What your AI canask Edge to do.

Edge gives your agent a small set of tools: find the skill that fits the task, load it for that task, say how it went, report a problem, and check a skill's security. Your agent decides when to call them; you never have to type a tool name.

The local connector (npx @getedge/mcp) lists eight tools. The Claude app connector (https://getedge.cc/mcp) lists seven: everything except benchmark, and its scan_skill checks catalog skills only, because both of those work on your machine. One of the eight, feedback, is a deprecated alias of report_issue.

Search, loading and scanning are annotated read-only. Rating and reporting send information to Edge, so some hosts ask for approval first; if a host blocks them, the agent skips the call and carries on.

The loop: find, use, rate

1. Find. Before specialist work, your agent calls find_skill with the task. Edge searches 254,734 public skills in its index of the skills.sh catalog, withholds skills that fail its security gate, ranks the rest for fit with this task, and returns up to five, best fit first.
2. Use. If one clearly fits, the agent calls use_skill and follows those instructions for this task. Nothing is installed.
3. Rate. Once the task's outcome is known, the agent may call rate_skill: helped, hurt, no_difference or unclear, with one short line on why.
4. Edge learns what helps. A rating counts only for a skill this caller loaded from a search it made, so each outcome is tied to the exact search and load behind it. Counted outcomes show Edge which picks help on which tasks, and they appear on your own Edge page.

Ratings are agent-reported outcomes: the agent's own judgement of how the task went, not human star ratings and not independently verified results. Today they do not change the order of anyone's search results. The connector already carries the controls for a personal nudge, where your own rating could move a skill at most two places in your own next similar search, but that is switched off on Edge's service.

find_skill

What it does. Searches Edge's index of the public skills.sh catalog, checks each candidate against available independent scanner audits, withholds skills that fail the security gate, and ranks the rest for fit with the task. It returns up to five candidates. Choosing none is always a valid answer.

When an agent calls it. Once per task, before work that involves UI or visual polish, documents or slides, images, video or media, data analysis or charts, a security review, or a specific framework or API. The agent compares the results with skills it already has, and keeps a local skill when Edge's best candidate is not clearly better.

What the result looks like

5 candidate skills for "review a pull request for correctness bugs", ranked by Edge for this task.
A good fit for this task: 4 of the top 5. In the catalog index's own order it was 2 of 5.

1. owner/repository@skill-name
   Source: owner/repository
   Skill: skill-name
   What it does: the skill's description, from Edge's index
   Edge moved this up from #9 in the catalog index's order.
   Installs: 1,234
   URL: https://...

2. ...

Next step: call use_skill for the best match before you start the task. ...

Shape of a result, with placeholder values. Real results name real skills.

The first line

How many candidates came back, the query, and how the list was ordered. ranked by Edge for this task means task-fit ranking ran; otherwise the line names the catalog order that was kept.

A good fit for this task

Shown when the ranker returned usable scores: how many of the five it judged a real fit, next to the same count for the catalog's own top five.

Edge moved this up

Appears only on a candidate that ranking promoted, with the position the catalog had given it.

Security note

Appears when fewer than three recent scanner audits are on hand, or when one scanner failed a skill that otherwise passed the gate.

Next step

Names the best match and tells the agent to pass its exact Source and Skill values, plus the search's request_id, to use_skill.

Edge this week

The first result each week can end with a one-line recap of your skill loads, searches where the loaded skill differed from the most-installed pick, and feedback marked helpful.

What you see. Your agent's tool call and, in most hosts, the list above. When nothing audited fits, the result says so and the agent continues without a catalog skill.

Privacy. The query and keywords go to Edge and are stored with your random install ID or Pro key label. When task-fit ranking runs, the query and candidate descriptions also go to the ranking service at api.typesafe.ai. No file contents, paths or conversation history are sent. Keep confidential detail out of the query. Full detail.

use_skill

What it does. Loads one skill that find_skill returned into the current task. Edge fetches the audited bundle at its pinned commit, checks every file against that audit, and returns the instructions. Loading is temporary: nothing is installed and nothing on your machine changes.

When an agent calls it. After a search, when a candidate clearly fits the task and before it starts the work.

What you see. The result starts with Loaded owner/repository@skill for this task only. Nothing was installed. It asks the agent to add one line to its final answer naming what the skill concretely changed, marks the text as a third-party skill that passed Edge's automated audit, and invites one optional rating after the task. The skill's own instructions follow.

Privacy. Besides the usual install, session and client fields, only the skill identifiers and the request_id are sent. The skill text is written by a third party and lands in your agent's context as instructions, so read it, and any commands, before acting on it.

rate_skill

What it does. Records how a loaded skill did on the task, as reported by the agent. This is the step that closes the loop.

When an agent calls it. Optionally, once the outcome is known (tests pass or fail, the result is accepted, the work is done), and only for a skill it actually followed. The agent judges the outcome itself and does not ask you. Bugs and ideas about Edge go to report_issue instead.

What you see. Feedback recorded. Thank you. when the rating counts. If the request_id is unknown, expired, or does not match a skill this session loaded, it is stored but not counted, and the result says why. A second rating for the same load is not added. Your host may ask you to approve the call.

Privacy. Edge stores the outcome, reason, note and rated skill with your install ID or key label. The note is kept with your searches to improve Edge until you ask for deletion. A rating for a search made by a different client is refused.

report_issue

What it does. Sends Edge a bug report or idea about Edge itself: a call that failed or answered wrongly, a suggestion that did not fit, setup trouble, or a feature idea.

When an agent calls it. When you ask it to, or on its own when an Edge call fails. A host that denies approval for a call is not an Edge bug and is not reported.

What you see. A short confirmation. If the connection to Edge is what broke, any agent can still report with a plain request: POST https://getedge.cc/api/feedback with {"kind": "bug", "message": "..."}.

Privacy. Reports are stored in a separate file with the key label, never the key. Anything shaped like a credential is redacted before it is written, and IP addresses are not stored.

feedback is the old name of this tool. It takes the same input and does the same thing, and stays so older setups keep working.

scan_skill

What it does. A security check for one skill, in one of two modes. Each answer says which mode produced it.

• A catalog skill (source and skill as find_skill printed them): Edge checks the exact pinned bundle or reuses a current audit of it, then shows the findings, the audit date, and whether Edge serves or withholds the skill.
• A skill on your machine (local, local connector only): the connector's built-in checks run right there, looking for patterns such as remote code execution, obfuscated payloads, credential theft, hidden instructions and committed secrets. The answer is labelled LOCAL SCAN, gives Result: PASS, WARN or FAIL, says whether anything was uploaded, and lists any findings.

When an agent calls it. When you want to know whether a skill is safe to use or install.

Privacy. A catalog check sends only the skill identifiers. A local scan uploads nothing, unless you agree to the Snyk scan, which sends that skill's files plus this machine's hostname and username to Snyk. Automated checks are not a guarantee: a pass means no check matched.

benchmark

What it does. Measures whether skills actually help on real tasks. It runs each task with each skill and with no skill, using your local Claude Code or Codex, grades every run blindly with the task's grader, and ranks the conditions on one leaderboard: mean, 95% interval, paired lift over no skill, runs, tokens and time.

When an agent calls it. Only when you ask. The first call returns a plan with its cost (tasks times conditions times runs) and a confirmation token; nothing runs until you approve that plan.

What you see. First the plan. After you approve it, the leaderboard and the path to a full local report with outputs and transcripts.

Privacy. Tasks, outputs and transcripts stay on your machine under ~/.config/edge/evals/. Scores, the category, skill IDs, run counts and the agent and model names are shared with Edge; set EDGE_EVAL_SHARE=off to share nothing. Published benchmarks.

Your Edge page

Ask your agent open my Edge page. It calls my_activity and replies with a private link that works once, within 10 minutes. In a terminal, npx @getedge/mcp you opens the same kind of link. There is no password and no permanent link, and you never need your key.

The page shows the last 7 days. At the top: the estimated time saved finding expert guidance (how we estimate it), the tasks your agent reported improved, the average impact your agent rated, the expert skills used and your streak. Then each task your agent used Edge on, the skill it loaded, what your agent says that skill changed when it reported it, and the outcome with a 0 to 10 score. Then your totals since you started and your most-used skills. Task names are short labels your agent writes, never your request; names, emails, links and keys are removed. Free and Pro installs both get it.

Opening the link signs that browser in with a secure, HttpOnly cookie for 30 days. A used or expired link shows one message and never reveals which. The page is not indexed, sends no referrer and runs no third-party scripts. From the page you can sign out every other browser, which also cancels unused links, and clear your activity history, which deletes every task and total after you confirm. Once a week your agent may add ⚡ Edge · your week is ready · ask me "open my Edge page" to a search result; that line never contains a link.

Limits

Search is free and has no limit, and so is task-fit ranking without a key, for now. With a Pro key, task-fit ranking covers 300 ranked searches a day per key, reset at midnight UTC. After that, the day's results come back in catalog order, still screened by the same security gate.

Errors

Failures come back as one-line MCP errors that tell the agent to keep working, for example Edge skill discovery failed: <reason>. Continue the task without Edge. Busy or temporarily unavailable responses are retried twice with a short backoff before the agent sees them.