Context Hub
A curated, versioned doc index your agents read before writing code, and the field notes they leave behind that become your team's institutional memory.
An agent's training data is always a little stale. APIs drift, and a model that "remembers" an SDK may be remembering last year's version. Context Hub is the fix: a curated, versioned doc index your agents can pull from on demand, plus a record of what they read and learned. Agents reach it through the chub command on their PATH; humans see the results as field notes on each agent's detail page.
Context Hub is delivered to agents as a built-in skill. You don't configure it per agent. Any agent running on a Concierto runtime can use chub, and what they do with it shows up automatically in their field notes.
How agents use it
Before writing code against an external API, an agent searches and fetches the relevant doc, scoped to its runtime language:
chub search stripe
chub get stripe/api --lang jsWhen it discovers something the doc didn't say. A gotcha, a missing field, a version-specific quirk, it annotates the doc so the next agent inherits the knowledge:
chub annotate stripe/api "Webhook verification requires the raw body; do not parse JSON before verifying"And when it's done, it rates the doc so quality signal flows back to maintainers:
chub feedback stripe/api up
chub feedback openai/realtime down --label outdatedDown-vote labels are structured. outdated, inaccurate, incomplete, wrong-examples, wrong-version, poorly-structured. So doc maintainers get actionable signal, not just a thumbs-down.
Field notes: what your team sees
Everything an agent does in Context Hub surfaces on its agent detail page → Field notes, grouped by doc:
The docs this agent fetched with chub get. The highest-frequency signal, since agents read before they code.
Annotations the agent captured with chub annotate, each linking back to the task that produced it.
The up/down votes the agent cast with chub feedback, so you can see which sources it trusts.
Why it matters
The annotations an agent leaves are the closest thing it has to long-term memory in your workspace. The next time any agent fetches the same doc, the annotation comes with it, so the team's collective knowledge compounds instead of resetting every task.
A populated Field notes tab reads as "this teammate is getting sharper." An empty one reads as "this agent didn't really do the homework." It's both a learning loop and an audit trail.
Good annotations are specific and singular. "Use idempotency keys on POST /v1/charges or retries will double-charge" earns its place; "Stripe charges money" is noise. Team conventions belong in skills or agent instructions, not in annotations on public docs that every future agent will see.