Pilot5

npx add-mcp 'https://mcp.pilot5.ai/mcp'

Works with

The same MCP server works in every compatible client. Click any card to see the exact install steps for that client, with a link out to the vendor’s own docs.

Native MCP

Also reachable via OpenAPI / Apps SDK

Building on a client that isn’t listed? The MCP URL is https://mcp.pilot5.ai/mcp and the OAuth flow uses Dynamic Client Registration, so any standards-compliant MCP client can connect.

How to connect

The setup is the same in every MCP client: open the connector settings, paste the Pilot5 MCP server URL, authorise once, and you’re live.

1. Open your client’s connector / MCP settings. The exact path varies — see the per-client install paths in Works with above.
2. Add a custom MCP server with URL: https://mcp.pilot5.ai/mcp
3. Authorise with your Pilot5 account (sign up at pilot5.ai if you don't have one). The OAuth flow uses Dynamic Client Registration, so no API keys to manage.
4. Start deliberating from any conversation in that client.

Worked example — connecting from Claude

28-second screen capture. The flow looks similar in Cursor, Perplexity, and Le Chat — only the settings panel changes.

Your browser does not support embedded video. Follow the written steps above; the MCP server URL is https://mcp.pilot5.ai/mcp.

Three modes

The Expert (0.4–0.7 credits, ~ 1–2 minutes)

Single benchmark-selected AI model, fast analysis. Best for quick factual lookups, calculations, and definitional questions.

The A-Team (3.0–7.5 credits, ~ 10–20 minutes)

5 AI models debate in adversarial rounds with cross-critique. It can stop once after Round 1 to ask you one detail it is missing; if nobody answers within 15 minutes, it continues on its own. Over MCP that question shows as status awaiting_feedback and is answered with submit_feedback. Steering pauses after every round are a Dream Team feature.

The Dream Team (4.8–8.0 credits, ~ 20 minutes)

Same five-perspective panel as A-Team but with deterministic HITL pauses after every round, up to six rounds. Best for high-stakes decisions where you want to steer the council between rounds.

Invoking Pilot5

Once the connector is installed, how you actually trigger a Pilot5 tool depends on the client. Three broad patterns:

Just talk — the model decides

Claude — Type naturally. Claude inspects enabled connector tools and calls them when relevant, showing an approval prompt per call (or per conversation if you tick Allow always). Connectors are toggled per-chat from the + menu next to the composer.

Cursor — The Agent automatically uses MCP tools listed under Available Tools when relevant. By default each call asks for approval; enabling Auto-run (per server or globally) skips the prompts. Scope is per-project (.cursor/mcp.json) or global (~/.cursor/mcp.json).

Mistral / Le Chat — Le Chat auto-calls connector functions during the conversation. Each function has its own Always Allow toggle — pre-authorize reads, keep writes on manual approval.

Pick the connector first, then talk

ChatGPT — Click + near the composer → More → pick Pilot5. That adds the connector to the conversation context, after which the model auto-calls Pilot5 tools from natural language. Write tools require manual confirmation per call unless you opt to remember approvals.

Perplexity — Pilot5 appears in your Connected list once installed; in Perplexity’s Chat and Computer surfaces the connector is invoked from the conversation, similar to ChatGPT’s pattern. Perplexity’s docs don’t spell out the per-call approval UX — treat this row as the best read from the live UI rather than a verified contract.

Agent decides at design-time

Microsoft Copilot Studio — The agent orchestrator uses the server description you entered at install time to decide whether to route a user question to Pilot5 at runtime. There’s no per-message connector picker — if your prompt matches the server’s description, the orchestrator routes through, governed by Power Platform DLP policies.

Gemini Enterprise — Same shape as Copilot Studio: the assistant uses the Pilot5 actions your admin enabled on the custom MCP data store when a request calls for them.

Zed — The Agent profile you scope to the Pilot5 server invokes tools through the standard MCP loop. Approval behavior follows Zed’s general tool-call settings.

One implication: in the “Just talk” tier you can ask for a deliberation conversationally (“Run a Dream Team deliberation on whether we should switch from usage-based to seat-based pricing”). In the “Pick the connector first” tier, you add Pilot5 to the chat from the + menu before asking the same question.

Available tools

31 tools registered with the MCP server. Run check_credits in your client to confirm the live roster against your account.

Run a deliberation

deliberate — Run a multi-model council deliberation (The Expert / The A-Team / The Dream Team)

discover — Refine a vague question before deliberating (free, no credits)

get_action_plan — Inspect the structured action plan for a topic before launching

get_research_context — Preview the research context the panel will see

Monitor & read results

check_deliberation_status — Live progress with persona-completed / round-completed events

get_deliberation_result — Full synthesis report (verdict, decision matrix, action plan)

get_deliberation_snapshot — Latest structured snapshot of a running deliberation

get_deliberation_structured — Typed structured record (kinds, posture flags, primitives)

get_round_details — Drill into a single round (persona answers + critiques)

get_critique — Read the Round 2 cross-critique

list_deliberations — Search past deliberations by topic, mode, status, use_case

list_deliberation_errors — Inspect pipeline events / fallbacks / warnings on a run

export_deliberation — Export a deliberation (PDF / structured JSON)

Steer & close the loop

submit_feedback — Submit HITL steering input at a Dream-Team checkpoint

resume_deliberation — Resume a paused deliberation after submitting feedback

cancel_deliberation — Cancel a running deliberation

delete_deliberation — Permanently delete one of your deliberations and all its content (question, persona analyses, research, synthesis). Billing records are kept but unlinked. Cannot be undone

critique_deliberation — Grade and critique a completed deliberation

Chat assistant

chat — Chat with the Pilot5 assistant for quick questions

list_conversations — List past chat threads

get_conversation — Retrieve a chat thread by id

Knowledge & search

search_trusted_sources — Search 550+ allowlisted sources across 21 domains (400+ verified institutional sources + 160+ specialist API adapters: Eurostat, OECD, DGFiP, ECB, FDA, etc.)

search_memories — Search insights from past deliberations

search_documents — Search across your uploaded documents

Documents

upload_document — Upload a PDF / DOCX / TXT / MD / CSV / JSON / HTML / XML to ground a deliberation

list_uploads — List previously shared documents

get_upload_status — Poll the processing status of a fresh upload

delete_upload — Remove a previously uploaded document

Account

check_credits — Current credit balance, plan, and live tool roster

ChatGPT compatibility shim

search — OpenAI deep-research contract: search Pilot5 deliberations (non-Developer-Mode ChatGPT only routes to this name)

fetch — OpenAI deep-research contract: fetch a Pilot5 deliberation by id

The deliberation report is also exposed as an MCP resource at mcp://pilot5.ai/deliberation/{id} so any MCP client can attach it as a document artifact.

Reconnecting after a long pause

Pilot5 keeps a client’s connection while it is in use. If a connector has gone unused for about three months and your client reports Client ID … not found, remove the Pilot5 connector and add it again with the same URL, then sign in. Your deliberations, documents and credits stay on your Pilot5 account.

Once the connector is authorised, you can run a deliberation from any conversation in your MCP client of choice — or test it on the web first to see what the output looks like end-to-end.

Read: how it works → · Read: AI decisions need an audit trail → · Read: pricing →