@vinciai/mcp

A stdio MCP server that lets your coding agent (Claude Code / Codex) sync product knowledge from your repository into your Vinci support Knowledge Base — the repo never leaves your machine; only the curated entries you approve are sent.

Setup

1. In the Vinci dashboard → Workspace Settings → Connect a coding agent, mint a workspace key (vinci_wk_…).

2. Add the server to Claude Code:

   claude mcp add vinci \
     -e VINCI_TOKEN=vinci_wk_… \
     -- npx -y @vinciai/mcp

   For Codex, add to ~/.codex/config.toml:

   [mcp_servers.vinci]
   command = "npx"
   args = ["-y", "@vinciai/mcp"]
   env = { VINCI_TOKEN = "vinci_wk_…" }

3. In your agent, run vinci_sync for repo knowledge, or vinci_sync_sources to sync from connected sources and local documents.

Tools

vinci_whoami · vinci_list_agents · vinci_list_knowledge · vinci_sync_knowledge (create/update, keyed on externalKey, lands as draft; accepts sourceRefs for traceability) · vinci_archive_knowledge

• vinci_get_sync_state — read the saved incremental cursor for a stable source key such as github:org/repo, jira:PROJ, or doc:policies.pdf.
• vinci_set_sync_state — save the next cursor after an approved sync run succeeds.

Syncing from your sources

Prompt vinci_sync_sources guides a quality-gated sync from:

• connected MCP sources such as Jira, GitHub, Notion, or Linear;
• local documents such as PDF, Excel, or text files the host agent can read;
• this repository or other local repo files.

The Vinci MCP is the sink. It does not connect to Jira or GitHub itself. Your agent reads the source locally, rewrites useful items into customer-facing knowledge, shows you creates/updates/skips for approval, then calls vinci_sync_knowledge with a stable externalKey, a tag, selected agent ids, and sourceRefs.

Each source has a per-source watermark stored by vinci_get_sync_state / vinci_set_sync_state. Re-run vinci_sync_sources after each release; the prompt uses the watermark to process only new or changed items, and advances the cursor only after the approved sync finishes.

Distilled vs raw files

The MCP path always creates distilled, customer-voice entries. To ingest a raw file verbatim, use the Vinci dashboard file uploader instead.

Agent configuration (scope: agentconfig:write)

• vinci_get_agent_config — read an agent's current config (tone, temperature, behaviour, system prompt, escalation).
• vinci_update_agent_config — apply a partial config update. Live immediately for end customers.
• Prompt vinci_setup_agent — a guided interview that reads the current config, asks about identity / chatty↔careful / brand voice / creativity / escalation, shows a current → proposed diff, and writes on confirmation.

Newly minted workspace keys include agentconfig:write. Keys minted before this feature must be re-minted to gain the scope — until then the agent-config tools return a 403.

Notes

• Requires VINCI_TOKEN. VINCI_API_URL defaults to https://api.joinvinci.ai and can be overridden for staging or local development.
• Entries land as drafts — activate them in the Vinci dashboard.
• Internal development can still run the package from this repo with pnpm --filter @vinciai/mcp start.