Coral
Coral is an open-source data engine that gives AI agents one governed SQL connection to APIs, databases and files. Each source becomes a read-only schema agents can join across, exposed over MCP or the command line.
What is Coral?
Coral is a single SQL interface that sits between AI agents and the systems a company already runs. Point it at an API, a database or a set of files and each one becomes a read-only schema: coral source add github turns GitHub into tables such as github.pulls and github.issues. Because every source is expressed the same way, an agent can join across them in one statement — matching PagerDuty incidents against the pull requests merged during the outage, or Slack messages against open Linear issues — instead of stitching several tool calls together by hand.
Twenty-six sources ship with the product, among them GitHub, GitLab, Slack, Linear, Jira, Confluence, Notion, Datadog, Sentry, Grafana, PagerDuty, Stripe, Gmail, Google Calendar and CloudWatch. Two backends are supported: HTTP APIs, where Coral translates SQL into paginated requests, and files in Parquet, JSONL, JSON or CSV. Anything missing is added through a YAML source spec, either community-maintained or written in-house.
Agents reach Coral over MCP or through the CLI, both served by the same local runtime. The MCP surface is deliberately narrow — sql, search, list_catalog, describe, list_columns and a handful of others — and SQL execution is configured read-only, with DDL, DML and multi-statement queries disabled. Credentials go into the operating system keychain, kept apart from non-secret configuration, and the internal gRPC server binds to loopback. In local and self-hosted mode, data, credentials and usage history never leave the machine.
The publisher's own benchmark put 82 real tasks and 243 factual assertions through Claude Code on Opus 4.6, against the data providers' own MCP servers. Across everything it reports 20% better accuracy, 42% lower latency and 52% lower median cost; on the complex multi-hop work typical of coding agents, 31% better accuracy and 70% lower cost; on simple lookups the gap all but disappears.
Coral ships as a desktop app for macOS, Linux and Windows, as a CLI, and as a server a team can run under Docker or Kubernetes. It is open source under Apache 2.0, and grew out of Phoebe, the software reliability agent the same team built before unbundling its data layer.
What it does
- Query APIs, databases and files in standard SQL as though they were a single database
- Join across unrelated sources in one statement, such as PagerDuty incidents against the pull requests merged during them
- Expose the same runtime to any agent over MCP, or drive it directly from the command line
- Handle authentication, pagination, rate limits and schema mapping on the agent's behalf
- Explore the catalogue with search, describe and DESCRIBE EXTENDED before a single query is written
- Learn relationships, recommended joins and schema hints from past queries, and cache hot paths
- Scope what agents can reach with per-source permissions, workspace isolation and read-only SQL
When to use Coral / When not to
A quick filter to help you decide if Coral is the right fit.
When to use Coral
- Engineering teams wiring coding agents such as Claude Code, Codex or Cursor into internal systems
- Site reliability and on-call engineers correlating incidents, deployments and telemetry in a single query
- Security and compliance analysts joining identity, cloud and code audit trails across providers
- Product managers and delivery leads tracking progress across Linear, Jira, Notion and Slack
- Platform and data teams that need agent access to company data without letting that data leave their own infrastructure
When not to use Coral
- Non-technical users, since every workflow runs through SQL, a command line or a YAML source spec
- Teams that need agents to write back to their systems, because Coral is read-only by design
- Anyone shopping for a finished AI assistant or an agent framework, which Coral explicitly is not
- Organisations looking for a managed warehouse or an ETL pipeline, as Coral materialises nothing
- Windows-first or ARM Linux shops, given the unsigned Windows preview and the x86_64-only Linux builds
How to use Coral
A typical end-to-end flow, from setup to results.
- Install the desktop app for macOS, Linux or Windows, or run brew install withcoral/tap/coral for the CLI
- Add your first sources with commands such as coral source add github, coral source add linear or coral source add slack
- Supply each source's variables and secrets — tokens, workspace identifiers, file paths — from environment variables or interactively
- Explore what is now queryable: SELECT * FROM coral.tables, then DESCRIBE EXTENDED, search and list_columns
- Write SQL against a single source, then start joining several of them in one statement
- Register Coral with your agents over MCP, for example claude mcp add coral -- coral mcp-stdio or codex mcp add coral -- coral mcp-stdio
- Let the engine accumulate schema hints, recommended joins and cached hot paths as queries build up
- For team use, deploy the Coral server and UI with Docker Compose or on Kubernetes behind a TLS-terminating ingress
- Set up workspace membership, member and owner roles, and agent-scoped MCP tokens
- Export traces, logs and metrics to any OTLP backend to see what your agents are actually querying
Pros & Cons
Pros
- A published benchmark with full methodology, per-task results and honest reporting of where the gains disappear
- Apache 2.0 open source, with unlimited sources, agents and queries when self-hosted
- Data, credentials and query history stay on the customer's own infrastructure in local mode
- Read-only by construction, so agents cannot mutate the upstream systems they read from
- Framework-agnostic: any agent that speaks MCP, SQL or a shell can use the same runtime
- Authentication, pagination, rate limiting and schema mapping solved once for every agent at once
- A security model documented in public, including what it deliberately does not protect against
Cons
- Requires SQL and comfort with a command line; there is no no-code entry point
- Read-only only, with approval-gated writes still listed as a future plan rather than a feature
- Barely worth the detour for simple single-source lookups, where the vendor's own benchmark shows 22% more tokens consumed
- The Windows desktop app is an unsigned preview that does not update itself, and Linux builds are x86_64 only
- Explicitly not a sandbox: any connected MCP client can query every source you have installed
- Secrets fall back to plaintext local files whenever the operating system keychain probe fails
- No published DPA, no subprocessor list and no stated position on training models with customer data
Pricing & Plans
Coral offers a permanent free plan: the Open Source edition is free forever under the Apache 2.0 licence, self-hosted, with unlimited data sources, agents and queries. The lowest paid entry point is the Team plan at 249 US dollars per month, which includes 50,000 queries and bills 0.005 dollars for each additional query. Enterprise pricing is quoted on request. Amounts are shown on the pricing page in dollars, without an explicit currency code.
- query engine
- access control plane
- unlimited data sources
- agents and queries
- self-hosted
- bring your own LLM keys
- community support
- everything in Open Source
- plus admin tools
- Google and GitHub social sign-in
- user roles and team filtering
- source connector evals
- hosting on Coral
- automatic upgrades with no maintenance
- your own LLM keys
- everything in Team
- plus self-managed options
- SAML and SCIM provisioning
- advanced RBAC and agent policy controls
- an immutable audit log
- custom source connector development
- managed evals and optimisation
- self-hosted
Data, GDPR & hosting
A consolidated view of how Coral handles your data.
GDPR overview
Coral is run by a UK company, so the privacy policy is written against UK data protection law with the ICO named as supervisory authority, while the terms list the EU GDPR (2016/679) by name among the applicable laws both parties must comply with. Four lawful bases are set out — consent, contract, legal obligation and legitimate interests — and seven data subject rights are enumerated, with a one-month response target. Transfers outside the UK or EEA rely on adequacy decisions, the UK IDTA or the European Commission's standard contractual clauses. There is no automated decision-making with legal effect, and analytics tags load only after consent. Gaps remain: no Article 27 representative, no named DPO, no published DPA and no subprocessor list. The single contact is privacy@withcoral.com. Policy last updated 13 April 2026.
Who owns the data?
Coral's terms define Customer Content broadly: the data, queries, configurations, credentials, schemas, metadata and results that pass through the Services. Intellectual property stays with whichever party already held it, and the client receives only a non-exclusive right of use for the duration of the agreement. Where Coral runs on your own machines, the publisher states it does not access, store, process or control that data at all. Where it does hold personal data, Phoebe Technology Limited acts as controller for what you supply directly and as processor for what a client uploads, and may share it with hosting, payment, IT, accounting, audit and legal providers, with public authorities, or with parties to a merger or sale of the business.
Reuse rights
The Coral software is published under the Apache License 2.0, and the terms are explicit that any copy obtained, used, modified or distributed outside the commercial Services is governed solely by that licence: you may run it, fork it, extend it and redistribute it without asking permission, and terminating a commercial agreement does not extinguish those rights. The restriction is narrow and covers only the proprietary, non-open-source components, which may not be modified, translated, reverse-engineered or decompiled. Query results belong to the customer. In local and self-hosted mode they never reach the publisher at all, and no clause limits what you do with them afterwards.
Data retention & training
Hosting summary
Where your data lives depends on which edition you run. The Open Source edition executes in your own cloud or on your laptops, and the publisher hosts nothing: query results, cached data and credentials stay in your environment by default, with trace sharing offered only as an optional debugging aid. The Team plan is hosted by Coral in an AWS-EU region, and there agent queries and data source API responses do pass through the publisher's infrastructure, although LLM keys remain yours. Enterprise customers choose between self-hosted, fully air-gapped and SaaS deployment. For personal data the publisher processes itself, the privacy policy allows storage or processing outside the UK and EEA under adequacy decisions, the UK IDTA or standard contractual clauses. Website analytics split between PostHog Cloud (EU), which stores within the European Union, and Google Analytics, which may transmit truncated IP addresses to servers in the United States. The publisher is a UK company; English and Welsh law applies and the ICO is the named supervisory authority. No hosting country is named for the product itself.
Things to keep in mind
Risks and trade-offs to weigh before adopting Coral.
- Any connected MCP client can query every source you have installed, so the exposed surface is the sum of all configured credentials
- The documentation states plainly that Coral is not a sandbox for untrusted agents, source specs or credentials
- A source token reaches whatever the upstream provider allows it to reach, which makes least privilege entirely the user's responsibility
- A third-party source spec can make HTTP requests and read local file paths it declares, so only trusted specs should be installed
- If the operating system keychain cannot be used, secrets are written to plaintext local files, which is easy to miss
- On a shared deployment, filesystem access to the state directory or the Postgres URL is equivalent to owning every workspace
- Making personal data trivially joinable across Okta, CloudTrail, Slack, Zendesk and Stripe invites internal surveillance and purpose creep
Setup & Integrations
Technical difficulty
Getting started is easy: a signed DMG on macOS, an AppImage or deb on Linux, an installer on Windows, or a single Homebrew command, and the site promises a first query within sixty seconds. Real use is harder. You need to write SQL, obtain and configure an access token for each source, and understand how the schema maps onto the underlying API. Team deployment adds Docker Compose or Kubernetes behind a TLS-terminating ingress, with authentication and workspace management. Connecting an unsupported API means writing a YAML source spec. Comfortable for engineers; out of reach without them.
Deployment
Integrations
Behind Coral
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
What is Coral, exactly?
Do I have to move my data into a warehouse first?
Does my data leave my environment?
Does Coral store my data?
How does it relate to MCP?
What does it cost?
Which sources are included out of the box?
Can an agent change my systems through Coral?
Where does it run, and on what?
Who is behind Coral?
Should you pick Coral?
Coral occupies a narrow, clearly stated position: it is the plumbing between agents and company data, not an agent and not a framework. That restraint is its strength. By turning every API, database and file into a read-only SQL schema, it removes the pagination loops, retry logic and glue code that make agent retrieval expensive, and it does so once for every agent rather than once per integration.
It is also unusually candid about its own numbers. The benchmark is published in full — 82 tasks, 243 assertions, per-task results — and reports plainly that on simple single-source lookups Coral costs slightly more tokens than the alternative. That honesty makes the headline figures on complex work, 31% better accuracy and 70% lower cost, considerably easier to trust.
The commercial model is equally legible. The open-source edition is genuinely complete and free forever, the Team plan sits at a flat $249 per month, and Enterprise is quoted on request. Running everything locally, read-only, with credentials in the operating system keychain, answers the objection most enterprises raise first about giving agents access to internal systems.
The reservations are those of a young product. Public launch was in April 2026; the Windows build is an unsigned preview, Linux is x86_64 only, SQL, NoSQL and vector sources are still in development, and enterprise SSO and audit logs are only planned. The legal documentation is serious — ISO/IEC 27001:2022, GDPR named in the terms, English law — yet silent on a DPA, on subprocessors and on whether customer data ever trains a model. For a team fluent in SQL and willing to run infrastructure, Coral is one of the more thoughtful pieces of agent tooling available. For anyone else, it is early.
- Choosing a selection results in a full page refresh.
- Opens in a new window.