AI Assistants (MCP)

Fleeta MCP is a remote Model Context Protocol server for your dashcam fleet. Register one address in the AI assistant you already use — ChatGPT, Claude, Cursor, or any other MCP client — sign in with your API key once, and ask about your fleet in plain language:

Which vehicles braked hard yesterday? · How many dashcams are offline, and since when? · Draw me last Tuesday's route for Truck 114. · Recall the 14:07 clip from Van 07's SD card.

The assistant picks the right tool, calls the Open API on your behalf, and answers from live data. There is nothing to install or host, and nothing an assistant can change on a vehicle without a person approving it first. It is included in your Fleeta subscription — the same API key, the same plan, no separate fee.

Production MCP URL   https://mcp.fleeta.io/
Authentication       Authorization: Bearer flt_…   or   OAuth 2.1 (automatic discovery)
Tools                28 — 20 read · 1 playable link (direct, uses mobile data) · 6 write (approval gate) · 1 confirm step
Transport            Streamable HTTP, stateless JSON — no SSE connection required

What you can do

Ask about	Tools the agent uses	From plan
Vehicle status — online, parked, offline, firmware, battery	list_devices, get_device	free
Safety events — harsh braking, impacts, speed-limit violations, the 18 event types	list_events, safety_event_stats	starter
Trips, live locations and GPS tracks	fleet_locations, list_trips, trip_track	starter
Fleet summaries and driving reports	fleet_summary, driving_report	standard
Event video (a presigned URL the agent hands back)	get_event_video	standard
Geofences and enter/exit alerts	list_geofences, geofence_alerts, create_geofence…	pro
Recordings still on a dashcam's SD card — list, play or download over LTE, delete	list_sd_recordings, play_sd_recording, recall_sd_recording…	pro
Your own API usage and audit trail	api_usage, audit_logs	starter / enterprise

Every tool needs a scope, and scopes follow your subscription tier — an agent reaches exactly what your own team already reaches, never wider. The full list with scopes and tiers is in the Tools reference.

How it works

 You, in chat            Your AI client            Fleeta MCP               Open API
 ─────────────           ──────────────            ──────────────           ─────────────
 "Which vehicles    ──▶  picks a tool         ──▶  https://mcp.fleeta.io/   ──▶  /v1/events
  braked hard            list_events(              validates your key /          ?type=harsh_braking
  yesterday?"              type=harsh_braking,     OAuth token, checks the       (your fleet only)
                           window=24h)             scope, forwards the call
                    ◀──  reads the JSON       ◀──  returns the API response ◀──
 "Truck 114, 3 times —   and writes the answer
  worst at 14:07…"

1. One URL. The client speaks MCP to https://mcp.fleeta.io/. No SDK, no local process, no ports to open.
2. Your key, your permissions. Every tool call is authorized by the API key behind the connection — the same scopes, tenant isolation and rate limits as the REST API.
3. Reads run, writes wait. The 20 read tools answer immediately. The 6 write tools return a proposal; nothing happens until a person approves it and the agent calls confirm_action — see the approval gate. One tool sits in between: play_sd_recording runs directly, but it is not a read — it makes the dashcam upload the low-resolution sub-stream of a clip over LTE (a few MB; full quality goes through the gated recall_sd_recording), and is annotated readOnlyHint: false so your client does not treat it as a free lookup.
4. How the answer looks is your client's call. The server returns data, not formatting: figures arrive as label-and-value pairs ready to plot, and the server's instructions ask the assistant to show trends, distributions, rankings and comparisons visually rather than only in prose — but they leave the choice of form to the assistant, because it knows what it can render and the server does not. Whether that becomes a chart, a table or a downloadable file depends on the client you connected, and on what your plan and organization settings allow it to do. Two clients can answer the same question in different shapes from the same numbers.

Try it in two minutes — no dashcam needed

The public sandbox keys work on the MCP server exactly as they do on the REST API, so you can connect an agent to a 12-vehicle demo fleet before you own a single camera:

URL     https://mcp.fleeta.io/
Key     flt_test_ent_0000000000000000001      ← public test key, enterprise scope set
Ask     "How many vehicles are offline, and for how long?"

Then follow the setup for your client in Connect your client. When your cameras arrive, replace the test key with a live flt_live_… key from Management › Open API in the webviewer — nothing else changes.

On this tab

• Connect your client — ChatGPT, Claude, Cursor, Claude Code, local stdio, other MCP clients
• Tools reference — all 28 tools by domain, with scope, tier and annotations
• Write actions & approval gate — the two-step flow behind every write
• Example prompts — what to ask, by job
• Authentication & security — Bearer vs OAuth, per-scope consent, key rotation
• Limits, troubleshooting & FAQ — quotas, errors, client quirks

Roadmap

• More tools — the catalog grows with the REST API.
• Agent skills and plugins — packaged instructions for coding agents, alongside the MCP server.
• Machine-readable docs — every page is available as Markdown; see llms.txt at the portal root.

Anything not listed as live above is planned, not shipped.