.env file is required for a native connector.
this is a different setup path from set up API with your AI agent, which wires the REST API into a local codebase. use MCP when you want edgeful available directly inside your chat client; use the API path when you’re building something that calls edgeful from your own code.
https://api.edgeful.com/mcp
Claude (Desktop or Web)
1
add the connector
in Claude, go to Settings → Connectors → Add custom connector.
2
point it at edgeful
name it
edgeful and set the URL to https://api.edgeful.com/mcp, then save.3
connect
click Connect. a browser window opens an edgeful sign-in and consent screen, not a bare API key prompt. sign in with your edgeful account and click Allow.
4
ask a question
back in Claude, the edgeful tools are available in the conversation. try something like “what are the top reports for ES right now?” and watch it make a tool call.
Codex (via ChatGPT desktop app) / Codex CLI
1
add the server
app (ChatGPT desktop): open ⚙️ Settings → Plugins (under Integrations) → Add → Add MCP server. in the “Connect to a custom MCP” form, name it
edgeful, set type to Streamable HTTP, and set the URL to https://api.edgeful.com/mcp. leave bearer token and headers empty, then save.CLI: the same config is shared through ~/.codex/config.toml, so entries added in the app show up there too.2
authenticate
the server shows up flagged as requiring OAuth. click Authenticate, sign in with your edgeful account in the browser tab that opens, and click Allow. in the CLI, run
codex mcp login edgeful if it asks you to reauth.3
use it
restart if Codex asks you to, then the edgeful tools are available in a thread.
/mcp in the composer lists active servers.clients without a native OAuth connector (Cursor, older setups)
some MCP clients can’t run the OAuth handshake yet and only support a static bearer token. for those, keep using an edgeful API key overmcp-remote, the same setup this replaces for Claude and Codex:
- generate a key in the API dashboard → API Keys (see authentication for the full key format and rotation flow).
- add this to your client’s MCP config:
mcp config
- set
EDGEFUL_API_KEYin your environment (never paste the key directly into a shared config file).
plan access
MCP calls are entitlement-gated the same way the REST API is: your edgeful account needs an active API-enabled plan (Essential, Pro, or All Access). signing in works for any account, but a tool call from an account without an API-enabled plan returns amissing_entitlement error. see authentication for the full plan breakdown.
build and save custom reports
writes are on by default, and one permission covers custom reports and account settings. in the API dashboard, under let AI agents make changes in your account, allow changes through the MCP connector covers AI apps signed in to your account and allow changes through your API key covers a client that uses your key. turn either off to keep that credential read-only. new and regenerated keys start with writes on. then ask Claude to build a report, review its methodology and sample, and tell it to save. the report appears under my custom reports in Edgeful AI, marked built via API. you can complete the whole conversation in Claude. Edgeful generates and validates the report code. builds and revisions use your Edgeful AI allowance. runs use the saved code without a model call. turning writes off keeps reading available. see build and save a custom report for the complete workflow and limits.manage your settings
with the same permission, Claude can bookmark reports, edit your watchlists, change your account defaults, save custom sessions, change report customizations and save, list and delete templates. the app shows each change after reload; an Edgeful tab opened before the change can overwrite bookmark and watchlist edits the next time it saves, so reload it first. a report customization is stored where the report page reads it, andcall_report_endpoint applies it only when you pass the query the customization tools return for that page. a template saved by Claude appears in its picker (the report sidebar, the What’s in Play page or the screener) after reload (or the app’s five-minute template refetch) and is applied from there; What’s in Play and screener templates are composed from the vocabulary list_report_catalog lists for them. see manage your settings with Claude for the tools, validation rules and budgets.
submit product feedback
the server asks agents to proactively submit feedback when an edgeful limitation blocks the user’s task, while respecting user instructions and client permissions. whether an agent follows this guidance depends on its model and MCP client. ask your agent to usesubmit_feedback for a bug, feature request, or confusing behavior.
provide a description and category (bug, feature_request, or confusing_behavior),
with the relevant tool name and expected/actual behavior when available. descriptions
and behavior fields allow up to 4,000 characters each; tool names allow 100.
feedback uses the same write permission as settings and custom reports. submissions
are saved for manual review and return a feedback ID with status new. each call
creates a separate submission; avoid repeating feedback you already submitted.
the limit is 10 submissions per minute and 100 per hour per member, across credentials.
leave out credentials, full conversations, and private trading data.