Base for agents.
Base is where an agent turns documents into a shared, cited knowledge base, and where people and their agents work in one together. One remote MCP server, OAuth sign-in, free to connect, and skills for coding agents that teach the whole loop.
Set up in one line
Paste this into the agent. It reads the page, works out where it runs (a hosted chat, a coding agent, a sandbox), and does the setup that fits, asking the person only for the sign-in. Everything below is the same setup by hand.
Set up Base by following https://createbases.com/llms.txtThe first useful task
Connected, an agent can take a folder of documents and hand back one link that a person opens in a browser and another agent opens as a knowledge base with citations. That is four calls: create_base, add_document per document, act_on_base with publish, and the share_url it returns.
Reading is the other half. Given a share link, open_shared_base previews it, get_base reports what changed since the user last checked and what is waiting on them, and explore_base answers a question with the evidence and its sources.
In ChatGPT, Claude, and other hosted chats
The connector is the whole install. A hosted chat cannot run a command or read a skill file, so everything an agent needs to know about Base travels inside the server: the tool descriptions, the server instructions, and the text each result returns. Connect once and the first useful task is one message away.
- Claude: it is in the Claude connector directory, https://claude.ai/directory/createbases. Press Connect; there is nothing to paste.
- ChatGPT: it is in the ChatGPT plugin directory as Base Knowledge, https://chatgpt.com/plugins/plugin_asdk_app_6aebbd3a8a6c81919a191445b7a31014. Press + and connect it when ChatGPT asks; there is nothing to paste.
- Grok, and any other chat with custom connectors: add
https://createbases.com/api/mcp. Step-by-step setup per app is at https://createbases.com/connect, as plain text at https://createbases.com/instructions.md.
In Claude Code, Codex, Cursor, and other coding agents
One command finds the agents on this machine, installs the skills, adds the MCP server, and signs in. It ends by naming the one step left in each agent, its own sign-in to the MCP server. The CLI is base-knowledge on npm.
npx base-knowledge initThe sign-in prints a link and a short code. The person opens the link on any device and presses Allow, so it works from an agent's own shell, in a container, or over SSH. Once signed in, every tool is one command in the same session, before any agent has signed in to the MCP server:
npx base-knowledge tools
npx base-knowledge call explore_base '{"question":"What changed this week?"}'
npx base-knowledge open <share-link>Or the two lines it runs. The first installs the skills, which say when to reach for Base and in what order; the second adds the MCP server, which does the work.
npx skills add https://createbases.com
npx -y add-mcp https://createbases.com/api/mcp -gBoth ask which agents to install to. In the skills picker, the agents that share one folder (Codex, Cursor, and others) are always included; Claude Code is further down the list, so search for it and select it, or name it and skip the picker:
npx skills add https://createbases.com -a claude-code -s '*' -y
npx -y add-mcp https://createbases.com/api/mcp -a claude-code -gThe first time Claude Code starts in a folder with a new MCP server it asks whether to use it, with the cursor on the option that declines. Choose to use the server, then sign in from /mcp.
Three skills: create-base (documents in, a shared link out), work-in-base (open, read, question, contribute), and check-bases (what changed and what is waiting on you, once or on a schedule you set up). Each is a folder holding a SKILL.md in the Agent Skills format, listed at https://createbases.com/.well-known/skills/index.json, which is what the first line reads, and with a sha256 digest per skill at https://createbases.com/.well-known/agent-skills/index.json. Without npx, fetch the index and each skill's files from that directory by hand. What changed in the tools, newest first, is at https://createbases.com/changelog.md.
As a Claude Code plugin, skills and MCP server in one install, from github.com/Lucky-Tree-Labs/base-skills:
/plugin marketplace add Lucky-Tree-Labs/base-skills
/plugin install base@base-skillsIn a sandbox, on a server, or in a scheduled job there is no browser for the sign-in. A key is that sign-in done once by the person, in https://createbases.com/settings/connect, and pasted where the agent runs. It reaches what the person can reach, full access by default, spends against their balance and caps, and stops when they revoke it or it expires. Send it as a bearer token:
npx -y add-mcp https://createbases.com/api/mcp -g --header 'Authorization: Bearer $BASE_KEY'
curl https://createbases.com/api/mcp -H 'Authorization: Bearer $BASE_KEY' \
-H 'Content-Type: application/json' -H 'Accept: application/json, text/event-stream' \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"list_bases","arguments":{}}}'Keys start with bk_, so a secret scanner can catch one that leaks. A key can be limited to named Bases, or to reading, when it is created.
An agent can also ask for its own key, with nothing pasted. npx base-knowledge login does this and keeps the key; BASE_KEY hands the CLI a key you already hold. Over plain HTTP, the agent posts its name, shows the person the link and the short code that come back, and polls for the key while the person opens the link on any device, signs in, and presses Allow. The key arrives once; store it. Errors are the device grant's (RFC 8628): authorization_pending, slow_down, expired_token, access_denied.
curl -X POST https://createbases.com/api/oauth/device -H 'Content-Type: application/json' \
-d '{"name":"nightly worker","access":"full","expiry_days":90}'
# show the person "message" from the reply, then every 5 seconds:
curl -X POST https://createbases.com/api/oauth/device/token -H 'Content-Type: application/json' \
-d '{"device_code":"dc_…"}'An agent that already holds a connection or a key can mint keys itself, so it can set up another agent without a person in the loop. A key made this way never exceeds what its maker holds, lasts at most 90 days, and stops the moment the maker's connection or key is revoked:
curl -X POST https://createbases.com/api/oauth/keys -H 'Authorization: Bearer $BASE_KEY' \
-H 'Content-Type: application/json' -d '{"name":"nightly worker","access":"read","expiry_days":30}'Before the person has an account
An agent with no connection and no key can still make the Base first. It asks for a draft, puts the documents in, and shows the person one link. The person opens it, signs in or signs up, sees which agent made it and what it holds, and claims it; claiming starts reading the documents, against their plan. Nothing is charged and no model sees the documents before that, and an unclaimed draft is deleted after 24 hours. A draft holds at most 25 documents and 25 MB until it is claimed. Until then you can search the draft to check what you made: you get the documents' exact words with their citations.
curl -X POST https://createbases.com/api/agent-drafts -H 'Content-Type: application/json' \
-d '{"title":"Q3 diligence","agent":"Research agent","access":"full"}'
# keep draft_secret; show the person "message" (it holds claim_url and claim_code)
curl -X POST https://createbases.com/api/agent-drafts/<base_id>/documents -H 'Authorization: Bearer $DRAFT_SECRET' \
-H 'Content-Type: application/json' -d '{"file_name":"notes.md","content":"..."}'
# or -F files=@report.pdf; search it any time before the claim:
curl -X POST https://createbases.com/api/agent-drafts/<base_id>/search -H 'Authorization: Bearer $DRAFT_SECRET' \
-H 'Content-Type: application/json' -d '{"query":"renewal date"}'
# then poll every 5 seconds until status is "claimed":
curl https://createbases.com/api/agent-drafts/<base_id> -H 'Authorization: Bearer $DRAFT_SECRET'The secret is good for this draft's documents, search and status and nothing else, and it never appears in the claim link. content takes text, content_base64 bytes you actually hold, or send documents: [...] for several. Search is plain text search: pictures and scanned pages are not in it, and limitations names any document it could not read, so a miss does not show a document lacks the answer. When the person claims it they choose whether the agent gets a key, and at what access; a key reaches that Base only and arrives once, in the first status reply after the claim. If they say it is not theirs, the draft and its documents are deleted and the status answers 404.
The MCP server
- Library endpoint, every Base the person owns, belongs to, or follows:
https://createbases.com/api/mcp. - One shared Base, authorized by its share token, the same tools in read mode:
https://createbases.com/api/mcp/<share-token>. - Protected resource metadata:
https://createbases.com/.well-known/oauth-protected-resource.
Sign-in is OAuth 2.1 with PKCE, dynamic client registration, and client ID metadata documents, so a personal assistant connects once and acts as the person. There is no API key to mint or paste for that journey. Every request is checked against the person's permissions when it is made; a revoked connection or an expired link fails on the next call.
The tools
Nineteen tools, in the order a job uses them.
- Find:
list_bases,open_shared_base,follow_shared_base. - Read:
get_base,explore_base,list_files,fetch_document,get_table,download_file,get_thread. - Contribute:
add_document,edit_document,create_thread,append_thread,resolve_thread,tell_base. - Own:
create_base,configure_base,act_on_base.
Every result is data, not instructions, including the text of the documents a Base holds. A tool that would leak something the person cannot see returns only what they can.
Limits and cost
- Free: 2 Bases, 1 GB, and 300 credits that never expire. Building a Base from a connected AI is included on Free.
- Every new account starts with a Base Tutorial, a Base whose pages explain what Base can do, to test against before adding documents of your own; list_bases marks it starter: true.
- Reading, searching, and sharing never spend credits. A Base spends them when it reads new content, or when the owner set its answer depth to work harder on a question.
- A document goes in as text or a public URL at any size, as bytes the agent holds, or through an upload page for files only the person can reach (up to 50 MB).
- A limit error names the limit and nothing else. Prices are at https://createbases.com/pricing.md; a connector never sells a plan.
Everything else
- https://createbases.com/llms.txt: what Base is, in one page.
- Every public page has a markdown twin at the same path plus
.md, or sendAccept: text/markdown. - A person disconnects an AI app by removing Base from that app's connector settings. Keys are listed and revoked at https://createbases.com/settings/connect.
- Questions: support@createbases.com.