---
title: CrewAI capabilities
description: >-
  Choose a CrewAI operation and interpret its risk, approval, idempotency,
  streaming, cancellation, retry, and side-effect contract
kind: index
audience: application-developer
appliesTo: "1.x"
writingStandard: "aip-docs/1.0"
lastReviewedRevision: "97be86e9efedf07ecf1783b03800f683f107fb04"
connector: crewai
aliases:
  - CrewAI capability index
  - CrewAI operation catalog
---

# CrewAI capabilities

Use this index to choose the CrewAI operation that owns a task and interpret
its AIP contract before invocation. The reviewed connector has ten frozen
operations, but each configured crew publishes only its admitted subset.

The source-derived machine snapshot owns the ordered inventory. Focused pages
own input details and task guidance; neither artifact replaces authenticated
tenant discovery.

## Catalogue identity

| Property | Reviewed value |
|---|---|
| AIP source | `97be86e9efedf07ecf1783b03800f683f107fb04` |
| CrewAI source | `bfa652a7be8637562cc9b0833f75d927a64552d1` |
| CrewAI package in default product image | `1.15.5` |
| Connector profile | `aip.connector.crewai.v1` |
| Frozen operations | 10 |
| Low, medium, and high risk | 3, 5, and 2 |
| Mutations requiring approval | 7 |
| Operations requiring an idempotency key | 8 |
| Streaming operations | 6 |
| Operations advertising cancellation support | 5 |

The catalogue count is not a manifest count. One fully enabled crew publishes
ten capabilities, while a descriptor using compatibility defaults publishes
four.

## Read capability IDs

The `run` operation keeps the historical base capability:

```text
cap:crewai:<crew_id>
```

Every other operation appends its frozen suffix:

```text
cap:crewai:<crew_id>:<operation>
```

For a crew named `support-primary`, examples include:

```text
cap:crewai:support-primary
cap:crewai:support-primary:status
cap:crewai:support-primary:batch_run
cap:crewai:support-primary:memory_reset
```

Do not append `:run` to the base ID. Do not derive IDs from Python module
names, CrewAI object names, sidecar URLs, or provider credentials.

## Choose a family

| Family | Focused page | Operations | Primary use |
|---|---|---|---|
| Run lifecycle and replay | `run-lifecycle-and-replay.md` | `run`, `status`, `events`, `cancel`, `batch_run`, `replay` | Execute work and control or inspect its durable job |
| Training and testing | `training-and-testing.md` | `train`, `test` | Create a training artifact or evaluate a crew |
| Knowledge and memory | `knowledge-and-memory.md` | `knowledge_query`, `memory_reset` | Read indexed knowledge or reset an explicit memory domain |

Select the focused page before constructing input. Operations sharing the
same sidecar job route still have different schemas and safety contracts.

## Compare all operations

| Operation | Kind | Risk | Mutation | Approval | Key | Stream | Cancel | Retry |
|---|---|---|---:|---:|---:|---:|---:|---:|
| `run` | Agent | Medium | Yes | Yes | Required | Yes | Yes | No |
| `status` | Tool | Low | No | No | Optional | No | No | Yes |
| `events` | Tool | Low | No | No | Optional | Yes | No | Yes |
| `cancel` | Tool | Medium | Yes | Yes | Required | No | No | No |
| `batch_run` | Agent | Medium | Yes | Yes | Required | Yes | Yes | No |
| `replay` | Agent | Medium | Yes | Yes | Required | Yes | Yes | No |
| `train` | Workflow | High | Yes | Yes | Required | Yes | Yes | No |
| `test` | Workflow | Medium | Yes | Yes | Required | Yes | Yes | No |
| `knowledge_query` | Tool | Low | No | No | Required | No | No | Yes |
| `memory_reset` | Tool | High | Yes | Yes | Required | No | No | No |

`Cancel` in the table reflects published connector implementation support. It
does not assert that a CrewAI provider method or every resulting side effect
stopped.

The explicit `cancel` capability requests cancellation of another run. The
cancel request itself is not a cancellable long-running job.

Streaming support means the connector can publish ordered job events and a
terminal result. It does not mean every underlying CrewAI method can be
interrupted.

## Understand discovery projection

The host starts from a public descriptor for each crew. Its
`allowed_operations` set controls which capability objects enter the manifest.

| Descriptor policy | Published surface |
|---|---|
| Explicit subset | Exactly that validated subset |
| Missing legacy field | `run`, `status`, `events`, and `cancel` |
| Empty set | Startup failure |
| Unknown operation | Descriptor deserialization failure |
| Crew absent from sidecar health | Host remains unready |
| Descriptor operation absent from sidecar health | Host remains unready |

Runtime discovery is also scoped by tenant, connector admission, route
eligibility, instance state, artifact identity, and manifest limits. A
well-formed capability ID alone does not authorize a call.

## Interpret methods and fixed sidecar paths

| Operation | Sidecar method | Sidecar path |
|---|---|---|
| `run` | `POST` | `/jobs` |
| `status` | `GET` | `/jobs/{run_action_id}` |
| `events` | `GET` | `/jobs/{run_action_id}/events` |
| `cancel` | `POST` | `/jobs/{run_action_id}/cancel` |
| Remaining startable operations | `POST` | `/jobs` |

Streaming startable operations use `/jobs/stream`. Event replay uses its own
event path with an SSE `Accept` header.

At this revision, published binding metadata still assigns `/jobs/stream` to
every streaming operation, including `events`. Runtime code ignores that value
for event replay and calls `/jobs/{run_action_id}/events`; the machine snapshot
retains the published metadata.

These paths are private bindings between the Rust host and the frozen sidecar.
AIP callers invoke capability IDs through the admitted host route.

## Apply approval and idempotency

Every mutation requires tenant-policy approval. The published approval window
is 900,000 milliseconds and requires a reason, input snapshot, and policy
decision.

| Contract | Read-only status or events | Knowledge query | Mutation |
|---|---|---|---|
| Approval | No | No | Required |
| Idempotency key | Optional | Required | Required |
| Key scope | Tenant | Tenant | Tenant |
| Collision behavior | Revalidate input hash | Revalidate input hash | Revalidate input hash |
| Published key TTL | 86,400,000 ms | 86,400,000 ms | 86,400,000 ms |
| Connector retry support | Yes | Yes | No |

`knowledge_query` is read-only for provider state but starts a durable sidecar
job. That coordinator boundary requires a key.

The reviewed failure mapper still sets `retryable: false` on every concrete
connector failure. The table describes published support and safety, not an
automatic retry instruction.

An idempotency key identifies one intent. It does not prove exactly-once model,
tool, memory, knowledge, or external-provider effects.

## Interpret execution and cancellation

`run`, `batch_run`, `replay`, `train`, and `test` support blocking and streaming
completion. `events` supports blocking JSON replay and SSE replay.

All five streamed mutations advertise sidecar cancellation. A stream failure
can therefore trigger a remote cancellation attempt for any of them.

Replay, training, and testing execute synchronous CrewAI methods in worker
threads. Their common cancellation route cancels the asyncio waiter, records
`cancelled`, and can return `remote_stop_confirmed: true` while provider work
continues.

Cancelling an `events` stream only detaches the read. It does not change the
target job.

## Apply data and side-effect rules

Every capability declares confidential data, possible personal information,
and required redaction.

Read-only operations declare read and external-network side effects. Mutations
declare read, write, external-network, and code-execution effects.

Every mutation publishes `rollback_not_supported`. The connector exposes no
transaction, reconciliation, compensation capability, or provider rollback.

Crew factories can invoke deployment-authored Python, models, tools, storage,
and networks. Capability risk does not replace workload isolation or least-
authority provider credentials.

## Use the machine snapshot

The JSON asset contains source identities, aggregate counts, contract defaults,
and one ordered record for each operation. Each record includes its capability
template, family, method, private route, kind, risk, mutation, approval,
idempotency, execution, input summary, and side effects.

Its reviewed SHA-256 is:

```text
d004e0dd77115304dfdf105817efaf630138d6b6d383147518271c14e61f359c
```

The snapshot is a source-derived review artifact. It does not prove host
admission, tenant authorization, artifact qualification, or live provider
compatibility.

## Choose the next page

- Use `run-lifecycle-and-replay.md` for execution, status, event, cancellation,
  batch, and task-replay inputs.
- Use `training-and-testing.md` for bounded iterations, artifact names, and
  evaluator selection.
- Use `knowledge-and-memory.md` for queries, thresholds, result limits, and
  destructive reset domains.
- Use the run and recovery guides for procedures rather than field lookup.

## Related documentation

- [Source-derived capability snapshot](../assets/capabilities.json)
- [CrewAI connector overview](../README.md)
- [CrewAI connector configuration](../reference/configuration.md)
- [Capabilities and contracts](../../../concepts/capabilities.md)
- [Approvals and policy](../../../concepts/approvals-and-policy.md)