Coral
Also known as: Phoebe, Phoebe Technology Limited
Coral, formerly Phoebe, is an open source data retrieval layer that gives the customer's AI agents one read only SQL interface across SaaS, internal systems and files, over SQL, a CLI and MCP.
Coral is the product of Phoebe Technology Limited, which built Phoebe, a vertical AI agent for software reliability. In April 2026 Phoebe became Coral, and its team now builds the data retrieval layer for AI agents. The phoebe.ai address redirects to withcoral.com.
Coral gives the customer's agents one SQL interface across data sources, translating each query into API calls or file reads across bundled sources such as GitHub, GitLab, Slack, Datadog, PagerDuty, Linear, Jira and Stripe, plus community and custom source specs, with cross source joins, catalog discovery and value search. SQL is read only. Agents reach it through the CLI or its MCP server, and each query emits OpenTelemetry traces with a span per source call.
The open source edition is free under Apache 2.0 and runs locally or self hosted on Docker or Kubernetes, with data and credentials kept on the customer's machine. A hosted Team plan and an Enterprise plan add sign in, roles, SAML, SCIM, audit logging and air gapped or SaaS deployment.
Vendor details
Canonical URL
https://withcoral.com
Category
Agent infrastructure
Company status
independent
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
Bundled sources include GitHub, GitLab, Slack, Datadog, PagerDuty, Linear, Jira, Stripe, Notion and Confluence, plus community and custom source specs. Agents reach them over SQL, a CLI and MCP.
In practice
Your team's agents need answers that span GitHub, Jira and PagerDuty. Coral gives them one read only SQL interface across those sources and joins the results in a single query.
Your security team will not let data or credentials leave your own machines. Open source Coral runs locally or self hosted on Docker or Kubernetes, and data, credentials and usage history stay on your machine.
Your team needs to see what its agents pulled from each system. Coral emits an OpenTelemetry trace for every query, with a span for each source call, that you can export to any OTLP backend.
Sources & related URLs
Research sources
Agentic Index coverage score
5.5 / 14 capabilities · 39%
| Integrations & Tool Calling | Partial |
|---|---|
|
The customer's agents get one SQL interface over bundled sources. These include code hosts (GitHub, GitLab), chat (Slack), monitoring and incidents (Datadog, PagerDuty), issue tracking (Linear, Jira) and billing (Stripe), and community and custom source specs add more. SQL execution is read only, with DDL, DML and multiple statements disabled, so agents read outside systems and take no action in them. Sourcewithcoral.com/docs/project/securityread 2026-09-29 |
|
| Workflow Orchestration | Not documented |
|
As a query layer the customer's agents call, Coral plans and joins one SQL query across sources but runs no agents, steps or workflows of its own. The former Phoebe agent is retired. Sourcewithcoral.com/docsread 2026-09-29 |
|
| Knowledge Grounding & RAG | Partial |
|
Agents retrieve live data across the customer's sources through one SQL query, with catalog discovery and a search over values Coral observed during earlier queries. Data is fetched from each source at query time. Coral describes no maintained index of the customer's content. Sourcewithcoral.com/docs/guides/use-coral-over-mcpread 2026-09-29 |
|
| Human Oversight & Guardrails | Partial |
|
Scoping keeps the customer's agents in bounds. SQL is read only, workspace membership decides what a caller may reach, and agent credentials are refused every owner level operation. Coral names no approval step before an action commits. Sourcewithcoral.com/docs/project/securityread 2026-09-29 |
|
| Security, Identity & Governance | Full |
|
Workspace membership is separate from login, with member and owner roles. Agent tokens are refused owner operations, and credentials are stored in the OS keychain. Enterprise adds SAML, SCIM, advanced RBAC and an immutable audit log. Coral publishes no compliance attestation. Sourcewithcoral.com/docs/guides/shared-server-access-controlread 2026-09-29 |
|
| Observability & Auditability | Full |
|
Each query emits an OpenTelemetry trace with a span for every source API call and every MCP tool call made on the agent's behalf, recording tool name, endpoint, status and error. Traces export to any OTLP backend, and owners can read them in a shared workspace. That gives a record, call by call, of what the customer's agents did through Coral. Sourcewithcoral.com/docs/guides/observe-with-opentelemetryread 2026-09-29 |
|
| Memory & State Persistence | Not documented |
|
Coral keeps local usage history and a search over values seen in earlier queries, and its query engine uses them to find tables. Coral describes no memory layer the agent writes and reads as context. Sourcewithcoral.com/docs/guides/use-coral-over-mcpread 2026-09-29 |
|
| Deployment & Data Residency | Full |
|
Open source Coral runs locally or self hosted with Docker or on Kubernetes, and the customer's data, credentials and usage history never leave its machine. Coral hosts the Team plan. Enterprise can run self hosted, fully air gapped or as SaaS. Sourcewithcoral.com/pricingread 2026-09-29 |
|
| Prebuilt Agents, Templates & Packs | Not documented |
|
Bundled and community source specs ship with Coral as connectors. Coral names no prebuilt agents, templates or packs a customer adopts. Sourcewithcoral.com/docsread 2026-09-29 |
|
| Triggers & Channel Coverage | Not documented |
|
Work starts when the customer's agent or script calls Coral over SQL, the CLI or MCP. Coral names no schedule, webhook or event that starts work. Sourcewithcoral.com/docsread 2026-09-29 |
|
| Model Flexibility & Routing | Not documented |
|
The customer's agent brings its own model and calls Coral for data. Coral runs no model of its own, so there is no model to choose. Sourcewithcoral.com/docsread 2026-09-29 |
|
| APIs, SDKs & MCP Extensibility | Partial |
|
A CLI, a SQL interface and Coral's own MCP server drive the product, and custom source specs extend it. The MCP server runs over stdio or a loopback workspace URL, with tokens minted at login. Its tools are list_catalog, describe, list_columns, search and a read only SQL query. The tools read and do not create or change Coral's own objects. No API reference is published. Sourcewithcoral.com/docs/guides/use-coral-over-mcpread 2026-09-29 |
|
| Testing, Debugging & Optimization | Partial |
|
The Team plan includes source connector evals, and Enterprise includes managed evals. Coral does not say what they test or how results are read. Coral also measures retrieval with its own benchmark. Sourcewithcoral.com/pricingread 2026-09-29 |
|
| Browser & Computer Use | Not documented |
|
Sources are read through their APIs, files and MCP servers. Coral describes no browser, desktop or computer control. Sourcewithcoral.com/docsread 2026-09-29 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Product updates
Coral 1.3.0 hands the deadlines for a Run, its approvals and its connections to timers kept by the Run Owner. A sweep now finds Runs whose lease has expired and recovers them, and Runs are drained when a server shuts down. OAuth sign ins can also finish on any Coral process, not only the one that started them.
Bears on: Workflow orchestration
View sourceCoral 1.0 is a rewrite that replaces the old SQL interface with TypeScript programs. Agents write short programs that call services from Coral's setup recipes or any service with an OpenAPI document or an MCP server, and Coral runs them in an isolated sandbox and adds the credentials itself, so agents never hold them. Each operation can be set to allowed, approval required or disabled, and a call that needs approval waits for a person. Version 1.1.0 on October 15 enabled safe HTTP methods by default.
Bears on: MCP / tool calling / API
View sourcePricing
Open source free (Apache 2.0, self hosted); Team $249 a month plus $0.005 per query above 50,000; Enterprise custom
Monthly platform fee plus per query usage above 50,000 queries (Team); custom (Enterprise)
What is public
The open source terms, the Team monthly fee and per query rate above 50,000 queries, and each plan's inclusions; Enterprise is custom.
Cost watchouts
Query volume above 50,000 a month is billed at $0.005 per query on Team, and both paid plans start with a Contact us step.
Variable cost rationale
Team adds a per query charge above 50,000 queries a month; the open source edition runs on the customer's own infrastructure.
Overage / add-ons
$0.005 per query above 50,000 a month on Team
Sales call required
Mixed (some tiers require a call)
Free / trial
Open source edition free forever (Apache 2.0, self hosted)
Lowest paid plan
Team, $249 a month plus $0.005 per query above 50,000
Related vendors
- AgentOps — Agent observability and debugging platform: open source SDKs trace…
- Agno — Python agent framework and AgentOS runtime (formerly Phidata) for…
- AIsa — Resource and payment gateway for AI agents, with one key to 110+…
- AlphaBitCore — AI control plane for regulated financial firms: one gateway enforces…
- Anchor Browser — Cloud hosted browser infrastructure that lets AI agents operate real…
- Apify — Cloud platform and marketplace of more than 73,000 ready-to-run…
Alternatives to Coral
The closest documented capability profiles to Coral among agent infrastructure platforms tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.
- Hamming AI6.5 / 14Adds documented Triggers & Channel Coverage
- Prefactor4.5 / 14Fuller documented coverage on APIs, SDKs & MCP Extensibility
- AgentOps5.0 / 14Adds documented Model Flexibility & Routing
- Lakera7.0 / 14Adds documented Prebuilt Agents, Templates & Packs
- OpenBox AI6.0 / 14Adds documented Prebuilt Agents, Templates & Packs
- Arcade7.5 / 14Adds documented Model Flexibility & Routing
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded