Changelog
AgentSDR changelog
What shipped, release by release, including the database migrations each release needs. This page is the project's CHANGELOG.md, read from GitHub.
Unreleased
Not yet tagged
Added
- Merge field fallbacks, 2026-11-21.
{{firstName|there}}sends "there" to a lead with no first name instead of an empty string. Choosing a field from Insert field asks for optional fallback text, and the list now always offers the standard fields (first name, last name, full name, company, job title, email, company domain) even when no lead has them yet, each marked with how many leads are missing it. The preview says when a lead gets the fallback. A merge field is never read as{A|B}spin text. No migration. - AI assistants (MCP), 2026-11-13. AgentSDR is an MCP server at
/api/mcp. Connect Claude (web and desktop), Claude Code, Cursor or any MCP client and work CRM, Action required from a chat: read the queue and records, review, edit and send drafts, correct classifications, move stages, snooze, close and set Do Not Contact. Each connection carries its own permissions (crm:read,crm:write,crm:send), actions are recorded as "name via AI assistant", and Settings, AI assistants shows the server URL, setup steps and each person's connections with Disconnect. API keys now have permissions too (campaigns:read,campaigns:write, and the CRM ones); keys created earlier keep full access. Seedocs/mcp.mdx. - AI assistants find any lead and write like the team, 2026-11-14.
crm_list_action_requiredis nowcrm_list_records: the queue by default, or every open lead (waiting, follow-ups exhausted) or closed ones, by stage, sub-stage, channel and search. New read-onlycrm_get_writing_guide(the drafter's reply rules, the team's draft instructions, the lead's sender, local time, sequence step and matching Knowledge) andcrm_get_knowledge. The rules are the Langfuse promptmcp/writing-rules; saving the Langfuse integration orscripts/seed-langfuse-prompts.tscreates it. ChatGPT apps created earlier need a refresh to see the new tools. - Public API with API keys. Owners and admins create keys in Settings, API keys; another application uses one to list email campaigns (
GET /api/v1/campaigns), add leads to a campaign (POST /api/v1/campaigns/:id/leads, the same duplicate and suppression rules as a sheet import) and stop a lead's sequence without suppressing them (POST /api/v1/campaigns/:id/leads/stop). Seedocs/api.mdx. - Edit a mailbox's display name and daily limit. Settings, Email, Mailboxes: a mailbox's menu has Name and daily limit. Both were set only when the mailbox was connected, and changing the "Emails per day" Sending rule does not touch mailboxes connected earlier, so there was no way to change either afterwards (
PATCH /api/outreach/mailboxes/:id). - Choose which mailboxes an email campaign sends from. A campaign's new Mailboxes tab: All mailboxes (as before, the default) or chosen ones, so separate campaigns can run from separate mailboxes. Each mailbox shows the other running campaigns that use it. New leads are only given a chosen mailbox; leads that already started keep theirs, so follow-ups stay in the same thread. Removing a mailbox takes it out of the campaigns that chose it.
Migrations
bun run scripts/create-api-keys.tscreates theapi_keystable.bun run scripts/create-oauth-provider-tables.tscreatesjwksand theoauth_*tables for AI assistants (MCP).bun run scripts/add-api-key-scopes.tsaddsapi_keys.scopes.bun run scripts/add-campaign-mailboxes.tsaddsoutreach_campaigns.mailbox_ids.
- Merge field fallbacks, 2026-11-21.
0.3.1
2026-10-08
Fixed
docker-compose.ymlnames its release's image (0.3.1) instead oflatest. A server that had pulled an earlierlatestkept running it with the new file: 0.2.0's image does not create the schema, so every query failed withrelation … does not exist. Upgrade by downloading the new file;AGENTSDR_VERSIONstill overrides the tag.- First-run setup no longer loops. On a fresh install the page bounced between
/sign-inand/setupforever: the signed-in frame wrapped/setupand sent the session-less visitor to sign-in, which sent them back. - A rejected database password is explained at startup: the bundled database keeps the password its volume was first created with, and the log now says so, with the two ways out.
0.3.0
2026-10-08
Changed
- Scheduled jobs run inside the app. There is no cron container any more: Docker Compose is two containers,
dbandapp. The jobs run in production only (never inDEMO_MODE);INTERNAL_SCHEDULER=falseturns them off for an external cron, which calls the endpoints as before. - Database schema is created when the app starts. The image creates it in an empty database before the server starts (there is no setup container). An existing database is only read, never changed. If this step fails the app does not start;
AGENTSDR_SKIP_DB_SETUP=trueskips it. - The daily rollover is per organization. Each organization's email send queues, Gmail push watch renewal and LinkedIn daily-limit reset run at the start of its own day, set in Settings, Organization, Defaults (Time zone, default
UTC, and New day starts at, default00:00). The first-run setup page stores the browser's time zone for the first organization. - LinkedIn outreach now runs every 30 minutes, and the LinkedIn webhook replay every 10. The LinkedIn search queue is no longer scheduled: it runs when someone clicks Run.
CRON_SECRETandOUTREACH_TICK_SECRETare optional: they are only needed to call the job endpoints from outside, and are no longer indocker-compose.yml. Unset, the endpoints refuse every secret call.- The job endpoints skip work the in-app scheduler already did (organizations whose rollover ran today; a slot that already ran).
?force=1skips that check.
Fixed
- LinkedIn job logs: concurrent jobs no longer leave the console patched, so later lines landed in an old run's log. Log capture is now per run (
AsyncLocalStorage).
Migrations
Upgrading a Docker Compose install from 0.2.0:
1. Replace
docker-compose.ymlwith the new one. 2.docker compose up -d --remove-orphans(removes the oldsetupandcroncontainers). 3. Run the migration above.On Dokploy, nothing needs creating or scheduling any more; existing Schedule Jobs are harmless and can be deleted. Keep
CRON_SECRETandOUTREACH_TICK_SECRETonly if you keep calling the endpoints.bun scripts/create-scheduled-job-runs.tscreatesscheduled_job_runs, where each scheduled run is claimed. With Docker Compose:docker compose exec app agentsdr-entrypoint bun scripts/create-scheduled-job-runs.ts. A fresh install gets the table automatically. Until the migration runs, no scheduled jobs run in-process (the app logs a[scheduler] Table scheduled_job_runs is missingwarning); the endpoints keep working.
- Scheduled jobs run inside the app. There is no cron container any more: Docker Compose is two containers,
0.2.0
2026-10-08
Upgrading a Docker Compose install from 0.1.0 (breaking): the new
docker-compose.ymlno longer loads your whole.envinto the containers; it passes only the settings it lists by name. Beforedocker compose up -d:1. Add
APP_URL=<your public address>to.env. It replacesBETTER_AUTH_URL/NEXT_PUBLIC_APP_URL, which the new file does not read. Without it the app assumeshttp://localhost:3000. 2. KeepPOSTGRES_PASSWORD,BETTER_AUTH_SECRET,INTEGRATION_CREDENTIALS_KEY,UNSUBSCRIBE_SECRET,CRON_SECRETandOUTREACH_TICK_SECRETin.envexactly as they are. 3. Any other variable you relied on (aPAUSE_*switch, worker tuning, Apollo) must be added to theenvironmentblock ofdocker-compose.yml.Added
- First-run setup page. On a fresh install every visitor is sent to
/setup: one form (name, organization, email, password) creates the first admin, already verified, and their organization, then signs them in. No email service is needed to get started. The page closes for good once an account exists. - Copy-and-run Docker Compose.
docker-compose.ymlruns the published image (ghcr.io/kandid-ai/agentsdr), with no repository checkout and no files to mount. Every setting is listed by name in the file; the ones to replace are marked# CHANGEMEand can be edited in place or set in a.env. It starts as-is onlocalhost; on any otherAPP_URLit refuses to start while a secret still has its placeholder. The scheduler runs from the same image.docker-compose.build.ymlbuilds from source instead. The Docker guide is rewritten as numbered steps. - Public demo.
DEMO_MODE=trueturns a separate deployment into a read-only AgentSDR anyone can open without signing in, andbun run db:seed:demonow fills it with Northwind, a fictional company's three months of email, LinkedIn and WhatsApp outbound, calls with transcripts, a full CRM pipeline with AI drafts awaiting review, Tables and Analytics. See docs/self-hosting/demo.mdx.
Changed
- License: MIT (was AGPL-3.0-only). Anyone may use, modify, self-host, distribute and sell AgentSDR, including in closed-source products and hosted services, as long as the copyright and license notice is kept.
- Email recipients: inbound Gmail keeps every To and Cc, replies can go to all, and a colleague's reply joins the lead's thread.
scripts/backfill-email-recipients.ts(optional, dry run by default) restores recipients on messages stored before. - Tables: about 70 fixes across columns, AI and enrichment runs and the grid.
- The app is only the app. The marketing website, blog theme and launch film moved to their own repository (agentsdr.ai).
/now opens the workspace (sign-in first), and the app asks search engines not to index it. - Call recorder 0.15.0: the built-in hosted address is now
https://app.agentsdr.ai. Self-hosted addresses are still added on the extension's Options page.
- First-run setup page. On a fresh install every visitor is sent to
0.1.0
2026-10-06
The first public release.
Added
- Accounts, organizations and teams with Better Auth: email and password sign-up with verification, Google sign-in, password reset, invitations, owner / admin / member roles, an organization switcher. Sign-up is invite-only by default (
AUTH_SIGNUP=openopens it); the first account on a fresh install can always be created. - Multi-tenancy: every record belongs to an organization; an isolation test suite (
scripts/e2e/tenancy-isolation.ts) proves one organization cannot see, change or count another's data. - Per-organization integrations configured in Settings — Unipile (LinkedIn and WhatsApp), Google Workspace (Gmail), Cloudflare R2, OpenRouter — with credentials encrypted at rest.
- Lead database with custom fields, email outreach with sequences and mailbox rotation, LinkedIn campaigns and search, WhatsApp messaging and recorded calls, an AI-assisted CRM inbox with reply classification and drafted responses, Tables with enrichment providers.
- Fresh installs:
bun run db:setupcreates the schema in an empty database fromdb/schema.sql;bun run db:seed:demoadds a demo organization with fictional data. - Unipile webhooks register themselves: saving the Unipile integration registers the LinkedIn and WhatsApp webhooks in your Unipile workspace, each carrying your organization and a generated secret.
- Settings by channel: Email, LinkedIn and WhatsApp each have their accounts, their Connection (the integration they run on) and their Sending rules — the limits that used to be hardcoded (emails a day and the gap between them, sending hours, LinkedIn invitations a day and per run, delays, search leads a day, WhatsApp warm-up, new chats a day, seconds between sends) — with today's values as defaults, hard ceilings and a warning past the safe edge. A LinkedIn account or WhatsApp number can set its own daily limit; the default phone country is per organization.
/settings/integrationsredirects to the new pages. - Call recorder Options page: add your own AgentSDR address in the extension; a self-hosted deployment no longer needs its own build.
- Docker Compose self-hosting, CI, and the project's community documents, including responsible use.
- Public website at agentsdr.ai: product pages for each channel, the AI CRM and the data tools, solutions for founders, agencies and sales teams, comparisons, guides and this changelog, with a sitemap and structured data for search.
Migrations
- Existing single-workspace installations: follow docs/multi-tenancy/cutover.md (
scripts/create-auth-tables.ts,scripts/add-organization-scope.ts, thenscripts/finalize-organization-scope.tsafter deploying). - Sending rules:
bun run scripts/add-channel-settings.ts(additive; run it before deploying). TheWHATSAPP_*sending variables are no longer read — set those limits in Settings → WhatsApp → Sending rules.
- Accounts, organizations and teams with Better Auth: email and password sign-up with verification, Google sign-in, password reset, invitations, owner / admin / member roles, an organization switcher. Sign-up is invite-only by default (
Blog
From the blog
Sales StrategyLemlist vs Instantly.ai: Which Sequence Tool Wins in 2026Lemlist vs Instantly: compare warmup, inboxing, pricing and channels to choose the right sequence platform, with a clear 2026 winner for your team.
Sales Strategy6 LinkedIn Outreach Software Tools for Lead GenerationCompare LinkedIn outreach tools for B2B lead generation, including AgentSDR, HeyReach, La Growth Machine and Clay, with safety and pricing guidance.
Sales Strategy6 B2B Sales Automation Workflows That Pay for ThemselvesCompare 6 B2B automation workflows that can pay back in under 4 weeks, from multichannel outbound sequences to instant lead routing and post-meeting follow-up.
Keep exploring
- GuidesPlaybooks for safe, multichannel outbound
- Open source & self-hostedMIT-licensed code on your server, with your own AI key
- Email sequencesCold email from your own Google Workspace mailboxes
- LinkedIn automationInvites and follow-ups across many accounts, safely paced
- WhatsAppRecorded calls and message campaigns from your own number
Follow along on GitHub.
AgentSDR is open source. Releases, issues and the roadmap all live in the repository.