Frequently asked questions

Alloy is the AI Robotics Engineer: agents that investigate every issue across your fleet, built on a robotics data platform that captures, stores and queries every mission. The agents search telemetry, logs and recordings across every mission and answer with the evidence, in the Alloy app, Slack, Jira or any MCP client. Underneath is a data platform built for robot data: recordings are stored as Parquet in an open Iceberg catalog and queried server-side across petabyte-scale fleet history. Sensitive fields can be redacted on the device before data leaves your network.

Yes, and this is a common workflow. You can ask Alloy to analyze your data and produce a result, such as a model or a root-cause report, then hand that result to a coding agent like Codex to implement the fix. You find the root cause in Alloy, grounded in your real recordings, and ship the change in your editor, without exporting data by hand between the two.

You do not need ROS, because Alloy works with any MCAP file regardless of what produced it. If you do use ROS, the Docker edge agent supports ROS 2 (Humble and Jazzy) out of the box, and the lightweight binary agent uploads MCAP from any folder.

Alloy handles petabyte-scale robot data. A single robot can log around a terabyte a day, and Alloy is built to search across your entire fleet's history in one query, because the work happens server-side instead of loading anything into the agent's context. The more you upload, the more cross-fleet patterns it can find.

There is no separate charge for the MCP. Alloy is usage-based with no per-seat fees, and Mesh Storage is $0.02 per GB per month with no egress fees, so you are not charged every time you or an agent reads your data. You can start free with no credit card.

No. With the Alloy MCP, anyone on the team can ask a question in plain English and get an answer, without writing SQL or maintaining their own analysis scripts. The agent turns the question into a query for you. This replaces the scattered one-off plotting scripts engineers usually build, and lets less technical teammates get answers straight from the data.

Yes. Alloy keeps context about your fleet, your scenarios, and your previous investigations, so the agent picks up where you left off instead of starting fresh each session. Raw Codex or Claude forgets everything between sessions, which is one reason repeated fleet analysis is faster through Alloy.

Yes. Alloy stores recordings as Parquet in an open Iceberg catalog, so the same data the MCP queries is available to DuckDB, Spark, Trino, and PyIceberg directly. The MCP queries this lake with read-only SQL, and your other tools can read it in parallel with no exports.

The Alloy MCP works with any MCP-compatible client. It has been used with Claude Code, Claude Desktop, Codex, Cursor, Windsurf, ChatGPT, and Notion AI. One server connects to many agents, so engineers can stay in whatever tool they already use.

Yes. Alloy coexists with Foxglove. Alloy Edge can watch the same recording directory as the Foxglove Agent, so you can add Alloy without removing the viewer your team already uses. Alloy adds query, AI access, and fleet-wide analysis on top of your existing workflow rather than replacing it.

The Alloy MCP uses OAuth. When you first connect your AI tool you authenticate in the browser with the same login as the Alloy web app, and the session refreshes automatically. There are no API keys to generate, paste, or rotate for the MCP connection.

No. Alloy stores your data in an open Apache Iceberg catalog with Parquet underneath, and your MCAP files go in unchanged. You can query the same data directly from DuckDB, Spark, Trino, or any Iceberg client at any time, with or without Alloy. If you ever leave, the data is yours in open formats.

Every answer from the Alloy MCP is grounded in a real query over your actual recordings, not the model improvising. The agent does not summarize raw logs from memory; it runs a query server-side and reasons over the returned result, so answers trace back to real data and timestamps. This is the main reason teams use the MCP instead of pasting logs into a chat window, where the model is more likely to guess.

Alloy indexes every recording into queryable tables and does the heavy lifting server-side. When your agent asks a question, Alloy runs the query against those tables and returns only the matching slice, so a terabyte of telemetry never enters the context window. The agent receives a compact result set it can reason over, which keeps answers fast and grounded even across a whole fleet.

The Alloy MCP exposes robotics-specific tools rather than generic file access. They include listing and summarizing missions, semantic search across instances such as images, logs, and signals, natural-language data queries that the agent turns into SQL, read-only SQL against your data lake, and report creation. The agent calls these optimized tools against your indexed recordings, which is why it returns grounded results instead of guessing.

Add the Alloy MCP server URL to your AI tool and authenticate once in the browser. For Claude Code it is a single command; for Cursor, Windsurf, and Codex it is one entry in the MCP config; for Claude Desktop it is added as a custom connector. Authentication uses OAuth, so there are no API keys to copy or rotate, and the connection persists for your session.

No. Alloy works with the MCAP you already record. You can upload from the web, push from devices, or point Alloy at a folder your recorder already writes to, with no changes to your capture stack. If your robot writes MCAP, it works.

No. The Alloy MCP is read-safe by design. Destructive actions such as deleting devices or files, rotating keys, and resetting configs stay in the dashboard UI and are never exposed to the agent. The agent can read and query your data, but it never holds keys that could change or delete it.

Raw Codex or Claude cannot open MCAP files and run out of context on large logs, so they tend to guess or surface irrelevant suggestions. Alloy connects to those same agents through its MCP and does the data work server-side, returning grounded results from your actual recordings. The difference is that Alloy keeps the agent's context clean and every answer traceable to a real query, which is what makes cross-fleet analysis and reliable root-cause work possible.

The Alloy MCP is a Model Context Protocol server that lets AI agents query your robot data in plain English. It acts as a search engine over terabytes of MCAP recordings: the agent asks a question, Alloy runs the query against your real recordings server-side, and returns just the slice that matters. It exposes robotics-specific tools for listing missions, searching instances, and querying data, and it can return a replay link that opens the run in the Alloy app when you want to see it visually.

Because raw robot logs are too large to fit in a context window. When an agent like Codex scans local MCAP files, it spends its context just figuring out which runs are relevant and has little left to reason about the failure. Alloy's MCP does the scanning and filtering server-side and hands back a compact, structured result, so the agent's context stays clean. This is why comparing ten or fifteen runs works through Alloy but stalls with a raw agent.

Not directly in a useful way. Codex and Claude cannot natively ingest MCAP files, and a single robot can log around a terabyte of data per day, far more than any model's context window can hold. The Alloy MCP solves this by querying your MCAP recordings server-side and returning only the relevant slice, so the agent gets a grounded answer without parsing raw logs or exhausting its context.

Connect the Alloy MCP server to your AI tool. Alloy exposes a Model Context Protocol server at one URL that Claude Code, Codex, Cursor, ChatGPT, and Notion AI connect to with a single command and a browser-based OAuth login, with no API keys to manage. Once connected, your agent can search missions, query recordings, and pull back results inside your editor, because the MCP queries your real MCAP recordings server-side instead of opening the files itself.

Alloy is the AI Robotics Engineer, built on a robotics data platform. It stores your MCAP and ROS bag recordings in managed, open-format storage and lets you analyze them in plain English through an MCP server that connects to AI tools like Claude, Cursor, and Codex. Unlike a general data warehouse, it is purpose-built for multimodal robot data such as telemetry, video, and sensor logs, and it lets an AI agent query across your whole fleet rather than one file at a time.

Your data stays in open formats. MCAP files go in. Underneath, Mesh Storage converts to Parquet and indexes with Apache Iceberg. You can connect DuckDB, Spark, Trino, or PyIceberg directly to your data at any time. There is no proprietary format and no lock-in. If you leave, your data is yours.

Usage-based, with no per-seat fees: you pay for the investigations Alloy runs, not for logins. Mesh Storage, if you use it, is $0.02 per GB per month with no egress fees, so reading your own data never costs extra. Try it free, or get an estimate for your fleet.

No. Mesh Storage works with any MCAP file regardless of what produced it. The Docker edge agent supports ROS 2 out of the box (Humble and Jazzy), but the binary agent uploads any MCAP files it finds in a directory. If your robot writes MCAP, it works.

Yes. Alloy has edge-side redaction built into the device agent. You define rules in a YAML file on the device. You can drop entire topics (e.g. /audio/**), hash specific fields (operator names, hostnames), or strip patterns like emails and IP addresses from free-text fields. Rules run on the device before data leaves your network. If a rule breaks, the file goes to quarantine, not to the cloud. Every redacted file carries an audit trail documenting which rules were applied.

Yes. Reports are shareable links that anyone can explore. Product managers, leadership, and sales teams can query data in plain English without needing to know Python, MATLAB, or ROS tooling.

Yes. Alloy is built for robotics-scale data. Storage is backed by Cloudflare R2 with no egress fees, so you are not penalized for querying your own data. The more data you upload, the more useful the search and analysis becomes.

Yes. Alloy is SOC 2 Type II audited, supports HIPAA requirements, and is aligned with NIST SP 800-171. Data is encrypted in transit and at rest, and it is never shared across customers. Sensitive fields can be redacted on the robot, and teams whose data can't leave their environment can run Alloy in their own cloud or on-premise.

Two ways, and neither is a project for your team. Upload a recording and get your first report in minutes, or give us access to the robot data you already have and our engineers connect it and run Alloy's first investigations on your recent issues. Teams usually get time back from the first week.

Alloy replaces the custom scripts, manual charting, and wiki-based reporting that most validation teams maintain. It works alongside your existing infrastructure. You can keep using Grafana, Jira, Slack, or S3. Alloy connects to what you already have.

Foxglove shows an engineer one recording. Alloy finds out why a robot failed: its agents search logs, telemetry and recordings across your whole fleet and answer with the evidence attached. Many teams keep Foxglove for viewing and run Alloy alongside it.

You don't have to take Alloy's word for it. Every finding shows its working: which mission, which signals, which time window and how it got there, one click from the raw data. When the data can't support a conclusion, Alloy says what's missing and hands the investigation to an engineer with the evidence attached.

ROS bags (.bag, .db3), MCAP, controller logs, Parquet, CSV, JSON and custom binary formats, even when schemas change between firmware versions. No conversion scripts or preprocessing required.

In August 2026 Alloy raised US$8 million (A$11.5 million) at a US$80 million (A$115 million) valuation. Square Peg led the round, with Blackbird and Airtree. That makes Alloy one of four companies backed by all three funds, alongside Canva, Airwallex and Lorikeet. Customer demand brought the round forward. The money builds out our US presence, grows the forward-deployed engineers who roll Alloy out with each team, and deepens the agents and knowledge base that make every robot run improve the next one.