Summary: Arka is both an MCP client (call external servers like SigNoz or filesystem) and an MCP server (expose Arka skills to Cursor). Config lives at
~/.config/arka/mcp.json in Cursor-compatible format. Quick start: arka mcp list, arka mcp add, arka mcp tools, arka mcp call. See also How to code with Arka for IDE workflow.arka plugin doctor and arka plugin inspect <name> to
check its normalized capabilities and permissions before routing requests to
it. Agent Hub exports the same catalog alongside Arka’s skill manifest.
Config lives at ~/.config/arka/mcp.json in Cursor-compatible mcpServers format.
Quick start
Config format
~/.config/arka/mcp.json:
command + args. Remote servers use url + optional headers. Optional env values support ${env:VAR_NAME} substitution.
Transports
Arka reuses the existing HTTP MCP client for remote servers and a built-in newline-delimited JSON-RPC client for stdio — no extra dependencies required.
Optional Python SDK
arka mcp status shows sdk on).
SigNoz MCP
For SigNoz specifically, Arka also ships traced helpers (arka signoz status, goal self-heal). Generic arka mcp works with any MCP server — not only SigNoz.
MCP logs
Use MCP logs when Cursor, Claude, or another MCP client reports production-like issues such asConnection closed, invalid tool arguments, unknown tools, or
server error states:
logs/mcp.jsonl; set ARKA_MCP_LOG_PATH to override it for
tests or hosted deployments.
Three.js model MCP
Arka has a built-in preset forbaryhuang/mcp-threejs, a Sketchfab-backed MCP server for finding downloadable Three.js-compatible 3D assets. This is useful when a scene asks for real things such as satellites, spacecraft, vehicles, rooms, desks, or humans — Arka should prefer a licensed GLTF/GLB asset over rough geometric placeholders.
Preview the config first:
Commands reference
MCP tool calls open a fresh connection per invocation. For high-frequency use, prefer keeping servers local (stdio) or on a low-latency HTTP endpoint.
Arka as MCP server
Other agents (Cursor, Claude Desktop, OpenClaw, Codex, etc.) can call Arka skills and memory over MCP — the same stdio transport Arka uses as an MCP client.Cursor setup
Use Arka as an MCP server inside Cursor so the editor agent can callarka_ask, arka_recall, arka_skill, and the rest of Arka’s exported tool surface.
1
Install Arka
From the repo (editable install) or your venv:
arka mcp doctor should end with summary ok and report the current tool count.2
Add MCP config
Option A — project-level (recommended when working in the Arka repo): copy or symlink Example snippet (paths vary by machine — use Repo helpers:
.cursor/mcp.json from the repo root. Cursor loads it automatically for this workspace.Option B — global: open Cursor Settings → MCP and merge the arka entry into ~/.cursor/mcp.json:arka mcp install for yours):3
Restart Cursor
Fully quit and reopen Cursor, or use Settings → MCP → Reload. The
arka server should show as connected with the current Arka tool count.4
Example tool calls
In Cursor chat, ask the agent to use Arka MCP tools:
- Recall memory: call
arka_recallwith{"goal": "project conventions"} - Ask Arka: call
arka_askwith{"prompt": "summarize this repo"} - Run a skill: call
arka_skillwith{"skill": "repo_map"} - Repo layout: call
arka_repo_mapwith{"depth": 2}
IDE and agent setup
Arka exposes a standard stdio MCP server, so the same generated entry works in most MCP-compatible IDEs. Generate a machine-specific snippet instead of copying a path from this page:mcpServers.arka entry to the client’s MCP JSON file, then
restart or reload its MCP panel. The entry is:
After configuring any client, verify the connection before asking the agent
to use Arka:
arka_route as the umbrella Arka MCP tool: pass the user’s full natural-language
request to it when you are not certain which narrower Arka tool should run.
Useful first checks are arka_heartbeat, arka_repo_map, and arka_skill.
Existing MCP servers are preserved when using arka agent_hub sync --unify; use
arka agent_hub adapters to preview which client files will be updated.
Quick start
arka on PATH:
arka mcp install to print the correct command for your machine.
Exposed tools
MCP-safe defaults
Arka MCP hides or blocks personal desktop/device skills by default so IDE agents do not unexpectedly open Brave, start Spotify, play media, or run a personalized daily brief. This does not change local Fish/CLI behavior. Disabled-by-default MCP examples:arka_spotifyplay_spotify,spotify_control,spotify_brave_debugplay_song,play_youtube,play_movie,stop_musicopen_url,open,browse,open_urlssearch_web,browse_web,agent_browserdaily_brief
browser_check, web_screenshot,
automate, repo tools, code tools, data tools, and arka_route remain
available.
Opt in only when the MCP client is trusted to touch your desktop/browser/media
state:
Agent Hub
agent_hub sync copies MCP config to the hub and ensures the arka self-server entry is present so other launch agents can merge it:
ARKA_MCP_CONFIG, ARKA_HUB_DIR) are documented in Agent Hub.
Import IDE memory via MCP
arka_agent_hub can list and import memory from Cursor, Codex, OpenClaw, Arka session notes, and hub exports:
Other
arka_agent_hub actions: status, adapters, detect, doctor, list. See Agent Hub — Memory import for source ids and formats.
Using Arka MCP to extend Arka
Dogfood the MCP workflow when adding skills: use Arka’s own MCP tools to discover gaps, then wire new capabilities back into CLI routing.Example workflow
arka_heartbeat or arka_jsonkit), add:
route_command()in the Python moduleroute_*()insrc/arka/routing/symbolic.py- Fish helpers
_agent_is_*_request/_agent_route_*inconfig.fish - Tests in
tests/test_*_routing.pyandtests/test_nl_routing_coverage.py
ROUTE_MODE=symbolic_only in sync.
Browser safety
MCP requests do not open websites or desktop browser windows implicitly. Use headless skills such asbrowser_check, web_screenshot, or automate for
inspection. If a workflow genuinely needs a headed tab, call the arka_skill
tool with allow_browser: true after obtaining user approval. This prevents a
model-generated URL from unexpectedly launching a browser.