docs / guides / claude & chatgpt

Claude & ChatGPT.

The chat connector is the whole team's door: nothing to install, and you ask in plain language.

Connect it

Add Balladeer as a connector using the URL from your setup page's "Get connected" step (chat tab). First use sends you to the browser to sign in, pick your workspace, and allow it. Leave the connector's OAuth Client ID and Client Secret fields blank: signing in handles authentication.

Ask, record, rule

• Ask: "What did we decide about order sync, and why?" "Why did we reject batch polling?" "What did we decide this week?" "Where does the payments initiative stand?" Answers are decisions, not search results, with their standing, and receipts on request. When what you propose contradicts a confirmed decision, it names the conflict and its reasoning rather than going along. Progress questions read the living roadmap's rollup: what merged work says is done and still open, never a typed status. That answer comes back as grouped work, each group named for what having it gives you, rather than as a reformatted copy of the rollup.
• Record: tell it a decision that got made elsewhere ("we decided pilots get 30 days, because...") and it records the call, unconfirmed, for a ruling. The why is written in your own words where you gave them, and anything it inferred is marked as inference rather than put in your mouth. If the decision entails work, it drafts the definition of done from what you already settled and shows you that draft to edit a line or wave through, instead of interviewing you for it; how much it drafts follows how settled the decision is, so a still-fuzzy direction has its open part named rather than filled with invented milestones. A policy or a "we won't do X" carries none at all. Say "just record it" or "walk me through the criteria" and it shifts at once.
• Rule: "confirm that one", "dismiss those two", "yes, this can ship", "that's stale." It executes your verdicts (confirm, dismiss, and acceptance, the ship-gate sign-off that a decision's work is ready to ship) and records your reactions. It never rules on its own initiative; a verdict you hand it is carried out, including on a decision it proposed itself. Dismissal is for a decision that was wrong from the start, which takes it out of retrieval; one that was true and has since been replaced is recorded as superseded, so it stays findable as history.

How answers come back

Decision ids, tool names, and field names stay out of the prose: a decision is named by its claim in a few words, and an id appears only when you are being asked for a verdict on that decision. Work is named for what having it gives you, both for a single decision and for a group of them, and counts cover what needs you rather than the length of a list you are reading.

When it puts decisions to you for a ruling, it opens with the count and the ask in one line (how many need you, and what saying yes makes binding), then tells the story of the need each one serves, written for someone who does not remember the conversation that produced it.

If a tool call fails on approval

A failure that mentions approval ("No approval received") is your chat client's own tool-approval step, not a Balladeer refusal. Balladeer never interrupts a call to ask for approval, and its own refusals name what refused, who can unblock it, and whether retrying can work. Approve the tool for the connector, or disconnect and reconnect the connector so its tool list refreshes.

Chat obeys the same workspace visibility as every surface: a restricted decision simply does not exist for anyone outside its audience.

Verified from live product as of 28 jul 2026.