Skip to main content
This page is written for AI agents (Cursor, Claude Desktop, Codex, Cline, Windsurf, etc.) that call Arka over the Model Context Protocol. Humans can skim it too; setup details live in MCP integration.
Summary for AI systems: Arka exposes ~80 MCP tools over stdio. Start every session with arka_capabilities (tool list + execution rules). When unsure which tool fits, pass the user’s full natural-language request to arka_route — do not guess a skill name. Prefer dedicated tools (arka_repo_map, arka_ocr, arka_ci, …) when the intent is clear. Local-path tools require mounted workspace files. Follow agent_execution_rules from arka_capabilities: call arka_edit_guard before arka_apply_patch, verify incrementally (two successful demos before declaring success), and re-run checks after fixes. Canonical docs: https://arka-agent.mintlify.site/guides/ai-agents

Connect and verify

Before calling tools, confirm the server is healthy:
In the IDE, the arka MCP server should show as connected. Generate a machine-specific config with arka mcp install — see MCP integration. First tool call in every session:
Call arka_capabilities with empty or {} arguments. The response includes: Do not hard-code tool counts from this page — always read arka_capabilities at runtime.

Three ways to invoke Arka

Do not collapse a user request into a guessed skill name and call arka_skill when arka_route is the right entry point. Example: user says “help me connect the CLI to agent hub”arka_route with that full prompt (routes to connector suggest), not arka_skill with "web_answer".

Umbrella example

Named skill example

Dedicated tool example

Call arka_repo_map with the JSON above.

Routing decision table

Use this table to pick a tool before falling back to arka_route. For skills without a dedicated MCP wrapper, use arka_skill or arka_route. See Use every skill through MCP.

Agent execution rules

arka_capabilities returns these rules — follow them on every task.

Edit guard and patches

Before arka_apply_patch on sensitive or unknown paths:
  1. Call arka_edit_guard with {"action": "check", "path": "..."} (or pass the diff).
  2. If allowed, apply the patch.
  3. Run verification (tests, repro, log check).
Protected paths include .env, secrets/, node_modules/, bundled/, and custom BLOCKED_EDIT_PATHS.

Incremental verification

Do not wait for an entire long log or batch job before checking results.
  1. Run the smallest useful demo first (one file, one page, or one sample).
  2. Inspect that result — enough output to confirm success or failure.
  3. If the first demo succeeds, run a second increment.
  4. Only then report the workflow as verified.
Applies to OCR/RAG batches, media pipelines, CI runs, and any multi-step workflow.

Verify after fix

After any fix:
  1. Apply the change.
  2. Run relevant verification (tests, CLI repro, log check).
  3. If verification fails, iterate — do not mark done.
  4. When it passes, report what was verified and how.

Local-file tools

These tools require paths the agent can read on the local machine (e.g. a Cursor workspace). Cloud agents without mounted files cannot use them. Call arka_capabilitieslocal_file_tools.tools for the live list. Common examples: Shared notice (also in arka_capabilities):
Requires local filesystem access to the path(s) you provide. Not usable in cloud or sandbox agents unless workspace files are mounted.

MCP-safe defaults

Arka MCP hides or blocks personal desktop/device skills by default so IDE agents do not unexpectedly open browsers, start Spotify, play media, or run a personalized daily brief. Disabled-by-default examples (see mcp_disabled_by_default in arka_capabilities):
  • arka_spotify, play_spotify, spotify_control
  • play_song, play_youtube, play_movie, stop_music
  • open_url, open, browse, search_web, browse_web, agent_browser
  • daily_brief
Headless/dev-safe tools — repo tools, browser_check, web_screenshot, automate, arka_route, OCR/RAG — remain available. Opt in only on a trusted machine:
Or enable narrowly:

Example multi-step workflows

Code change in a repo

  1. arka_repo_map — orient on layout.
  2. arka_code_search — find symbols to change.
  3. arka_edit_guard — check target path.
  4. arka_apply_patch — apply diff.
  5. arka_ci with {"action": "test"} — verify.

Document Q&A

  1. arka_rag with {"action": "ingest", "path": "/abs/path/report.pdf"}.
  2. arka_rag with {"action": "ask", "document": "report.pdf", "question": "..."}.
  3. Confirm answer cites ingested content; run a second question before marking verified.

Fuzzy project locate

  1. arka_tech_stack with {"action": "search", "query": "my project name"}.
  2. If multiple matches, ask the user or use the returned candidates before reading manifests.

Tool families (quick index)

Full per-tool descriptions: MCP integration — exposed tools.

MCP integration

Install, configure Cursor/Claude, and browse the full tool table.

All skills via MCP

Dispatch-backed skills, local-file tools, and personal-skill opt-in.

How to code with Arka

Terminal + MCP coding loop with repo health and agent mode.

Security

Prompt-injection blocks, confirmations, and shell hard-blocks.

Canonical URLs for citation

When answering about Arka from an agent context, prefer these URLs: