Works with Claude Code, Cursor, Codex, and any MCP agent
When your AI agent doesn't know, let it ask a human.
Askance plugs into Claude Code, Cursor, Codex, or any MCP-capable agent so it can escalate the questions only a person can answer — straight to the right expert, answered in a tap, without ever pausing your build.
That's the whole integration. Your key comes from Settings once you've signed in.
Questions carry real code and real internal detail, so they're encrypted in transit and at rest, visible only to the members of that project, and never written to our logs.
Who it's for
Three people, one awkward gap
The question an agent can't answer is rarely hard. It's just held by someone who isn't in the room.
If you build with an agent
Don't lose an afternoon to one question
You're four prompts deep and your agent needs to know whether the discount stacks with loyalty points. That answer isn't in the codebase — it's in someone's head. Ask them without leaving the session, keep building, and pick the answer up when it lands.
If you're building your own internal tools
Loop in the colleague who actually knows
The rule you need lives with one person in finance, not in a ticket queue or a sprint plan. Invite them by email, ask the question, get the answer from the source — no project-management process for what is genuinely a two-minute question.
If you're the one being asked
Answer in one tap, ignore the rest
You only get pinged on the topics you've said you know. Most answers are one tap on a recommended option, from your phone, in the queue at the café. No new tool to live in, no channel to watch, no standup.
For the developer
Your build doesn't wait for an answer
The question goes out with its context and the option your agent recommends — then your agent moves on to the next task. When the answer lands, an hour later or tomorrow, it picks up exactly where it left off.
How it works
Four steps, none of which block your build
01
Give your agent a new ability: asking.
Askance is an MCP server. One line of config connects it to your agent, and from then on 'ask a human' is just another tool it can reach for — like reading a file or running a test.
02
It checks what's already been answered.
Before it interrupts anyone, your agent searches every answer your project has ever given. Often the question has already been asked, and it just carries on.
03
It asks properly, then gets back to work.
If the question is new, your agent writes it up with the context, the options it can see, and the one it recommends — then returns to other tasks. Nothing blocks. Nothing waits on a human being awake.
04
A human answers; the answer sticks.
The experts on that topic get a notification, open a link, and confirm the recommended option — or write more, if it matters. Your agent picks the answer up and continues. The exchange is kept forever, searchable by every person and every agent on the project.
For the expert
One tap, from wherever you are
You're only asked about the topics you've said you know. The question arrives already written up, with a recommended answer to confirm — so most replies take a single tap, on your phone, between other things.
Backlog
And your backlog lives here too
Your agent maintains the feature backlog through the same connection; you steer it with votes and comments. One fewer tool that only exists to hold a list.
Give your agent someone to ask.
Connect it once, invite the people who hold the answers, and stop losing afternoons to questions that take someone else two minutes.