Knowledge infrastructure for AI agents

The knowledge base your AI agents can cite.

Distill for Agents turns your company's documents into a governed source of truth. Connect any agent over MCP: every answer is grounded in your documents and cited by construction — and when the documents don't contain the answer, your agent says so instead of guessing.

What is our refund window for annual plans?

Annual plans can be refunded in full within 30 days of purchase[1]. After 30 days, refunds are prorated to the unused months and require finance approval[2].

[1] Billing policy v4 · §2.3 Refunds [2] Finance handbook · Approvals
Answered via mcp · support-bot 1.8s

Cited by construction

An answer without a source never leaves the system. Every response names the exact documents it came from, so anyone — human or agent — can check the work.

Honest abstention

When your documents don't contain the answer, that's the answer. No guesses delivered in your company's voice — an out-of-scope question gets an explicit "the documents don't cover this."

Proof, not vibes

Pin real questions as eval cases and re-run the suite as documents change. A scorecard shows what passed, what failed, and why — so "the bot is right" is something you measure, not hope.

How it works

From documents to provable answers in three steps.

[1]

Import your documents

Upload PDFs, Markdown, text, or point Distill at your help-center URLs. Each document is indexed and versioned — replace a policy and the old version leaves retrieval instantly.

[2]

Connect any agent over MCP

Mint a scoped API key and point your agent at the MCP endpoint — kb_search, kb_fetch, kb_ask. Works with any MCP client; plain REST if you prefer. Revoke a key and that agent is out, immediately.

[3]

Prove it stays right

Every question is logged with its answer, sources, and which agent asked. Turn any of them into an eval case, then run the suite whenever the corpus changes.

Bring your own agent

We're the data layer. The agent is yours.

Distill for Agents doesn't ship a chatbot, and doesn't want to be one. It's the layer underneath: a governed, versioned corpus with one job — giving whatever agent you already run answers it can cite. A Slack bot, a support widget, an IDE assistant: any MCP client connects in minutes, and anything else can use plain REST.

Example: an internal Slack bot with knowledge you control

  1. [1]

    Create a project and import the documents your bot may speak from — the HR handbook, the security policy, the runbooks. That set is the whole universe it can cite.

  2. [2]

    Mint a key named for the bot — slack-hr-bot — and point its agent harness at the MCP endpoint. Slack's agent framework, a LangChain app, or a dozen lines of Bolt calling kb_ask over REST: your harness, your choice.

  3. [3]

    Ship it. Every question your team asks in Slack comes back grounded and cited — or honestly declined. Each one lands in the activity log, attributed to the bot's key.

  4. [4]

    When HR updates the handbook, replace the document and re-run your eval suite — proof the bot still answers right, before anyone notices it was wrong.

#people-help via mcp · slack-hr-bot

maria @kb how many vacation days do we get in Sweden?

kb 30 days of paid vacation per year, per the collective agreement. Sources: [1] Employee Handbook §4.2 · [2] Sweden Addendum

tomas @kb can I expense a standing desk for my home office?

kb The handbook doesn't cover home-office equipment purchases — ask #people-ops.

Abstained

Answer quality

A scorecard for your agents' knowledge.

Your team pins the questions that matter — including the ones the corpus shouldn't answer — and Distill grades the live answering path against them. A retrieval miss tells you to add a document; a wrong synthesis shows the retrieved, cited, and expected sources side by side; an answer in protected scope is flagged as exactly that.

24
Passed
2
Failed
3
Correct abstentions
What is the refund window for annual plans? Pass
Which regions is SSO available in? Retrieval miss
What's the CEO's home address? Abstained
Current SLA for P1 incidents? Synthesis fail

Activity

See every question your agents ask.

The full log — which agent, which answer, which sources, how fast. Nothing your bot says on your behalf is invisible to you.

Question Status Source Latency When
How many vacation days do employees in Sweden get? Answered mcp · hr-bot 1.4s 2 min ago
Can we offer net-60 payment terms to new customers? Answered mcp · sales-assistant 2.1s 11 min ago
What's our policy on side projects? Abstained api · intranet 0.9s 26 min ago
Which subprocessors handle EU customer data? Answered ui · member 1.7s 1 hr ago
Project isolation Each project's knowledge is its own tenant. Agents scoped to one project can never read another's documents.
Scoped, revocable keys A key is bound to exactly one project and named for its agent. Revoke it and access ends immediately.
Everything attributable Every answer traces to its documents; every question traces to the key that asked it.

Early access

Give your agents answers you can audit.

We're onboarding design partners now — teams putting AI agents in front of employees or customers, who need those agents to be right.