What leavesyour computer.

Your query text goes to Edge and is stored there with your install id or key label. When task-fit ranking runs, it also goes to an external ranking service, with or without a key. This page says exactly what each tool sends, what the connector reads on your machine, where the data goes, how long Edge keeps it, and how to request deletion.

The short version

• Every search sends your agent's task text (up to 300 characters; searches stored before 2 October 2026 run up to 500) and optional keywords to Edge. Your agent is told to write it generically, the kind of work without names, figures or confidential details, so treat a query the way you would treat a public search box.
• Before storing it, Edge removes anything shaped like a credential, email address, link, file path, phone number or long number. It stores what remains encrypted (AES-256-GCM, with a key kept apart from the logs, archive and backups), with your random install id or Pro key label, and keeps it indefinitely.
• Search is answered from Edge's own index on Edge's server. No search goes to skills.sh or to any catalog analytics.
• When task-fit ranking runs, your query and candidate descriptions also go to api.typesafe.ai, with or without a key.
• Search includes your installed skills' names and descriptions only if you set EDGE_LOCAL_SKILLS=1. Search does not send their file bodies. A local scan_skill reads the selected skill's files on your machine; a separate Snyk scan can send those files to Snyk as described below.
• use_skill puts third-party catalog instructions in your agent's context. Edge withholds skills that fail its security gate after available provider reports and its own exact-bundle checks, but unverified skills require an exact-bundle scan before load, and the checks do not verify every instruction.

What the connector sends

The @getedge/mcp package is a protocol connector. Per tool call it sends one HTTPS POST to https://getedge.cc/api/mcp/<tool> with a JSON body. Every body carries these fields:

On top of that, per tool:

report_issue and feedback are aliases for the same report flow. scan_skill can look up a catalog audit or check a local skill offline; an optional Snyk scan shares local skill files, hostname, and username only when share_with_snyk=true is explicitly set. The stdio-only benchmark tool runs a local evaluation and shares only its scores. Your agent chooses what task text and reports to send.

What the connector reads and writes on your machine

• Environment: EDGE_KEY, EDGE_LOCAL_SKILLS, EDGE_INSTALL_STATE_DIR, EDGE_SKILL_PATH, CLAUDE_CONFIG_DIR, CODEX_HOME, EDGE_TELEMETRY, and EDGE_BACKEND or EDGE_MCP_API_URL, which point it at a different backend for development.
• ~/.config/edge/: it creates install-id and, after the first successful hello, hello-sent. It keeps the last day it sent the ready signal in ready-day. The prompt hook leaves an empty nudge-day-YYYY-MM-DD file on the first nudge of a day; the connector sends it and renames it to .sent. The setup flow records your answer about local skills in local-consent here.
• Edge setup skill: it can read the first 400 characters of ~/.claude/skills/edge/SKILL.md or ~/.codex/skills/edge/SKILL.md to find skill_version. The connector does not read CLAUDE.md or AGENTS.md to compute a routing flag.
• Setup markers: after the first successful search it deletes skills/edge/FIRST-TASK-PENDING under ~/.claude and ~/.codex, if present.
• Installed skills, before opt-in: with a Pro key it checks only whether any SKILL.md exists in your skill folders, to decide whether to ask once. It reads no skill text.
• Installed skills, after EDGE_LOCAL_SKILLS=1: it reads the first 4,000 characters of each SKILL.md in .claude, .codex, .opencode and .agents skill folders, from your project up to its Git root and in your home directory, and keeps only the frontmatter name and description. Never the body, never scripts.

The remote connector, getedge.cc/mcp

Hosts that add Edge as a remote MCP server can use six tools on Edge's server: search, load, rate, report_issue, feedback (a compatibility alias), and scan. They send the same tool arguments, with client set to remote-mcp and the MCP session id as session. There is no install id, no local skills, and nothing is read from your machine. The exception is the Claude Code plugin after Edge setup: it sends the install id setup created as an X-Edge-Install header, so its searches are stored, shown and deleted with that install. ChatGPT sends an anonymous user id (openai/subject) with each tool call; Edge never stores it and keeps only a keyed one-way hash of it as the install id, so one ChatGPT user's searches across conversations are counted, shown and deleted as one install. A Pro key arrives as an Authorization: Bearer header or a ?key= URL parameter; prefer the header, because the URL is written to the web server's access log.

Where the query goes after that

• Cloudflaregetedge.cc is served through Cloudflare, which terminates HTTPS in front of Edge's server. Every request, body included, passes through it.
• Edge, getedge.ccLooks up the task sentence and the keywords in Edge's own skill index, a search service on the same server that encodes the query locally. There is no live call to skills.sh per search, no catalog CLI, and no third-party analytics beacon with your query. Candidate descriptions come from the index itself.
• api.typesafe.ai, task-fit rankingThe ranking service (model jev-latest). One request per search, authorized with Edge's own service key, not yours. It receives your task sentence (not the keywords) and, for each candidate, its source@skill and description: the 20 catalog candidates plus any of Edge's own skills added for the task, and on opt-in your installed skills' names and descriptions. A key is not required for catalog ranking. Installed skill names and descriptions are included only after local-skills opt-in.
• PostHog, usage analyticsEdge's server sends PostHog (US) one usage event per tool call: which tool ran, the skill, the outcome and timings, client, version and referral code, under one-way hashes of your install id, session and key label. No query, keywords, task text, rating notes, IP address or key. A deletion request covers these events too.
• GitHub, on use_skillLoading a skill makes Edge's server fetch the audited bundle at its pinned commit from where it is published, usually raw.githubusercontent.com. That request comes from Edge's server and carries no part of your query.

What Edge stores, and for how long

Retention. The event log rotates daily. Each day stays on the service for 31 days, and before it leaves, a compressed copy goes to an archive that is never pruned; an encrypted copy of that archive is mirrored nightly to offsite backup. So search text, keywords and rating notes are kept indefinitely, tied to your install id or key label, because Edge uses them to improve search. Nothing is deleted on a timer; data is deleted when you ask.

Key label. The key label is the first 8 hex characters of your key's SHA-256. Keys are stored only as SHA-256 hashes, and the key itself is never written to the event log or feedback file.

Other logs on the server. The events contain no IP address. The web server in front of Edge keeps an access log of every request, with your IP address, time, path and user agent, for 60 days; connector requests carry no query in the path. The index service logs only the path and status of each lookup, not the query. Entries written before 26 September 2026 still contain the query or keywords until the system journal rotates.

Referral codes. When your setup link carries a referral code (?ref=, a public creator or partner code), Edge stores it as ref on your events and in a ledger row with your install id (as came_from when it was matched by network instead of sent by setup), and to make that match the ledger keeps the code with an HMAC of your IP address under a salt that changes daily and is deleted after two days, never the IP itself; forget deletes these rows with the rest.

Request ids. The search's request id is printed in the result and passed back on use_skill and rate_skill, so a load and a rating can be tied to the search that produced them. That link is held in memory for 12 hours.

What comes back in

use_skill returns a document written by whoever published that skill, and it lands in your agent's context as instructions. This is the part worth being deliberate about.

• Edge does not review these skills. The catalog is a public, third-party ecosystem. Anyone can publish to it. Edge withholds skills that fail its security gate, and every load carries an advisory, but that is not a review of what the text tells your agent to do.
• The text is written to be followed. Skills are instructions for an agent, and they can name commands, scripts and tools. Read what came back before acting on it, and do not let an agent run a skill's scripts unexamined.
• Loading is temporary. The retrieval happens on Edge's server, the text is returned over the same HTTPS call, and nothing is written to your disk. No skill ends up permanently in your agent's skill directory through Edge.

Output is stripped of terminal control codes and capped at 2 MB, beyond which it is cut and marked truncated.

Edge is published by Floom, Inc. For support, email fede@getedge.cc.

Delete and opt out

• Stop sending: remove the server, claude mcp remove edge or codex mcp remove edge, or delete the edge entry from your host's MCP config. That also removes EDGE_KEY and EDGE_LOCAL_SKILLS.
• Stop sending installed skills only: set EDGE_LOCAL_SKILLS=0 or remove it, and restart the connector.
• Delete local state: rm -rf ~/.config/edge. This removes the install id on your machine, not the records already on Edge's server.
• Delete it yourself: npx -y @getedge/mcp@0.6.56 forget deletes this install’s searches, loads, ratings and reports (with EDGE_KEY set, everything under that key too) from the service, the archive and its offsite copy. It prints the counts by type and the date the last encrypted daily backup expires (31 days). A deletion record keeps only one-way hashes, so a restore from backup deletes the data again.
• Request review or deletion: email fede@getedge.cc with your install id (cat ~/.config/edge/install-id, before you delete it) or your key label (printf %s "$EDGE_KEY" | sha256sum | cut -c1-8). Never send the key itself. We delete the matching records, including the archive and its offsite copy, within 30 days of your request. A key label is a short hash prefix and may not identify every record. Remote connector searches carry no install id (the Claude Code plugin’s do after setup), so without one or a key they cannot be matched to you. In ChatGPT, ask your agent to delete your Edge data instead: its forget_my_data tool deletes everything stored under that id.

What Edge never does

• Install a skill, or call the catalog's installer, on your machine.
• Send your project's files during search or skill loading. A local scan reads the selected skill's files, which may be inside your project, and the optional Snyk scan can upload them as described above.
• Write outside ~/.config/edge, apart from deleting the setup marker above. Launching it with npx does put the package itself in your npm cache, the same as any npm package.
• Require a login. There is no account or password on either plan; a Pro key is checked against its hash. Your Edge page opens from a one-time link your agent gives you and sets a 30-day sign-in cookie for that browser.

The practical rule

Keep confidential detail out of the query. A query naming a client, an internal system, an unannounced product or a specific vulnerability is a query you have handed to Edge's stored logs, and, when task-fit ranking runs, to a ranking service as well. Describing the kind of work rather than the subject of it gets the same candidates back: review a pull request for correctness bugs finds exactly what review the Acme billing PR would.