Tools

The Model Context Protocol is an open standard for exposing tools to a language model. Register a server once and every tool it publishes becomes available to any agent you bind it to — MongoDB, Playwright, design tools, internal services, anything that speaks the protocol.

For database servers you can define a constrained tool rather than exposing raw query access. You pick the tool type — Query Tool, Aggregation Tool or both — list exactly which collections are reachable and which fields within them are allowed, and then set hard constraints:

• maxResultLimit — ceiling on rows returned
• allowSorting — whether sorts are permitted
• allowProjectionOverride — whether the model may change the projection
• readOnlyEnforcement — reject anything that writes

Several things happen between "the agent has an MCP server bound" and "a tool actually ran":

A Knowledge Base turns your documents into something an agent can search by meaning rather than by keyword. Upload files or point it at the web; the platform extracts text, splits it, embeds it and indexes it. At query time the question is embedded too, and the closest passages are injected into the prompt — Retrieval-Augmented Generation.

Connection strings and other sensitive fields are encrypted at rest on the knowledge-base record rather than stored in the clear.

Every file goes through the same pipeline: detect the format, extract text, chunk it, embed the chunks, store and index them. Each stage is more careful than it first appears.

Three storage backends are supported, and the search path checks the index before querying it.

• Atlas Vector Search — classic per-collection knowledge bases
• Atlas V2 — a single default collection and index
• Chroma — self-hosted collections
• Index status: READY, PENDING, BUILDING or MISSING
• Candidate pool scaled from topK with a floor
• Score threshold plus topK capping
• Single and multi-collection search paths
• Auto-embed or pre-computed query vectors

Content can also come from the open web. Five fetch strategies are available — plain fetch, an HTTP client, two headless browsers, and a naive full-body extractor as a baseline — because a static page, a JavaScript-rendered app and a listing page each need different handling. Readable-content extraction falls back to naive body text when a page has no article structure, and content is capped with truncation flagged so the interface can say so.

Feed URLs get special treatment: Google News and similar redirect wrappers are resolved to the real article URL first, so the crawler fetches the destination rather than the redirect shell.

The API Registry is the lightest of the four: register a REST endpoint and it becomes callable. No server to run, no protocol to implement — a URL, a method, headers, query parameters and an optional payload template.

Any part of the URL, query or payload can contain {{token}} placeholders. Token names are restricted to letters, digits, underscore and dot, so stray braces in a payload never match by accident.

At call time the tokens are collected and one cheap model call extracts their values from the user's message. The extraction prompt is deliberately strict: return only compact JSON, and omit any field whose value is not clearly present. Unresolved tokens are left intact in the string rather than blanked, which is what lets the caller detect a missing required value and skip the endpoint entirely instead of sending a half-filled request.

Channels are the write surface. A channel registration binds a platform account to one or more agents; messages arriving on that account are answered by a tagged agent, and the platform's actions are exposed to the model as tools. A channel is unique per platform and account id, so the same number or mailbox cannot be registered twice.

Agents are attached and detached without touching the agent itself, and a short platform cache keeps repeated tool binds from hitting the database on every turn.

Phone numbers are sanitised before sending — spaces, dashes and brackets stripped, a leading + removed — and anything suspicious is flagged with a specific warning: a leading zero suggesting a missing country code, too few digits, or leftover non-digit characters.

Connection uses a browser OAuth flow with PKCE, a callback secured by state, stored tokens and automatic refresh. Status and disconnect are available without touching the agent.

Anything arriving from the internet is treated as untrusted:

• HMAC SHA-256 signature verification per channel secret
• Timing-safe comparison, never a plain equality check
• Missing signature or missing secret both fail closed
• Duplicate message suppression on a one-hour TTL set
• The TTL set is swept on a timer so it cannot grow unbounded
• Delivery status tracking — sent, delivered, read, failed
• Human-readable diagnosis for platform error codes
• Inbound media resolved into agent attachments