Documentation

Bring your stack into Bolt.

Bolt is the superagent your users talk to. Three integration surfaces let you bring your stack to it: Apps (declarative, no LLM tax), Agents (LLM reasoning, delegate to subagents your teams build), Retrieval (Aeira-indexed corpus search). Pick the surface per query. Your identity, your ACLs, and your audit chain apply to all three.

Start with Apps → See the three surfaces

The three surfaces

Three ways to bring your stack in.

Same security model, same audit chain, same identity. Different cost, latency, and authoring profile. Pick per integration, not per platform.

Mode 1 · Available now

Apps

Declarative integrations. No LLM in the loop. Zero LLM tokens on standard flows. Instant.

  • CostZero LLM tokens
  • LatencyInstant
  • AuthoringWizard or YAML
  • Best forKnown-shape queries

Explore Apps →

Mode 2 · Rolling out

Agents

The agent platform for your org. Define agents in Bolt, sync from your team's repo, or wrap services you already run. Bolt routes, composes, and audits.

  • CostTokens per query
  • LatencyBound by LLM
  • FlavorsLocal, Remote, System
  • Best forReasoning + routing

Read about Agents →

Mode 3 · Production today

Retrieval

Indexed corpus search at scale. Powered by Aeira. ACLs gate the data at the index, not as a postprocess filter.

  • CostIndex once, query cheap
  • LatencyFast at query time
  • AuthoringConnector config
  • Best forMillions of docs with ACLs

Read about Retrieval →

One unified integrations catalog. Apps and Agent tools run on MCP servers; Retrieval runs on data ingestors. Both kinds appear side-by-side in the catalog, and many vendors (Jira, Slack, Confluence, GitHub, Salesforce, and more) offer both. Connect to a source once and, where the vendor supports it, get a launcher chip, an indexed corpus, and a tool any Agent can call. Browse the full integrations catalog.

Decision guide

Which surface fits your use case?

The same question can land in any of the three surfaces. Pick by how structured the query is, how much latency you tolerate, and whether the answer is worth tokens.

Reach for Apps when

Apps

  • You have Jira, Slack, GitHub, or Salesforce APIs your team uses daily and want one-click access from the launcher.
  • You have a custom 1P internal API or MCP server and want low-code authoring with instant queries.
  • The question has a known shape and you don't want to pay tokens to answer it.

Reach for Agents when

Agents

  • The question is vague or open-ended and needs an assistant to plan and execute.
  • Your teams are shipping specialized internal agents and you want Bolt to discover and route to them.
  • You already have an agent service (LangChain, OpenAI Assistant, internal platform) you want Bolt to call.

Reach for Retrieval when

Retrieval

  • You have millions of internal documents and want to search them with ACLs respected.
  • The corpus is large and context matters, and you want RAG with real authorization gating.
  • The right document isn't predictable from the question.

The three surfaces compose. A workflow can use Apps to fetch structured data, Retrieval to ground the answer in your corpus, and an Agent to reason across the result.

Architecture as governance

Governance is the architecture, not a feature.

Bolt and Aeira were built with governance as the first design decision, not the last. Four commitments hold across every surface, every agent, every subagent call. They are properties of the system, not options you turn on.

Per-user OAuth only. No service accounts.

Every call carries the requesting human's identity through the full chain, including subagent delegation. An integration cannot escalate beyond what the human running it can do. There is no service-account mode.

Authorization runs before the LLM, not after.

ACLs gate retrieval at the index. Tool calls run with the user's permissions. Documents the user cannot read never enter the candidate set in the first place, so there is no postprocess-filter leak surface.

The audit chain holds across composition.

Every routing decision, every reasoning turn, every tool call, every subagent invocation, every emit. Tamper-evident receipts. The same chain across Apps, Agents, and Retrieval. When Bolt delegates to a subagent, the chain follows.

The LLM is at the boundary, not at the core.

Tools see real values so they work. The model sees redacted ones, only when you opt in to a model. Bring your own model. Your data does not leave your perimeter to train a vendor's. The system works without an LLM, and works better with one.

Start with what's live today

Apps and Retrieval ship. Agents are rolling out.

Five minutes from gallery to chip-in-launcher for a bundled App. Fifteen minutes for your first custom manifest. Agents grow into the same surface as they land.

Write your first App Browse 36 integrations

Or read the full manifest spec.