memory, answers, returns a summary, and your question + answer go into a recall-able Q&A history.

Who is this for?

You are…	And Harbormaster gives you…
A solo developer with 10+ side projects	One console for all of them, with cross-project search and a graph of who calls whom
A small-team CTO	One MCP endpoint your agents talk to across local + production hosts, with budget caps per host/tool/project
A platform engineer	A drop-in Bridge daemon for FleetQ — Platform Tool, A2A Agent Cards, federated KnowledgeGraph
An AI agent researcher	A real-world multi-agent orchestration target with a working web UI for observability
A vibe coder with too many tabs	A keyboard-first dashboard (Cmd-K palette, 1–5 tab nav) and a calmer life

---

What it does

You work across many projects, each with its own CLAUDE.md and Serena memories. Switching cwd loses context. Harbormaster lets one Claude Code (or Codex) session ask any project a question without changing directory — the project's subagent loads its own memory, answers, and returns a summary.

Eight MCP tools cover the day-to-day flow: list projects and hosts, inspect project status, ask a single project, fan out to many, recall prior Q&A, walk the inter-project dependency graph. Optional SSH fan-out targets remote VPS hosts. Optional FleetQ adapter exposes Harbormaster as a first-class Bridge daemon (Platform Tool, A2A Agent Cards, federated KnowledgeGraph).

The companion web UI turns the same MCP server into a local operator console: dashboards, trace waterfall, network graph, memory editor.

---

Quick start

Install with the [ui] extra so you also get the operator dashboard:

uvx --prerelease=allow 'harbormaster-mcp[ui]' --version
# or with pipx:
pipx install --pip-args='--pre' 'harbormaster-mcp[ui]'

Register in Claude Code:

claude mcp add --scope user harbormaster harbormaster-mcp

Or in Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "harbormaster": {
      "command": "/opt/homebrew/bin/harbormaster-mcp",
      "env": {}
    }
  }
}

Run the operator UI alongside (separate process — both read the same TOML config so projects discovered by one are visible to the other):

export HARBORMASTER_UI_TOKEN=$(python -c 'import secrets; print(secrets.token_urlsafe(32))')
harbormaster-ui --port 7531
# then open http://127.0.0.1:7531/ — paste the token at the prompt

Zero-config by default: Harbormaster auto-discovers projects under ~/htdocs/* if it exists. For any other layout, see Configuration.

---

Tools

Twelve MCP tools, all optionally targetable at a remote host via host="<label>".

Tool	Purpose	Cost
list_projects(host=None)	Enumerate configured projects (local) or remote dir listing (SSH).	~50 ms / ~1 s
list_hosts()	Configured [hosts] plus ~/.ssh/config Host aliases.	~5 ms
project_status(name, host=None)	Git log, Serena memory list, log tails.	~200 ms / ~2 s
ask_project(name, question, max_turns=5, host=None)	Spawn claude -p (or configured backend) in project cwd, return ≤ 800-word summary.	~30 s / ~90 s
delegate_task(name, task, deliverable, allow_writes=False, mode="sync", inbox_id="default", host=None)	Delegate a task; allow_writes=true lets the subagent edit files directly, false (default) keeps it read-only. mode="async" enqueues and returns immediately; the caller polls with get_delegated_task or drains the inbox with recall_pending_results.	sync: ~60 s / async: ~5 ms enqueue
get_delegated_task(job_id)	v22.0.0 — status of one async delegated job.	~5 ms
recall_pending_results(inbox_id, mark_read=True, limit=50)	v22.0.0 — drain completed/failed async jobs from one inbox (FIFO on completion).	~10 ms
await_delegated_task(job_id, timeout_seconds=900)	v22.1.0 — block until one specific job lands. Returns immediately if already done; on timeout returns the current row with status="running" or "queued".	up to timeout_seconds
await_inbox(inbox_id, timeout_seconds=900, since=None)	v22.1.0 — block until ANY job in the inbox lands. since skips completions already seen. Caller drains via recall_pending_results after.	up to timeout_seconds
fan_out_ask(question, project_filter=None, host_filter=None, max_concurrency=5, max_turns=3)	Parallel multi-project Q&A; one section per target.	~max_turns × backend_time × ⌈targets / max_concurrency⌉
recall_qa(question, top_k=5, host=None, project=None, min_similarity=0.6)	Semantic recall over prior Q&A answers. Opt-in via [history] enabled = true.	~50 ms (FTS5) / ~150 ms (vec, after warm-up)
project_graph(format="json", include_dev_deps=False, transitive=False)	Cross-project dependency graph from manifest parsing. Returns nodes + edges + optional Mermaid markup.	~100 ms / ~10 ms cached

Full design notes for each tool live in docs/architecture-harbormaster.md.

---

UI overview

The companion dashboard (harbormaster-ui --port 7531) is a local-first operator console. All surfaces speak the same SSE event stream that powers the MCP transport, so what you see in the UI matches what your MCP clients see.

Workspace shell (v19+)

Every page renders inside a fixed three-column workspace borrowed in spirit from VSCode / Postman / Linear:

┌─ topbar (h-12, fixed) ────────────────────────────────┐
├──────────┬──────────────────────────┬─────────────────┤
│ sidebar  │ main                      │ inspector      │
│ (240px)  │ (1fr, scroll-y)           │ (320px,        │
│ projects │ page-specific content     │  collapsible)  │
└──────────┴──────────────────────────┴─────────────────┘

• Topbar — brand-mark + page title + Cmd-K hint + theme toggle + auth indicator.
• Sidebar — all hosts ▾ filter + Filter projects… search + RECENTLY ASKED + language groups (GO, JAVASCRIPT, PHP, …) with ★-pinned project markers.
• Main — page content; layout choice (single column / card grid / split pane / tabs) belongs to the page.
• Inspector — context-aware widgets per page (KPI summary + live activity feed on the dashboard, project metadata + budget gauges on the project page, etc.). Collapse via « button (state persisted in localStorage); auto-collapses below 1280 px.

Linear-violet OKLCH palette, compact density, light/dark theme toggle.

Per-page surfaces

• Dashboard (/) — Quick Ask card at the top (project picker + question → navigates to project page with pre-filled query), KPI strip (projects, active embeds, recent queries, bridge state, dispatcher health, host budget headroom), 2-column card grid (recent activity, Auto-reembed, FleetQ Bridge, Plugins, Recall Q&A history full-width, Project graph full-width). Inspector shows a KPI mini-strip plus a live SSE-driven activity feed with pulse-on-new.

• Project page (/projects/<name>) — five-tab system: Overview / Memories / Trajectories / Q&A History / Settings. Number keys 1–5 jump tabs; #tab=<id> URL hash makes shareable deep links. Inspector shows project metadata + 24h budget consumption.

  

• Memories tab — split-pane editor: file list (left) + raw markdown textarea (centre) + bleach-sanitised live HTML preview (right). Toolbar: Save / Undo (Cmd-Z) / Redo (Cmd-Shift-Z) / diff vs: revision selector. Last-20 per-file history, side-by-side HTML diff, optional tag chip editor.

  

• Network (/network) — inter-project call graph rendered with a vendored Cytoscape build. Edge weights track real Harbormaster MCP calls (caller → target). Filters by host / project / tool / window; switchable graph / chat list view; SSE-driven live append. Aggregate stats at /api/network/stats?window=….

  

• Dispatcher trace (/dispatcher) — live in-flight spans + last-100 completed spans rendered as a waterfall with parent / child nesting. Each span exposes tool, project, host, duration, and (where the backend emits it) tool-call sub-spans for the model's own tool use. Real backend token usage in the SSE usage event.

  

Mobile (< 1024px)

Below 1024px the 3-column shell collapses into drawers. A hamburger (☰) toggles the sidebar; an info button (ⓘ) toggles the inspector. Click the backdrop or press Esc to dismiss.

Drawers closed	Sidebar drawer open

- **Cmd-K command palette** — bigram fuzzy-matched action launcher; shareable URLs via `?q=` pre-fill; pulls actions from a single catalog so every page surface adds itself for free.

Operator-facing reference: docs/operator-guide.md.

---

Configuration

Zero-config by default. For any other layout, drop a TOML file at ~/.config/harbormaster/config.toml:

[projects]
glob = ["~/code/*", "~/work/*"]
exclude = ["**/node_modules/**", "**/vendor/**"]

[hosts.friday]
ssh_host = "katsarov-server.local"
remote_htdocs = "~/htdocs"

[hosts.hetzner-1]
ssh_host = "hetzner-1.example.com"
remote_htdocs = "/var/www"

# Optional — opt in to Q&A history / recall
[history]
enabled = true

# Optional — daily call budgets
[budget]
daily_call_budget_per_tool = { ask_project = 200, delegate_task = 50 }

A per-project override at ./.harbormaster.toml in your cwd takes precedence over the user-level config. Validate at any time with:

harbormaster-mcp config check

Full schema (every section, key, type, default, valid range): docs/operator-config-reference.md.

---

Backends

Harbormaster's backend abstraction is a Protocol; the project ships two first-party implementations:

Backend	Default for	Notes
claude (claude -p)	[backends.default]	The reference backend. Per-token streaming, real backend-reported token usage in the SSE usage event, tool_use sub-span instrumentation.
codex (Codex CLI)	opt-in via [backends.codex]	Token instrumentation parity (v12.0.0a1). Tool-use sub-span instrumentation parity (v17.0.0a2). Same ask / delegate / fan-out surface.

Switching a project's backend is a TOML edit — no code changes. SSH hosts each carry their own backend setting (backend = "codex" on the host block) so you can mix backends across the fleet.

Pre-flight on each remote host: install the backend's CLI, authenticate once, ensure project paths exist with their CLAUDE.md / .serena/ in place, and confirm passwordless SSH from your machine (BatchMode=yes is enforced by Harbormaster).

---

FleetQ Bridge integration (optional)

Install with the [fleetq] extra and Harbormaster registers itself as a Bridge daemon, advertises its MCP tools as Platform Tools, publishes an A2A Agent Card per project, and writes back semantic triples to the federated KnowledgeGraph:

pipx install 'harbormaster-mcp[ui,fleetq]'

[fleetq]
enabled = true
register_as_bridge = true
base_url = "https://app.fleetq.net"
api_token_env = "FLEETQ_API_TOKEN"
heartbeat_interval = 30

export FLEETQ_API_TOKEN=...
harbormaster-mcp

Harbormaster shows up in the FleetQ Connections UI as harbormaster on <hostname>. Reverse-tunnel calls flow from FleetQ → Bridge → Harbormaster transparently with text/event-stream forwarding (X-Accel-Buffering: no so reverse proxies don't buffer).

Discovered contract reference: docs/fleetq-bridge-contract.md.

---

How it compares

	Harbormaster	A single MCP per project	cd between projects	Bash aliases
Number of MCP server processes	1	N (one per project)	0	0
Q&A history searchable across projects	✅	❌	❌	❌
Inter-project call graph	✅ Cytoscape network view	❌	❌	❌
Live trace waterfall	✅ /dispatcher	❌	❌	❌
Multi-axis budgets	✅ per-host + per-tool + per-project	❌	❌	❌
Remote SSH project routing	✅ same tools target host=...	requires N SSH MCPs	manual ssh + cd	manual ssh + cd
Memory editor (CLAUDE.md, .serena/memories/*.md)	✅ split-pane + bleach preview + revision history	edit on disk	edit on disk	edit on disk
Web operator console	✅	❌	❌	❌
Setup per new project	0 (auto-discovered)	new MCP entry	n/a	n/a

FAQ

Does Harbormaster work without Claude Code? Yes. Any MCP-compatible client works — Claude Code, Claude Desktop, Codex, Cursor, Aider, or your own integration via the MCP spec.

Do I need FleetQ? No. Harbormaster is fully functional standalone. The [fleetq] extra is purely additive — register as a Bridge, publish A2A Agent Cards, mirror trajectories into FleetQ Memory. All opt-in.

Where is my data stored? Locally. SQLite at ~/.harbormaster/qa_local.db (Q&A history) + ~/.harbormaster/network_log.db (inter-project calls) + ~/.harbormaster/memory_revisions.db (last-20 memory revisions per file). No telemetry, no phone-home. The optional [telemetry] config block is opt-in only and routes to your own endpoint.

What about security? Bearer-token authentication on every UI endpoint + cookie-backed SSE (v12+). Memory editor goes through bleach.clean() on render (allowlisted tags, blocks javascript:/data: schemes). Path-traversal protected per project (only CLAUDE.md + .serena/memories/*.md).

Can I run it as a long-lived daemon? Yes. harbormaster-ui --port 7531 & then bookmark http://127.0.0.1:7531/. Or wrap in launchd / systemd. Or use a Tauri shell (planned, separate project: harbormaster-desktop).

How is this different from the MCP Inspector? MCP Inspector is a debugging UI for a single MCP server. Harbormaster is a multi-project router plus operator console — it exposes MCP tools that target many projects, and visualizes their interactions.

---

Versioning

Harbormaster ships on a proven alpha-cadence: each major (vN.0.0) is preceded by a sequence of vN.0.0aK alpha tags, every alpha is a PyPI-published release, and the GA tag vN.0.0 is a no-code promotion plus a cumulative retro doc.

The cadence has shipped GA 18 times (v1 through v18) without a single forced rollback. Current head: v21.0.0a3.

Every behaviour change lands in one alpha and shows up in CHANGELOG.md under the corresponding major release. Per-major retros under docs/sprint-retro-harbormaster-v*.md are the source-of-truth narrative if you want the why behind a change.

PyPI publishing is automated via Trusted Publishing (OIDC) — no API tokens in the repo. Tag-pushes to v* trigger the publish workflow.

---

Pre-commit hooks

Two repo-local hooks ship in .pre-commit-config.yaml:

• harbormaster-config-check — validates examples/harbormaster.toml against the TOML schema. Fails the commit on any schema error.
• harbormaster-config-doc-parity — fails the commit if a Pydantic field is added to src/harbormaster/config.py without a matching mention in docs/operator-config-reference.md. On failure, the hook emits a copy-paste-ready markdown stanza so adding the missing doc takes seconds.

Install once:

uv sync --extra dev
bash scripts/post_sync_install_hooks.sh

Or with a system / pipx pre-commit: pipx install pre-commit && pre-commit install.

---

Contributing

Harbormaster is primarily a single-operator tool. Pull requests are welcome when they fit the alpha-per-feature cadence:

1. One feature per branch — feat/v<N>.<P>-<phase-name>.
2. Tests + mypy --strict + ruff clean before requesting review.
3. Update docs/operator-config-reference.md for any new config field (the doc-parity hook will tell you).
4. Keep changes additive when possible — Harbormaster's invariant is zero breaking changes per release line.

Operator workflows and deployment scenarios are documented in docs/operator-guide.md.

---

Lineage

Harbormaster v1.0 grew out of project-router-mcp v0.1 (2026-05-08). v0.1 git history is preserved on this repository — the v0.1 single-file server lived at src/server.py in commits prior to the v1.0 scaffolding refactor.

---

License

MIT — see LICENSE.