Airbyte

What it is

Airbyte is ELT software: it extracts data from SaaS apps and databases (Salesforce, HubSpot, Greenhouse, Zendesk, Postgres) and loads it into Snowflake, BigQuery, Databricks or a lake, where you model it. It is the open-source alternative to Fivetran, the category leader. The platform is source-available under the Elastic License 2.0, which lets you self-host it for free but not resell it as a managed service. Airbyte’s own connectors stay open source. The pricing page lists 700+ connectors on every plan.

In 2026 Airbyte added a second product line. The agentic data platform went into public beta on February 26, 2026 with 20+ agent connectors, among them Salesforce, HubSpot, Gong, Linear and GitHub. Airbyte Agents launched generally on May 5, 2026. The core of it is a context store, now marketed as a “context layer”: Airbyte copies a curated subset of your business data into an index that agents search, so a question doesn’t hit the Salesforce or HubSpot API on every call. Writes and actions still go through live connectors. The Airbyte MCP server is free on every plan, including self-hosted Core. The context layer and real-time queries require a paid plan, and Airbyte’s context-layer page says it is “available today for Snowflake and BigQuery users.”

Why it shows up in ops stacks

• You can run it inside your own network. For legal ops landing matter and billing data, or recruiting ops landing candidate records from an ATS, self-hosted Core or Enterprise Flex (hybrid and on-prem) keeps records out of a vendor’s cloud. Fivetran cannot match that at zero license cost.
• Capacity pricing beats row pricing on noisy tables. On Pro and Enterprise Flex you pay for Data Workers (processing capacity, typically about three concurrent syncs each), not for rows changed. A Salesforce object that a nightly workflow rewrites end to end costs the same as one that barely moves.
• One vendor for the warehouse and the agent data layer. A RevOps team that already syncs CRM and support data into Snowflake can point Claude, ChatGPT or Cursor at the same sources through Airbyte’s MCP, instead of wiring and governing a separate vendor MCP per app. The MCP classifies tools as read, mutate or destructive, requires explicit confirmation for destructive operations, and logs the acting user, tool and target object.

Pricing reality

Checked against airbyte.com/pricing on November 4, 2026:

• Core: free to self-host. You pay for the infrastructure and the engineer who upgrades and patches it.
• Standard: from $20 a month with 5 credits included (about 1 GB of syncs), hourly syncs at most. This rose from $10 in late September 2026.
• Plus: credit packages from $189 a month for 40 credits up to $4,999 a month for 2,000 credits, with SSO, 15-minute syncs, a 24-hour support SLA and up to two workspaces. Airbyte’s FAQ says Plus usually costs less than Standard above about 40 credits a month.
• Pro and Enterprise Flex: custom quotes priced on Data Workers. Pro adds 5-minute syncs and RBAC.
• Trial: 14 days with 400 credits, cut from 30 days in October 2026.

What teams pay: Vendr data cited by CostBench puts the median Airbyte contract near $16,000 a year, on a small sample of 13 deals, against a Fivetran median above $44,000. Read it as a mid-five-figure difference for a typical CRM, marketing and billing footprint. A team syncing a few gigabytes a month lands on Plus at $189 to $449; anything with database CDC or 5-minute freshness is a Pro quote. Airbyte Agents bills separately on agent operations (reads, searches, actions and reasoning calls). Get that rate in writing; the public pricing page does not list it.

Best for

Data and RevOps leads at companies with a warehouse and one engineer who is comfortable running containers, who want Salesforce, HubSpot, ATS and support data modeled in one place and either need data to stay in their own environment or have high-churn tables that make row-based billing unpredictable. It is the stronger pick when the same team will also feed internal agents from those sources.

Do not buy it if nobody on the team can own pipelines. A self-hosted Airbyte nobody upgrades is worse than a Fivetran bill. Do not buy it as the agent layer if your warehouse is not Snowflake or BigQuery yet.

Versus the alternatives

Fivetran is the volume leader and, since its June 2026 merger with dbt Labs, also owns the transformation layer. Pick Fivetran when you would rather pay more than run or debug connectors, and when one contract for ingestion, dbt and reverse ETL matters more than cost. Pick Airbyte when data residency or MAR unpredictability is the deciding constraint.

dlt, the open-source Python library from dltHub, is the fastest-growing entrant. Pick dlt when a data engineer already writes Python pipelines and wants ingestion as code in the repo with no platform to host. Pick Airbyte when you want a UI, a connector catalog and scheduling that a non-engineer can monitor.

Building your own agent data layer, with per-app vendor MCPs plus a Glean or Dust index, is the alternative to Airbyte Agents. Airbyte’s own benchmark claims 73% fewer tokens and 33% fewer tool calls than native vendor MCPs; it is vendor-run, so test it on your queries. Pick the DIY route when your agents mostly need documents and chat, not CRM records. The GTM context layer entry covers the design choice.

If the activation half is the problem, pushing warehouse scores back into Salesforce, pair either ELT tool with Hightouch.

Watch-outs

• Community connectors carry no SLA. Airbyte’s docs say certified connectors are maintained by Airbyte, while community connectors are maintained by contributors and should be used in production with caution. Guard: before you commit, list every source you need and check its support level in the connector catalog; for any community connector on a revenue-critical source, run a two-week trial sync and budget engineer time to fix it.
• The pricing page moves. Standard doubled, the trial shrank and per-credit overage rates disappeared from the plan cards within six weeks in late 2026. Guard: fix the credit rate, overage rate and annual uplift in the order form, and set a usage alert at 80% of your monthly credits.
• Self-hosting is not free. Upgrades, Kubernetes resources and connector failures land on your team. Guard: name an owner and an upgrade cadence before you deploy Core, and compare a year of that engineer’s time against a Plus or Pro quote.
• The context layer refreshes on a schedule, not in real time. An agent reading a deal stage from the index can be behind what a rep just changed. Guard: route anything that triggers a write or a customer-facing message through a live connector read first, and read the MCP write-access guidance before granting mutate tools.