Connect an agent

You need Node.js 22 or newer and a key from a Profile with access to at least one indexed document. If you do not have those yet, follow Your first cited answer.

1. Download takibi.mjs and save it as tools/takibi/takibi.mjs in your working directory. The CLI is a single file with no package dependencies.
2. Run it with Node, or create an alias using its absolute path:

alias takibi='node /absolute/path/to/tools/takibi/takibi.mjs'
takibi --help

Download the Takibi agent skill into your tool's skills directory, or give its instructions to your agent. Keep the CLI at the path above if using the skill without an alias. Tools may run in different working directories, so check the path from the agent's environment too.

The CLI talks to an API on your own machine unless told otherwise. If your workspace lives on Takibi's servers, set the endpoint in the environment where the agent runs:

export TAKIBI_BASE_URL="https://app.takibibase.com"
mkdir -p ~/.takibi
touch ~/.takibi/key
chmod 600 ~/.takibi/key

Use an editor to save the complete publicId.secret key as the only line in ~/.takibi/key. Do not paste the secret into shell history or chat. Set TAKIBI_KEY_FILE if you store it elsewhere. The account running the agent must be able to read that file.

Run takibi version to check the endpoint. This probe does not need a key, so a successful version response does not prove the key or grants are valid. Test those with a permitted search or question next.

If the Profile has one project, the CLI can select it automatically. Otherwise pass --project <project-uuid> for the project you want to use. You can save a local name with takibi projects --add <name> <uuid>.

A local project name is a convenience, not an access grant. If the project is unavailable, check the Profile's Access settings and that you copied the correct project ID.

takibi ask -q "What is our refund policy?" --json
takibi search -q "refund eligibility" --json
takibi doc text <document-uuid>

Ask returns exact source passages and citations. Search returns ranked snippets so you can find relevant material. Doc text reads the permitted converted text for a document.

spans and citations

The retrieved passages and their document/chunk references. Check the source context before drawing conclusions.

support and supportVersion

How strongly the passages back the question, scored by the version named in supportVersion. It is not a probability that a claim is true.

answerability

Can be answerable, partial, unanswerable or unknown. Partial evidence may cover only part of the question.

conflict

Flags detected contradictions. A false flag does not prove that no contradiction exists.

abstained

When true, no passages were returned. This is a successful response with no evidence, not a transport failure.

If no passages come back, check the Profile's access and document status, try a narrower question, or search for a specific phrase. Do not treat missing evidence as a negative answer. If the AI ranking step is unavailable, answers fall back to plain keyword matching. Search keeps working either way.

Have the agent cite the returned documents and distinguish source evidence from its own interpretation. Give it explicit approval before task-changing actions, then follow the task workflow.

For failed requests, see error diagnosis. For a direct integration, use the API reference and its static OpenAPI download.