Bigger tasks, without starting over each time

A skill is a small package of instructions, scripts, and reference material an agent loads only when a task calls for it: how to use a specific API, format a document, follow your team's own process. Write it once, and your agent stops needing the same explanation in every prompt. Graph-of-Skills doesn't replace skills. It's what finds the right one once your library has enough of them that two start looking alike.

No. The public catalog, 594 skills harvested from skills.sh, is searched by default alongside your own, so a new account gets results with zero uploads. You can add up to 200 of your own skills on top, searched together with the catalog.

Capability Pages compares each new skill against your existing library when you add it, and resolves any that look alike, such as a file-write tool and a file-append tool. That resolution happens once, at index time, not on every query, so your agent's retrieval call never has to guess between confusable skills.

Charging more as your skill library grows would reward the exact problem this product exists to fix: large, uncurated libraries hurt an agent's accuracy, not skill count on its own. The free tier includes 200 skills and 1,000 retrievals a month, no card required. The paid tier is $19/mo (USD) for 100,000 retrievals a month, and scales with how often you call the retrieve endpoint, not with the size of your library.

The retrieve endpoint returns an error asking you to upgrade once you pass 1,000 retrievals in a calendar month, and the count resets at the start of the next month. Upgrading to the $19/mo (USD) tier raises the limit to 100,000 retrievals a month. Indexing new skills checks a separate, unrelated cap on library size, so adding skills never counts against your retrieval limit.

Every Pro subscription starts with a 30-day guarantee. Email info@akilima.tech within 30 days of subscribing and we refund your first month in full, no questions asked. After that, cancel any time from the billing portal in your dashboard and you won't be charged again.

A plain vector-search tool ranks skills purely by how similar their text embeddings look, which is exactly what breaks when two skills are near-duplicates. Graph-of-Skills adds keyword search on top of that similarity search, so an exact term like an error code or function name still gets found, resolves confusable skills before your agent ever queries, and trims the result set to fit your model's context limit instead of handing back a fixed list.

Use the "Connect your agent" section on this page. In Claude Code it takes two commands: export your key as GOS_API_KEY, then run claude mcp add-json with the server URL and a ${GOS_API_KEY} placeholder, so the key stays in your environment and out of the config file. That registers search_skills, retrieve_skill_bundle and hydrate_skills as tools your agent can call. Any other MCP client that accepts a remote HTTP server with a custom header connects the same way, using the .mcp.json shown there.

Save the "No MCP" file from the "Connect your agent" section as .claude/skills/graph-of-skills-lookup/SKILL.md. It tells your agent to call POST https://www.akilima.tech/api/graph-of-skills/retrieve with curl before a task it may not handle well from memory. Skills are the same format your own library is already in: a name and a short description your agent scans upfront, with the full content loaded only when a task matches. It reads your key from the GOS_API_KEY environment variable.

Yes, by default: we log the query text and the names and scores of the skills it matched, so we can review real retrieval misses and improve accuracy. We never log full skill content. Turn logging off any time from your account settings, and it stops for every future query.

No. Claude Code connects over MCP with one command. Claude Desktop, Cursor, and any other client that speaks MCP connect the same way, pointed at the same URL with your API key. If your agent doesn't speak MCP at all, or runs on a different model entirely (OpenAI, an open-weight model, one you built yourself), it can call the same retrieve endpoint directly over plain HTTP: send your query, get back ranked skills. Nothing here is tied to one vendor or one model.