model context protocol • Live
Okareo in
your editor.
Connect Claude Code, Cursor, or your editor to the hosted Okareo MCP, run fifty voice calls in parallel and know if a change worked without leaving your copilot.
0
packages to install
8+
editors supported
WORKS WITH YOUR EDITOR
CC
Claude Code
CD
Claude Desktop
CU
Cursor
VS
VS Code 1.101+
GH
GitHub Copilot
CL
Cline
WS
Windsurf
GM
Gemini Code Assist
Voice Agents • The Real Ask
Are you
really
going
to
pick
up
the
phone
and
make
50
calls
yourself?
Ask the Okareo MCP to run dozens of synthetic voice conversations in parallel — in the languages and personas your real customers actually call from. Audio, transcript, and trace come back in moments. You'll know if today's change to the agent worked before the meeting ends.
3:12
to know if it worked not three days
Project Awareness
Reads your code. Picks the right Drivers.
Ask for tests and the Okareo MCP doesn't make you fill in a form. It opens your repo, reads your agent's system prompt, the tools it can call, the policy strings hard-coded into the flow — and proposes Drivers that fit this agent. Not generic personas. Not a template. Drivers shaped by your code.
Natural • Language Control
Click
through
the
UI.
Just
ask.
Everything Okareo's web app does, the MCP exposes as tools. Generate Drivers. Run simulations. Pull traces. Compare runs. Open a scorecard. Replay a failed call. None of it requires you to leave your editor — you describe the outcome, your co-pilot calls the right tools in the right order.
Closed Loop • In IDE
Production failures become PR-blocking tests without ever leaving your editor.
Something broke last night. You don't open a dashboard, copy a call ID, fill in a scenario form, or paste a yaml file. You type one sentence to your co-pilot. The MCP fetches the call, generalizes it into a Driver, adds it to your suite, and wires the gate into your CI. The next PR will fail until you fix it.
Production fails at 3:42 PM
You ask Claude Code, in passing
Driver lives in your suite. Every PR runs it.
Connect
Install.
Just
point
and sign in.
The Okareo MCP is hosted at tools.okareo.com/mcp. Nothing to install, nothing to run locally. Point your editor at the URL, sign in through your browser once, and the tools appear in the picker.
One-line install with the claude CLI:
A browser tab opens to app.okareo.com on first connect. Sign in, consent, and you're done — open any project and ask "list my Okareo scenarios" to verify.
Or paste into .mcp.json at your project root (or ~/.claude.json globally):
{ "mcpServers": { "okareo": { "type": "http", "url": "https://tools.okareo.com/mcp" } } }
Wire Okareo into your editor.
Run 50 calls before lunch.
Common Questions
Do I need to install anything locally?
No. The Okareo MCP is a hosted endpoint at https://tools.okareo.com/mcp over MCP's streamable-HTTP transport. There's no package to install, no Python environment to set up, no container to run. You point your editor at the URL, sign in once through your browser, and the tools appear.
How does authentication work?
OAuth on first connect — a browser tab opens to app.okareo.com, you sign in and consent, and your editor stores the token. If your editor doesn't yet implement MCP OAuth discovery, set OKAREO_API_KEY in your environment and use the Bearer-header fallback shown in the docs.
Which editors does it support?
Claude Code, Claude Desktop, Cursor, VS Code 1.101+, Cline, Gemini Code Assist, GitHub Copilot, and Windsurf are officially supported. Any MCP client that speaks streamable-HTTP transport works — the server is one URL and one JSON config block away from any compatible tool.
Can I really run 50 voice calls in parallel from a prompt?
Yes. Ask your co-pilot to list driver voices, create a few personas across languages, and call run_simulation against your voice target. Okareo runs the calls in parallel, the trace + transcript + audio land on the same timeline, and you can re-score the whole batch against new checks without re-running the model.
How does the MCP know what to test?
Your co-pilot reads your project — the agent's system prompt, the tools it calls, hard-coded policy strings — and proposes Drivers whose personas stress the parts most likely to break. It then uses tools like save_scenario, create_or_update_driver, and list_targets to wire everything up. You approve the plan before any run goes out.
Does it work with private codebases?
Yes. Your editor reads your local files the same way it always does — nothing about your repo is sent to Okareo beyond what your co-pilot decides to include in a specific tool call (typically just the agent code or transcripts you've explicitly named).
Can I ingest production voice calls through the MCP?
Yes — connect a Retell, Twilio, Vapi, or ElevenLabs integration via connect_voice_integration, paste the webhook URL into the provider, and completed calls flow into Okareo for scoring and trend analysis. You can also pipe transcripts in directly via ingest_conversations.
Are there rate limits?
The hosted endpoint is throttled per credential at 60 requests/minute/organization by default. Tool calls that exceed it return a 429 with a retry_after field. If your traffic profile warrants a higher limit, Okareo support can raise it.