eve-mcp
Provides BioBASE tools for RoxyBot, a voice-controlled Android launcher, including contact resolution, profile lookup, memory search via EVE, and append-only logging.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@eve-mcp search my clipboard history for the Amsterdam address I copied"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
eve-mcp
MCP-server voor EVE (klembord- en transcriptiegeschiedenis op de Mac), uitgebreid met drie tools voor RoxyBot, de spraakgestuurde Android-launcher. In RoxyBot heet deze bron "BioBASE".
De server roept de eve-CLI aan en geeft de JSON terug. Alle data blijft op de Mac.
Tools
Tool | Doet |
| De bestaande EVE-tools (ongewijzigd) |
| naam → |
| sleutel ( |
|
|
RoxyBot vertaalt memory_query naar eve_search. Via de relay mag de bot alleen
contacts_resolve, profile_get en memory_query gebruiken; log_append roept alleen de telefoon
zelf aan. De telefoon dwingt dat af.
Related MCP server: altic-mcp
Vereisten
macOS met de EVE-app en de
eve-CLI (standaard/opt/homebrew/bin/eve)Node.js 20 of nieuwer
Voor RoxyBot: Tailscale op de Mac en op de telefoon
Installeren
git clone https://github.com/netvistamedia/eve-mcp.git
cd eve-mcp
npm install
npm test # compileert (tsc) en draait de tests; raakt geen draaiende dienst
npm run buildvoert na het compilerenlaunchctl kickstart … ai.eve.mcpuit. Gebruiknpx tscofnpm testals je alleen wilt bouwen.
Als MCP-server voor Claude Code (stdio)
Wijs de bestaande eve-server in je Claude-config naar deze build:
"eve": {
"command": "/opt/homebrew/bin/node",
"args": ["/pad/naar/eve-mcp/dist/index.js"],
"env": { "EVE_BIN": "/opt/homebrew/bin/eve" }
}Of kopieer dist/ over een bestaande installatie, bijvoorbeeld ~/.claude/eve-mcp/dist.
RoxyBot-gegevens
De drie RoxyBot-tools lezen en schrijven in ~/Library/Application Support/eve/roxybot/
(of EVE_ROXYBOT_DIR). Maak de twee JSON-bestanden zelf aan:
// contacts.json
[{ "naam": "Sarah de Vries", "nummer": "+31612345678", "aliassen": ["Sarah"] }]// profile.json
{ "thuis": "Keizersgracht 1, Amsterdam", "werk": "Stationsplein 2, Utrecht" }log_append schrijft regels naar log.jsonl in dezelfde map (alleen leesbaar voor jou).
Past een naam bij meer dan één contact, dan geeft contacts_resolve null: liever niets dan
de verkeerde persoon bellen.
HTTP-modus voor RoxyBot (via Tailscale)
RoxyBot bereikt EVE via Streamable HTTP op je Tailscale-adres, met een bearer-token. Zonder
EVE_MCP_TOKEN weigert de HTTP-modus te starten.
Eenmalig proberen
export EVE_BIN=/opt/homebrew/bin/eve
export EVE_MCP_HOST=$(tailscale ip -4) # 100.x.y.z — alleen het tailnet
export EVE_MCP_PORT=8787
export EVE_MCP_TOKEN=$(openssl rand -hex 32)
node dist/index.js
# eve-mcp: Streamable HTTP on http://100.x.y.z:8787/mcpBlijvend via launchd
Maak ~/Library/LaunchAgents/ai.eve.mcp.plist (vul paden, adres en token in):
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
<key>Label</key><string>ai.eve.mcp</string>
<key>ProgramArguments</key>
<array>
<string>/opt/homebrew/bin/node</string>
<string>/pad/naar/eve-mcp/dist/index.js</string>
</array>
<key>EnvironmentVariables</key>
<dict>
<key>EVE_BIN</key><string>/opt/homebrew/bin/eve</string>
<key>EVE_MCP_HOST</key><string>100.x.y.z</string>
<key>EVE_MCP_PORT</key><string>8787</string>
<key>EVE_MCP_TOKEN</key><string>JOUW-TOKEN</string>
</dict>
<key>RunAtLoad</key><true/>
<key>KeepAlive</key><true/>
<key>StandardErrorPath</key><string>/tmp/eve-mcp.log</string>
</dict>
</plist>chmod 600 ~/Library/LaunchAgents/ai.eve.mcp.plist # het token staat erin
launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/ai.eve.mcp.plist
launchctl kickstart -k gui/$(id -u)/ai.eve.mcp # herstarten na een nieuwe buildClaude Code start zijn eigen stdio-exemplaar en heeft deze dienst niet nodig.
Koppelen met RoxyBot
Met dezelfde variabelen als hierboven:
EVE_MCP_HOST=… EVE_MCP_PORT=… EVE_MCP_TOKEN=… npm run roxybot-qrDit toont een QR-code (roxybot-eve://<host>:<poort>?token=<token>&v=1). Scan hem in RoxyBot
via instellingen → BioBASE (EVE) → QR-code scannen, of vul host, poort en token daar handmatig
in. Het script accepteert alleen Tailscale-adressen.
Beveiliging
De HTTP-modus luistert alleen op het opgegeven adres; gebruik je Tailscale-IP, niet
0.0.0.0.Het token geeft toegang tot je EVE-geschiedenis: bewaar het niet in gedeelde bestanden en laat de QR-code niet open in beeld staan.
RoxyBot bewaart het token versleuteld in Android Keystore en praat alleen met Tailscale-adressen.
Licentie
MIT — zie LICENSE.
Available Tools
15 toolscontacts_resolveResolve a contact (RoxyBot)A
Find a phone number by name for RoxyBot calls and messages. Matches name or alias, ignoring case and accents; a unique first-name match also counts. Returns {contact: {naam, nummer}} or {contact: null} when nothing or more than one person matches.
Inputs: naam (required).
eve_activityActivity aggregationA
Aggregate counts of clips and transcriptions grouped by app, origin, content_type, hour, or day. Use for 'what apps did I work in today', 'when am I most active this week'. Optionally scope to one app.
Inputs: limit, range, group_by, source_app.
eve_app_statsPer-app statisticsA
Statistics for a single source app: total clips, breakdown by content type, first/last seen, favorite count, total copies. Use for 'how much did I copy from Safari this month'.
Inputs: bundle (required), since.
eve_categoriesList categoriesA
All user-defined categories with entry counts and purpose descriptions. Categories group mixed history entries (clips, transcriptions, recordings, live talks) into projects; the description states why a category exists. Use the names with eve_recent/eve_search category filters or eve_categorize.
Inputs: no named parameters.
eve_categorizeAssign or remove a categoryA
Adds a category to (or removes it from) history entries — the write path for organizing entries into projects from the agent side. Targets are type:id pairs from eve_recent/eve_search/eve_get_selection results. The running app refreshes automatically.
Inputs: targets (required), add, remove.
eve_export_clipExport a clip from a recordingA
Cut a shareable mini-clip out of a screen recording, second-exact. Workflow: call eve_get with type 'recording' to read transcript_segments (millisecond word positions on the video timeline), pick one or more cut ranges, then call this tool. Multiple ranges are stitched into ONE clip in order. The master video is never modified; the clip lands in a clips/ folder next to it and appears as its own row in the app's Aufnahmen tab, carrying a remapped timed transcript of its own.
Inputs: title (required), segments (required), recording_id (required), format, +1 more.
eve_getGet full clipA
Fetch one clip, screen recording, screenshot, or transcription by id with full, untruncated content. Clips: html, rtf size, OCR-extracted blob_text, optionally the linked voice transcription. Recordings (type 'recording'): the full recording row plus transcript_segments — millisecond word positions on the video timeline, the data needed to pick cut ranges for eve_export_clip. Screenshots (type 'screenshot'): full OCR text and the blob hash — resolve it to bytes via the blob-path CLI command / /api/blob route. Transcriptions (type 'transcription'): the full transcription row plus a linked screenshot (blob hash + OCR text), when one was captured at dictation stop. Use after eve_search/eve_recent returns an id.
Inputs: id (required), type, include_context.
eve_get_selectionWhat is the user looking atA
The history entries the user has currently selected in eve's history board — their live focus context. Returns each selected item with a snippet/title. Check updated_at: it is the moment of the last selection change; if it is old, the selection may be stale and worth confirming. Empty items means nothing is selected right now.
Inputs: no named parameters.
eve_ingest_recordingIngest a persona answer videoA
Deliver a finished persona answer video into eve: copies the file into the user's recording folder, writes the history row and pops the persona bubble in the app. Use after generating a talking-head video for the user. Requires an absolute path to a local mp4/mov file.
Inputs: path (required), title (required), persona (required), transcript_file.
eve_live_talkLive-Gespräch turn (voice loop with the user)A
Voice-conversation loop with the user. Call it in a loop: pass your answer to the previous utterance as reply (the app speaks and stores it), then the call blocks up to ~50s and returns the user's next utterance. Rules: after status utterance answer briefly and speakably, then call again with reply. After silence call again immediately (no reply). Stop calling on paused, ended or busy. Pass talk_id verbatim if you received one in the start instruction; on a session-initiated start omit it and send agent_label once.
Inputs: reply, talk_id, agent_label.
eve_recentRecent clips or transcriptionsA
List the most recent clips or transcriptions, newest first. Use for 'what did I copy/dictate recently' or, with filters, 'recent links from Safari' / 'my longest dictations'. Supports sorting (recency|created|copy_count for clips, duration for transcriptions), favorites, time bounds, and keyset pagination via meta.next_cursor.
Inputs: sort, type, limit, order, +14 more.
eve_restoreRestore soft-deleted entriesA
Brings soft-deleted history entries back (clears their deleted flag). Find candidates with eve_recent's deleted option. The running app refreshes automatically.
Inputs: targets (required).
eve_searchSearch clips & transcriptionsA
Full-text search (FTS5) across copied clips and voice transcriptions on this Mac. Use to find a specific item by its words — e.g. 'the URL I copied from Slack', 'what I dictated about the budget'. Returns clips[] and transcriptions[] with match snippets. Default sort is relevance (best matches first, top-N). For chronological paging through many results, set sort to recency/created and a single source (clips OR transcriptions), then page with the returned cursor.
Inputs: query (required), sort, limit, order, +14 more.
log_appendAppend a log line (RoxyBot)B
Record one call or message line from the RoxyBot phone: type (call, sms, whatsapp, …), doel, ts (epoch ms), status and optional inhoud. Written to log.jsonl on this Mac only. Returns {ack: true}.
Inputs: ts (required), doel (required), type (required), status (required), +1 more.
profile_getRead a profile value (RoxyBot)A
Read a saved personal value such as the 'thuis' or 'werk' address. Returns {sleutel, waarde} with waarde null when unknown.
Inputs: sleutel (required).
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 15 tools
Most tools target clearly distinct resources or actions: search vs recent vs get vs selection vs export are separable, and the RoxyBot helpers (contacts_resolve, log_append) and profile_get stand apart. The main overlap is eve_activity vs eve_app_stats, which both produce aggregate counts and differ only in grouping scope; the descriptions help but an agent could pick the wrong one for 'how much did I copy from Safari'.
Consistent snake_case throughout, with domain prefixes (eve_, profile_, log_, contacts_) applied predictably to their respective areas. Minor deviation: some eve_ tools are verb-led (eve_search, eve_get, eve_restore, eve_categorize) while others are noun-led (eve_activity, eve_categories, eve_app_stats), so the action isn't always the second token, but the set stays readable.
15 tools covers a genuinely broad surface — clipboard/dictation history, recordings, export, categories, live voice conversation, and phone-side logging/contacts. It sits at the top of the comfortable range but each tool maps to a distinct operation with no obviously redundant entries.
Read and organize workflows are well covered (search, recent, get, stats, categories, categorize, restore, export, live talk). Notable gaps: no category create/delete or rename, no way to write profile values (profile_get is read-only), and no entry delete — only restore of soft-deletes — leaving dead ends for common lifecycle operations.
Related MCP Connectors
Your AI apps as one team: hand work to Codex or Claude Code on your Mac, with a shared logbook.
Search, read, and write your Apple Notes from ChatGPT/Claude via a local Mac agent + MCP relay.
Mac & Windows: let ChatGPT, Claude & Cursor use your email, calendar, iMessage, Teams, files.
Persistent, verbatim, searchable memory for AI assistants — one memory across every MCP client.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceA standalone clipboard history MCP server for macOS that monitors clipboard in real-time, stores history locally in SQLite, and exposes it to any MCP-compatible AI assistant.5MIT
- AlicenseBqualityCmaintenanceA macOS automation MCP server providing 40+ tools for controlling Messages, Contacts, Notes, Reminders, Calendar, Files, Clipboard, Windows, Safari, Chrome, screenshots, and system settings via natural language.91181Apache 2.0

CrossPasteofficial
AlicenseAqualityAmaintenanceOpen-source cross-device clipboard manager with a built-in MCP server. Search and retrieve clipboard history synced across all your devices, read OCR text extracted from copied images, and add new items to the clipboard.52,713AGPL 3.0
Pastea MCPofficial
AlicenseAqualityBmaintenanceGives AI assistants local access to a Mac's clipboard history, letting them search past clips and lists, read a clip in full (including text recognized inside screenshots), and place a clip back on the system clipboard for the user to paste. Everything runs over 127.0.0.1 with explicit user approval, and assistants cannot delete clips or see items concealed by password managers.156 npmMIT