Mellos Mapping MCP Server

Mellos Mapping

English | 简体中文

A live, terminal-native map of bottom-up development for Claude Code, ChatGPT desktop Codex mode and Codex CLI.

While Claude builds your system, a split pane beside the conversation shows the system's layered dependency map: primitive layers at the bottom, dependency edges that may only point downward, ghost nodes for what is designed, a spinner on what is being built right now, and solid green for what is built and verified.

A real session: Claude Code on the left, the map pane on the right. L0 is verified, L1 just lit up, everything above is still a ghost.

Why

Most progress reporting is a task list — a top-down worldview. A Mellos map grows the other way: an upper node can only stand on nodes below it, and the picture makes the discipline visible:

• The ghost design appears before any code. Claude declares the whole intended structure as dashed ghost nodes first; you can veto a bad design while it is still only a picture.
• The spinner is where Claude's attention is. One glance answers "what is it doing right now, and on top of what?"
• Done means verified. A node turns solid green only with evidence (a passing test run). If later work cracks a foundation, the node turns red — a cracked foundation under a spinning upper floor is the most honest status report there is.
• The map is a ledger, not a judge. The tools refuse only structural corruption (an edge pointing upward, a duplicate rank). Workflow is Claude's discipline, defined in the bundled skill; violations are made visible, never silently blocked.

Install

Clone the branch for your host, then run one command. No build is required.

Requires Node.js 18.17+ on the 18.x line, or 20.3+, and the corresponding host CLI on PATH. Supported platforms are Windows 10+, macOS 13.0+ and Linux with glibc 2.28+, each on x64 or arm64. The OS must also meet the selected Node.js version's requirements. Native lock bindings ship with the release; no local compilation is needed. The installer checks release integrity and all eight MCP tools, retains the runtime outside the clone, and configures the host. Start a new conversation after installation. See release and branch instructions.

Claude Code's marketplace installation is also available:

Two lines inside any Claude Code conversation:

/plugin marketplace add GuangminJu/mellos-mapping
/plugin install mellos-mapping@mellos-mapping

Or one line in a terminal:

claude plugin marketplace add GuangminJu/mellos-mapping && claude plugin install mellos-mapping@mellos-mapping

Requires the same Node.js and platform versions on PATH (install Node separately when using native Claude Code). No build step: dist/ is committed, so a clone runs as-is — dist/server.mjs (the MCP server), dist/watch.mjs (the pane), dist/mmap.mjs (the mmap toggle), dist/hook-session-start.mjs (the SessionStart hook that hooks/hooks.json registers) and dist/store-paths.mjs (the store's path vocabulary, which the plain-node pane launcher imports instead of restating filenames).

omp (Oh My Pi) reads the same plugin layout the Claude Code edition ships — the same marketplace catalog, .mcp.json, skill and slash command — so one release serves both hosts:

omp plugin marketplace add GuangminJu/mellos-mapping
omp plugin install mellos-mapping@mellos-mapping

That install follows main, where a release is a version bump. To freeze a version instead, add a checkout of one — the Claude Code edition ZIP from the GitHub Release, or a clone at the release tag:

omp plugin marketplace add "<checkout or extracted release directory>"
omp plugin install mellos-mapping@mellos-mapping

omp never reads hooks/hooks.json, so the session paragraph arrives through the plugin's omp host adapter (dist/omp-extension.mjs, declared in package.json#omp.extensions): the same store, the same policy text, and the same mmap shim install as the Claude hook. The pane is the same terminal split beside the session. Install it as a plugin: neither omp plugin link nor the npm package is a plugin, and both would leave the tools behind. See omp installation and limitations.

pi reads none of that layout: it loads an extension and has no MCP of its own, by design. So the plugin speaks MCP to itself — the extension declared in package.json#pi.extensions (dist/pi-extension.mjs) starts the same dist/server.mjs in the session's directory, which makes the eight mmap_* tools the same eight, under the same names, and builds the session paragraph from the same store and the same policy text the Claude hook prints:

pi install npm:mellos-mapping

Add -l to install it for the current project instead of every one, then start a new session: pi discovers packages at session start. The package's skills/ and commands/ travel with it — the same map discipline, and /mmap for the same slash command. See pi installation and behaviour.

The first session after installing asks you one question — how eager mapping should be — and records the answer for every project you will ever open. From then on the hook carries it into each new session by itself; there is no per-project setup step. See Setup: choose when maps open.

The mmap terminal command (the pane's open/close toggle) installs itself on Windows: the same hook notices on session start when the shim is missing or points at an older install, writes mmap.cmd (cmd, PowerShell) and mmap (git-bash) into %LOCALAPPDATA%\mellos-mapping\bin, appends that one directory to your user PATH, and tells you so through the assistant. The PATH change reaches only new processes — and a new tab of a running Windows Terminal inherits the old environment, so close the terminal app entirely and reopen it before the first mmap. The PATH edit keeps the installer's guarantees: nothing happens when the entry is already there, and a PATH that setx would damage (flattened %VARIABLE% references, truncation past its limit) is refused outright.

A refused PATH is not a dead end: the same two shims are then written into %LOCALAPPDATA%\Microsoft\WindowsApps (or ~/.local/bin) — a directory your PATH already names — so mmap is runnable in a new terminal with no PATH change at all. Machines with a long PATH are exactly where that happens, and where it matters most. The command takes the page to open as an argument (mmap omp-host-support), and closes the pane when it is already open.

The step behind it is still a command of its own, for the cases the hook does not cover — --uninstall, or re-adding a PATH entry you removed while the shims stayed put:

node "<plugin dir>/scripts/install-mmap-command.mjs" [--uninstall]

(--json prints the install outcome as one JSON line instead of prose — the mode the hook itself calls it in.) npm i -g mellos-mapping provides the same mmap via bin, no shims involved.

Update

claude plugin marketplace update mellos-mapping && claude plugin update mellos-mapping@mellos-mapping

Two steps because plugin update compares against the locally cached marketplace clone — the first command is what actually pulls this repo. Restart Claude Code to apply. Releases are version bumps on main. (In-app, /plugin opens the same management UI.)

omp updates with its own two steps:

omp plugin marketplace update mellos-mapping && omp plugin upgrade mellos-mapping@mellos-mapping

Close the pane first (q in it): the watcher runs from the plugin copy, and on Windows an open file cannot be renamed away, so an upgrade attempted while a pane is open can fail with EPERM and leave the plugin cache empty — repaired by omp plugin install mellos-mapping@mellos-mapping --force. Then: the first command refreshes the catalog, the second reinstalls from it. omp also refreshes a catalog entry it has not updated for a day at startup, unless marketplace.autoUpdate is off — in the default notify mode that check writes its finding to the debug log only, so marketplace update is the step that makes an update visible. upgrade does not compare versions: it force-reinstalls whatever the catalog names, which is why a release is identified by its version bump rather than gated by it. Restart omp afterwards, and reopen a pane still showing the old bundle (q in it, then mmap_open again).

pi updates the package it installed, with no marketplace in between:

pi update npm:mellos-mapping

That is the npm source; a git source is reconciled to the ref your settings name by pi's own update, and a versioned spec (npm:mellos-mapping@0.27.0) is pinned and skipped. Restart the session afterwards: a session owns the map server process it started, so one still running keeps the copy it began with until it ends.

Upgrading from 0.19

0.20 moved the store out of .claude/ — the map belongs to this tool, not to one client — into .mellos/. The server and the watcher each perform the move once at startup, and print exactly one line on stderr when they do (mellos-mapping: moved the legacy .claude map store to .mellos/ — commit the move.):

Nothing is merged and nothing is overwritten: a project that already has a .mellos/ store (map, pages or config) is left untouched, whatever the legacy directory still holds. If you keep your maps in git, commit the move — git add -A .claude .mellos records it as renames rather than as a pile of deletions plus untracked files.

That one move is the only time either process touches .claude/. Afterwards the tools write nowhere but .mellos/, and never outside the project directory they resolved at startup.

ChatGPT App · Codex mode

This is Codex mode in the ChatGPT desktop app (also called Codex App). From the source branch run the following command; on chatgpt-app, omit the host argument. It configures the desktop skill, marketplace and eight MCP tools.

node install.mjs chatgpt-app

Start a new conversation. The desktop skill defaults to web-terminal: mmap_open {surface: "web-terminal", page: "<slug>"} starts a local service, then the AI opens its URL in the current conversation's right browser panel. The mmap terminal starts automatically, with its own font-size selector; no manual paste or Computer Use is needed. Graphical SVG, native terminal and Markdown remain available. With an older MCP schema, run node "<plugin root>/dist/web.mjs" "<project>" --terminal --page <slug>. A queued panel request does not prove that the user can see it. See desktop installation and limitations.

For Codex CLI inside Windows Terminal (not the desktop integrated terminal), run node <plugin root>/scripts/open-pane.mjs <project dir> — it splits the terminal window hosting the session, keeping keyboard focus on the conversation. If that window cannot be identified or focused, it reports failure without opening elsewhere. --window explicitly chooses a separate window. Add --page <slug> to open on a particular page — and with a pane already open, rerunning with --page retargets the pane belonging to this console. A map open in another session or a separate window does not count as this split. The pane auto-follows the page being written — the map the agent is operating on right now; press f to toggle that (a manual page switch also turns it off), or start with --no-follow. Elsewhere run node <plugin root>/dist/watch.mjs from the project directory in a second terminal (or any terminal split). Both take the same flags — see Pane flags.

Desktop Markdown map

When the user chooses the document surface, the skill uses mmap_open {surface: "markdown", page: "<slug>"} and opens the returned file beside the current conversation. The document embeds a colored SVG dependency map plus module details, evidence, and child-page links. It needs no web server, browser rendering process, Mermaid support, or additional runtime dependencies.

Successful MCP map writes regenerate enabled previews in .mellos/previews/; JSON remains the source of truth. Image nodes are static. File visibility and automatic viewer refresh belong to the desktop host. See the desktop map guide for regeneration, cache behavior, and limitations.

Optional interactive web viewer

An additional interactive web viewer is available with mmap_open {surface: "web", page: "<slug>"} or node "<plugin root>/dist/web.mjs" "<project directory>" --page <slug>. Open its returned URL in the host's right browser panel. It provides live updates, pan/zoom, hover and pinned details, search, dependency highlighting, filters, group overview, submaps and themes. Stop it with node "<plugin root>/dist/web.mjs" "<project directory>" --stop. Markdown/SVG and terminal workflows remain available; all surfaces share the same map data. See web viewer details.

Any MCP client

The server ships on npm, so any MCP client (Cursor, Windsurf, Zed, Gemini CLI, …) can run it with a standard stdio entry:

npx -y mellos-mapping

The map file lands in the client session's working directory (.mellos/map.json). Open the live pane from the same project:

npx -y -p mellos-mapping mellos-mapping-watch

The server picks its project directory in this order: MELLOS_MAPPING_CWD (an explicit override, for clients that spawn servers from a fixed directory), then CLAUDE_PROJECT_DIR (what Claude Code sets for plugin MCP servers), then the server process's own working directory. Set MELLOS_MAPPING_CWD when your client would otherwise start the server somewhere other than the project you are working in.

The skill/discipline layer is Claude Code + Codex specific; other clients get the eight mmap_* tools and the pane, and bring their own prompting.

Use

1. Ask Claude to build something non-trivial. The bundled skill has Claude declare the ghost design and keep the map current as it works.
2. The pane opens itself. Every write tells Claude whether anybody is actually looking (see Who is watching), and Claude opens or retargets the pane with mmap_open when nobody is — you never have to remember to. Open or close it yourself with mmap in any terminal, or /mellos-mapping:mmap in the conversation (Windows Terminal split on Windows, tmux split inside tmux, or a printed command to run in any second terminal). On Windows the pane opens in the terminal window hosting YOUR session, even with several windows open; pass --window to put the map in its own dedicated window instead. Prefer --ascii if your font lacks box-drawing glyphs.
3. Watch nodes light up from the bottom. Interrupt when the picture worries you — that is what it is for.

On Linux and macOS, mmap_open and mmap automatically use tmux. The launcher targets the inherited TMUX/TMUX_PANE; if the tool process lost those variables, it discovers the single attached session on the default tmux server. The default is a right split that preserves input focus; --window creates a new tmux window. Repeated opens reuse the watcher bound to that source pane or session window.

For multiple attached sessions or a custom tmux socket, set these variables in the environment of the MCP server (restart it after changing them):

Ambiguous sessions, detached sessions and missing tmux produce a concrete error and a fully quoted watcher command, including the project and requested page, to paste into any visible terminal. Map writes remain available; after an open failure the assistant retries only when the environment changes or you ask.

The pane is mouse-aware (xterm SGR any-event tracking — the same protocol htop and tmux speak):

Every other key is inert, on purpose: an escape sequence the pane does not know (F-keys, Home/End, PgUp/PgDn, Insert/Delete, modified arrows) is consumed whole and does nothing, rather than having its payload bytes read as hotkeys.

Zooming scales the picture first and switches display mode only at the ends of the ladder, so every level still shows meaningful data:

detail+ ← detail ← 100% ← 85% ← 70% ← 55% ← overview

• zoom in past 100% — detail unfolds evidence and the first three lines of the design notes inside the boxes; detail+ widens them into reading cards (up to twelve note rows);
• 85–55% — whitespace tightens and labels truncate proportionally, boxes stay boxes;
• below 55% — labels would stop meaning anything, so the map AGGREGATES: each declared group (a labeled subsystem within a band) becomes one box named foundation subsystem 1/2 with its status derived from the members, edges collapse onto the groups, ungrouped nodes stay themselves. Like a real map, zooming out shows province names — not anonymous dots. (A map with no groups falls back to a pure glyph constellation with per-band counts.) The footer always names the level.

Below the map, between it and the hint line, sits a fixed-height detail panel: a separator you can drag, a status-colored header, the focused node's evidence, both wire directions (uses → … · used by ← …, each neighbour carrying its own status glyph) and its design notes, word-wrapped. With nothing focused it shows the map's dashboard instead. Fixed height — details never float over the map and the layout never jumps.

Glyphs

One status, one glyph, everywhere a map is drawn — the pane's boxes, its tab strip and detail panel, and any other client reading the same store:

The legend under the picture names the four statuses; □ done, no evidence joins it only on a map that actually contains one — the four are the vocabulary, that one is a rule being broken here and now. Documentation diagram kinds replace the status legend with their node-kind glyphs.

Pane flags

Both the pane launcher (scripts/open-pane.mjs <project dir>) and the watcher (dist/watch.mjs) take the same watcher flags; the launcher forwards them verbatim and rejects anything it does not know rather than dropping it.

Launcher-only: --window opens the dedicated "mellos-mapping" window instead of splitting the session's window, and --force opens another pane even though one is already running for this project. Watcher-only: --file <path> names the default page's state file (the launcher derives it from the project directory).

Pane ownership is carried by the internal watcher flag --owner <token> and the optional owner field in viewer reports. Launchers derive the token from the source console process and its creation time; manual watchers can omit it. Focus and quit requests to a bound viewer use its PID, so another window of the same project cannot consume them. Bare mmap toggles only the pane bound to the current console; mmap --window toggles the separate project window.

The mmap command

mmap, typed in any terminal, is a toggle: it opens the map pane for the project you are standing in, or closes the one that is already open.

The project is found the way git finds its root: from the current directory upwards, to the nearest one holding a .mellos/ store. Standing in a project that has no map yet is fine — the pane opens on its standby screen and says so until the first mmap_declare.

Closing goes through the store rather than through a signal: mmap writes a one-shot request beside the map, the pane consumes it on its next poll (250 ms by default) and exits, handing the terminal back exactly as it found it — mouse reporting off, cursor visible. A pane still on the standby screen closes the same way. The request is deleted as it is read, and a leftover from a pane that died is swept when the next one starts, so a stale request can never close a fresh pane.

Every watcher flag above works here too, forwarded verbatim; an unknown one is a usage error, never dropped in silence. mmap needs installing once unless you have the npm package. Inside a Claude Code conversation, /mellos-mapping:mmap opens the same pane.

Pages

A project can keep several maps side by side — one effort = one page. Claude targets a page by passing page to any mmap_* tool; the pane grows a tab bar as soon as a second page exists. The active tab is bold in its map's aggregate status color. Each page remembers its own pan, zoom and pinned node.

By default the pane follows the page being written — the map the agent is operating on right now — so declares and updates bring the audience along by themselves. f toggles that, a manual page switch turns it off, and --no-follow starts it off; with follow off, a background page's change lights its tab in status color instead of stealing your view. An explicit --page outranks follow, and a page requested before it exists is shown the moment it appears.

Deleting a page. An effort ends; its page does not have to stay. In the pane, x — or the × the active tab carries when the mouse is on — asks: the footer says press x again to delete <page>, and a second press within three seconds removes that page's file. Switching page, Esc or simply waiting takes the request back. The × sits on the active tab only, so clicking an inactive one switches to it first and offers its × on the next frame. From a tool call it is mmap_remove {pages: ["slug", …]}, applied after that call's map edits. Either way the file is gone for good — the maps are plain JSON, so committing them is the only undo there is.

State lives in the tool-owned .mellos/ directory at the project root:

The map files are plain JSON, safe to commit if you want the maps' history in git. Keep focus, quit, viewers/ and .write-lock out of Git. The first three carry runtime messages; the lock file must remain in place, including when no process holds its lock.

Concurrency, stated plainly. Every save is atomic — written to a private sibling temp file and renamed over the target — so a reader polling the store sees the previous complete map or the new one, never a torn write. Every graph writer (MCP, the HTTP viewer, the watcher) uses the same non-blocking OS exclusive lock for the project. The lock has its own fixed file, separate from the map files being replaced; it stays in place after the OS releases the lock on completion or process exit. Contention returns a retryable BUSY.

The lock serializes each operation's read, change and save. To protect an edit computed from an earlier read, pass expectedRevision: it is compared inside the lock, and a stale revision returns CONFLICT. Without it, a later write can still overwrite the same field or a whole value such as context. Pages separate unrelated efforts; sharing a page still requires revision-checked edits. Low-level saveMapFile, hand edits and older running processes do not participate in this contract. Stop all old MCP, viewer and watcher processes for the project before upgrading; an old lock directory needs a one-time migration. See the persistent-map API guide and project locking for the boundary and upgrade steps.

Diagram kinds

The default kind, dev, is the living progress ledger described above. The same layered-DAG machinery also draws documentation diagrams: pass kind in mmap_declare and the page renders neutrally — plain solid boxes, no ghosts, no spinners, no progress counts.

Node kinds and edge labels work on dev maps too. State machines are out of scope on purpose: transitions cycle, and edges here only point downward.

Sub-maps

A node can link a child page with submap: <page-slug> — the pane badges it ⊞; double-click dives into the child map, Backspace climbs back out. A map of maps, built entirely from pages: no new storage, no new invariants. Whether a node deserves a sub-map is the AI's judgment call — most don't.

Sub-maps are interior detail, not siblings: a page some other page dives into never occupies a tab. Two refinements keep the tab strip from erasing itself — a page whose own node names itself hides nobody, and a link cycle keeps its tabs unless a page outside the cycle dives in, because a cycle has no outside to climb back to. Inside a dive the tab row becomes a breadcrumb — ⌫ parent map ▸ node — and clicking it (or Backspace) climbs back out. When a hidden sub-map changes in the background, the footer says so.

MCP tools

Start with mmap_read to resume an existing page; a new conversation is not a new effort. Use stable IDs, filtered reads and revision-checked writes. The persistent-map API guide covers complete CRUD, atomic mixed batches, context checkpoints, source hashes and worktree boundaries.

A batch applies bands → groups → lanes → node updates, and within one node update layer moves the node before its other fields, so a node can move and join a group on its new band in one item.

What the boundary refuses, so the ledger never records something it did not mean:

• an unknown key, naming it — a misspelled evidance is an error, not a silently dropped field, at every nesting depth;
• control characters in text fields — an ESC sequence stored in a label would let a map repaint the terminal of everyone who opens it. detail is the exception: newlines and tabs are how a note is written, everything else (ESC, BEL, lone CR) is still refused. File loading and library saves apply the same control-character rules, retaining multiline evidence in older files. Terminal rendering also neutralizes controls in maps constructed directly by library callers;
• an empty string where a field is optional — null is how a field is cleared, never a blank that renders as a box nobody can tell from a real one;
• a node whose submap names the page the call itself targets — a link with no bottom, not a parent link;
• a page deletion aimed at the page the same call targets, or at a slug the project does not have — one call must not edit a map it is deleting, and a name matching no page is a typo far more often than a race; the refusal lists the pages that do exist.

A write that does not land answers save failed, nothing changed (retry): the previous file is intact and calling again is the whole recovery.

Who is watching

A map nobody has on screen is a file, not a map — and nothing in the system used to be able to tell the difference. An assistant would declare a design, light nodes up as it built them, and report all of it into a store you had never opened a pane for.

Every pane now publishes a small report while it runs — .mellos/viewers/, one file per pane, refreshed once a second — and every write and every view ends with what those reports say:

The same reports answer "is a pane already open?" for mmap and for the launcher — a question that used to cost a Windows-only process scan and could not say which page was on screen. A report whose pane stopped refreshing it is ignored after five seconds and deleted after a minute, so a killed pane cannot go on claiming an audience.

A heartbeat does not establish screen visibility. An explicit tmux open also selects the existing watcher's window, reveals it if another pane was zoomed, and verifies the attached session's active window before reporting visibility. Focusing that watcher in a shell without TMUX reuses its original owner.

Setup: choose when maps open

How eager mapping should be is a habit, not a property of a repository — so it is chosen once, for you, in the first session after you install:

• always — map every structured task: workflows, designs, architecture, technical dependencies. The assistant opens the pane on its own initiative; your recorded answer is its standing consent, so it stops asking.
• complex — the same, but only for medium or complex tasks: several modules, a new subsystem, roughly an hour of work or more.
• on-request — map only when you explicitly ask. In a project with no map, the plugin then says nothing at all — zero noise is the point.

The answer lands in <your home>/.mellos/config.json and reaches every session through the plugin's SessionStart hook, which reads it and hands the assistant the matching instruction before you have typed anything. Nothing has to be set up per project, ever again.

One project can still differ: mmap_setup {policy, scope: "project"} records a policy in that project's .mellos/config.json, and a project policy overrides the user one. /mmap setup re-runs the question for either scope whenever you want to change your mind. The policy guides the assistant; it never blocks the tools, and asking for a map explicitly always works under any policy.

Hosts without hooks (Codex CLI, a bare MCP client) get the question another way: while no policy exists in either scope, every mmap_declare reply carries a note telling the assistant to ask you. That note goes quiet for good — in every project — the moment you have answered anywhere.

Structural invariants enforced by the tools: layers form a total order by rank (an integer in 0..99, 0 = bottom, unique per map); every node lives in exactly one layer; edges point strictly downward — which makes the graph acyclic by construction; nodes may not depend on same-layer siblings (if A needs sibling B, either B is really a lower concept or A and B are one node); a group clusters nodes within one band; and node ids and group ids share one namespace — an id names a node or a group, never both, because both render as boxes and one id must mean one box.

Development

See contributing and the project structure, branch and recovery guide.

npm ci
npm run verify

Use Node.js 22.12+ for development. verify runs typecheck, test, build (bundles dist/ and emits lib/ with declarations), check:package, check:codex, and check:release. These check npm entrypoints, the Codex package and both installable editions, including their MCP handshake.

The repo is itself layered bottom-up, and each layer has its spec:

Documentation truncated — see the full README on GitHub.