Agents that use your integrations and propose the improvements
A Draftbase agent runs on a schedule, pulls from a connected integration, and turns what it finds into a change proposal against your entries. A human still reviews and approves every write. Use the built-in content gap FAQ miner, or build your own pipeline.
Updated
What a Draftbase agent is
An agent is one row: a type, a config, a run frequency, and an enabled flag. On its schedule — daily or weekly — it fetches data from a connected integration, matches it against your content, and proposes an edit. Nothing about that path writes to a live entry directly.
Two kinds exist today. A prebuilt agent — currently the content gap FAQ miner — ships as a fixed job with a config form: point it at a template and a field, and it runs. A custom agent is a pipeline you assemble yourself from five typed step kinds, for a case the prebuilt catalog doesn't cover.
Every agent instance can be triggered on demand too, not just on its schedule, and its last 50 runs are logged with how many proposals came out and how many AI tokens it spent. Nothing runs invisibly.
The content gap FAQ miner
Reads real search demand
Pulls Google Search Console queries with real impressions over a configurable window, up to 90 days.
Clusters near-duplicates
"How to reset password" and "how do i reset my password" collapse into one proposal, not two.
Matches to your best entry
Semantic search finds the closest-matching entry across up to 10 target templates per agent, and writes the proposal there.
Capped per run
A run proposes at most 50 changes, so a busy week of search data turns into a review queue, not a flood.
In practice: an agent watches your content model for gaps between what people search and what you've published, and hands your team a drafted answer instead of a raw keyword list. The content gap FAQ miner needs the Growth plan or above.
Building a custom agent
When the prebuilt catalog doesn't cover your case, a custom agent lets you chain five fixed step types into your own pipeline. No code, no eval, no arbitrary HTTP calls — each step is a small, typed handler validated by its own schema before the pipeline ever saves.
| Step | Does |
|---|---|
| fetch_integration_data | Pulls rows from a connected integration |
| filter | Keeps rows passing a numeric comparison |
| semantic_match | Matches a row to an entry by meaning, not keyword |
| ai_generate | Rewrites a field through Draftbase's AI model |
| propose_change | Turns a matched row into a reviewable proposal |
Custom agents are gated to the Enterprise plan (customAgents feature key). For the full step reference, config example, and troubleshooting, see Building custom agents in the technical docs.
Why agents propose instead of write
Every agent run — prebuilt or custom — produces change proposals, never a direct write. A proposal names the entry, the field, the old value, and the new one, and sits in a review queue until a person approves or rejects it. That review step isn't a setting to disable; it's how the feature is built.
The same discipline extends to AI spend: every AI-generated proposal is charged against the org's monthly token quota before it's counted as done, and a run that hits the cap skips the remaining rows instead of failing outright. Nothing about an agent bypasses the checks a human editor is bound by, including template validation.
Agents vs. MCP-connected AI
Both put an AI in your content, but they're not the same feature. MCP connects an external client — Claude Code, Claude Desktop, any compliant host — to your org through your own login, for one interactive session, calling the same tools a person would. An agent is a job Draftbase itself runs on a schedule against a data source you connect, with no client or session on the other end.
| Agents | MCP-connected AI | |
|---|---|---|
| Who acts | Draftbase, on a schedule | An external AI client, in a live session |
| Trigger | Daily/weekly, or on demand | Whenever the client connects |
| Can it publish? | No — proposals only, human approves | Yes, if the connecting user's role can |
| Plan | Growth+ (custom: Enterprise) | All plans |
Which plan you need
Agents are an add-on feature, not available on every plan. Hobby and Startup have none; the FAQ miner opens up at Growth; custom pipelines are Enterprise-only.
| Plan | Content gap FAQ miner | Custom agents |
|---|---|---|
| Hobby (free) | — | — |
| Startup ($49/mo) | — | — |
| Growth ($299/mo) | Included | — |
| Scale ($499/mo) | Included | — |
| Enterprise (custom) | Included | Included |
See the full pricing breakdown for environment and storage limits per plan.
Put an agent on your content backlog
Connect Google Search Console and let the content gap FAQ miner turn unanswered queries into review-ready proposals — or build your own pipeline on Enterprise.
Hobby is free, no card. Startup is $49/mo when you outgrow it. The price is on the pricing page, where prices go.
No migration quarter, no kickoff workshop. Define a template and ship something today.
Frequently asked questions
What are AI agent skills in Draftbase?
Draftbase doesn't use the term "skill" — an agent here is a scheduled automation with a fixed job: pull data from a connected integration, match it to your content, and propose an edit. The built-in content gap FAQ miner is one such agent; a custom agent is an org-built pipeline of five step types (fetch, filter, match, generate, propose) rather than a set of individually toggled skills.
What is a custom AI agent, and how do I build one?
A custom agent is a pipeline you assemble from five typed step kinds: fetch_integration_data, filter, semantic_match, ai_generate, and propose_change. There's no code to write — each step is validated by its own schema when you save the pipeline. See the full build walkthrough in the technical docs.
Does an agent publish content on its own?
No. Every agent run — prebuilt or custom — ends in a change proposal, not a write. A human reviews and approves each proposal in the dashboard before it touches a live entry. That review step is fixed platform behavior, not a setting you can turn off.
What's the difference between a Draftbase agent and an AI agent connecting over MCP?
An MCP-connected agent is an external AI client (Claude Code, Claude Desktop) acting through your own login, on your schedule, for one session. A Draftbase agent is a scheduled job Draftbase itself runs on your data, daily or weekly, whether anyone is at the keyboard or not. See the MCP pillar for the external-agent side.
Which plan includes AI agents?
The content gap FAQ miner needs the Growth plan ($299/mo) or above. Building your own custom agent needs the Enterprise plan. Hobby and Startup have no agent access.
What does the content gap FAQ miner actually do?
It reads your Google Search Console queries, finds ones with real impressions that your indexed content doesn't answer, clusters near-duplicate questions together, and proposes an FAQ addition to your closest-matching entry. You approve or reject each proposal.
Can an agent use any integration to fetch data?
As of the current API version, Google Search Console is the only supported source for fetch_integration_data. That covers both the prebuilt FAQ miner and any custom pipeline you build.
Is building a custom agent a security risk?
No arbitrary code runs. A custom agent's pipeline is limited to five fixed, fully-typed step handlers — no eval, no user-supplied function, no HTTP calls beyond the one integration path. The attack surface is exactly those five step types, whatever order you chain them in.
Why did my agent run produce zero proposals?
Usually one of three things: the semantic match step found nothing above the similarity floor, the org hit its monthly AI token limit mid-run, or a filter step's threshold was set against the wrong data window. Each run's detail shows a proposedCount and an error, if any, so a clean zero reads differently from a failed run.