LoginSign Up
← Back to all sparks
T

Twenty

CRM
Velocity6.3

Modern open source CRM with a fresh interface designed as an alternative to Salesforce and HubSpot.

twenty.com ↗

Twenty's open-source CRM hits v2.5 while wiring AI agents and credit-metered billing into the workflow core.

open-source-crmai-agentsworkflow-automationbillingmigrationsself-hosted
◆Current state
Twenty is shipping fast on its v2.x line, with five releases across April and May pushing AI agents as first-class workflow nodes and rolling out a billing v2 that meters AI credit usage. The release cadence shows the cost of that ambition: a string of cross-version upgrade hotfixes, agent-node execution bugs, and modal-loading regressions has accompanied the new surface area. The team is leaning into incremental hotfixes (v2.5.0 to v2.5.3 within four days) rather than batching.
◆Where it's heading
AI agents and credit-based metering are becoming structural to the product, not optional add-ons — the architecture is being reshaped to gate billing at AI entry points rather than per workflow step. Meanwhile the workspace migration runner keeps surfacing cascade-dependency bugs as the schema evolves, suggesting an underlying brittleness that will need a structural fix. The pattern is: new capability ships, upgrade paths break, hotfix lands.
◆Prediction
Expect a consolidation release that hardens the workspace migration runner against cascading column dependencies — the recurring pattern of fixing this case-by-case (v2.5.0, then the band-aids in #20581/#20583) signals a refactor is overdue. AI agent capabilities will continue expanding as the credit-cap architecture matures.

◆Recent moves

  1. 3d ago

    Revert version constants for v2.5.3 patch tag

    Cosmetic revert so the admin-panel inferred-version view reports the correct 2.5.x minor — v2.5.0 to v2.5.2 binaries had been picking up a stale 2.6.0 constant. No user-visible behavior change.

    View source ↗
  2. 6d ago

    Fix cross-version upgrades by seeding subFieldName column in 2.3 sequence

    Addresses the third in a sequence of cross-version upgrade failures: a cascading metadata-dependency lookup pulled the viewSort cache provider, which queried a column that did not exist yet on pre-2.3 instances. The structural fix — adding the column via a 2.3 instance command before the cascade fires — is more durable than narrowing the select, and matches the pattern Twenty's migration runner expects.

    View source ↗
  3. 7d ago

    Fix AI agent node execution after billing v2 added workspaceId requirement

    Regression introduced when billing v2 landed: agent steps were reading workspaceId from a workflow context object that never carried the key, silently passing undefined to the cache layer. A direct followup to the credit-metering architecture — the kind of integration cost that surfaces when billing becomes structural rather than peripheral.

    View source ↗
  4. 14d ago

    Fix plan-required modal stuck on loading skeleton

    Auth UX fix: users in plan-required state were getting an indefinite loading skeleton instead of the auth modal after a recent metadata-loading refactor. Restored by adding the plan-required paths to the metadata gater's exclusion list, mirroring how the invite flow already works.

    View source ↗
  5. 14d ago

    Automated i18n translation update

    Routine machine-generated translation sync from the GitHub action — no functional change.

    View source ↗
  6. 21d ago

    v2.1.1 — Hotfix: gate AI credit-cap at AI entry points

    ⚡ SPARK

    This is the architectural commit that defines how Twenty meters AI: credit-cap enforcement now lives at AI entry points (agent execution, generate-text REST, title generation) rather than gating per workflow step. The framing in the release notes — recognizing that per-step gating was too coarse — signals the team has settled on the right axis for billing v2.

    View source ↗