Most CRMs were built for humans clicking buttons.
Salesforce, HubSpot, Pipedrive β great dashboards. But try giving one to an AI agent and it falls apart immediately. There's no API-first CRM built for the agentic era. So I built one. And its free
Headless CRM β open source, self-hostable, MCP-native.
What you get out of the box:
β 29 typed MCP tools (contacts, deals, pipelines, approvals)
β Full RBAC β give your agent exactly the permissions it needs
β Request human approval before any high-value action
β Works with Claude, Codex, Cursor, Windsurf
β SQLite for local dev, Postgres for prod. Zero config swap.
β Barebones UI, if you still want to track stuff Old School
Hand an agent a JWT and an MCP endpoint. It manages your entire pipeline by end of day.
Early beta β πͺ¨rough edges included. Would love brutal feedback from builders.
https://t.co/zQL25jgIeq
Joshua Kushner's Thrive Capital is putting $1B into buying local accounting firms and rebuilding the back office around AI. The acquisition arm is a company called Current.
The pitch to a decades-old CPA firm is unusually candid: sell us a majority stake, keep a meaningful piece for yourselves, and we'll re-engineer your operations with AI.
The ownership model is the part worth studying:
β Traditional PE buys to sell inside a fixed window
β Thrive says it plans to hold for the long run, the way Berkshire Hathaway does
β Local partners keep real, meaningful stakes
β Patient capital pointed at a fragmented, unglamorous, $100B+ industry
The proof point so far is Larson Gross β one accountant, one office in Bellingham, WA in 1949, now a regional firm with five offices and 200 staff. In 2025 its partners sold control to Current.
Forbes reports the in-house, self-improving models are hitting up to 98% accuracy on data entry.
Two things worth holding in view:
β Be precise. Data entry is the high-volume floor of accounting, not the judgment work clients actually pay for. The automation is real, but it's automating the bottom of the stack first.
β Be skeptical. AI roll-ups have been hyped for years and mostly underdelivered. The gap between the pitch and the operating reality is still wide.
The bet underneath all of this: that permanent capital plus AI can run a professional-services firm better than the people who spent decades building it.
If it works in accounting, the same template is waiting for law, insurance and consulting.
A billion-dollar wager on the back office of American business.
https://t.co/PUcRYeQOq5
Microsoft Build 2026 is done. The story isn't a feature β it's a posture shift, well for them π
Three years of "Copilot waiting for a prompt." Now: agents that act without one.
The stack they shipped to back it:
β Project Solara β chip-to-cloud platform for agent-first devices, built on Android (not Windows). Desk hub + wearable badge. Reference designs, not products. Pilots: Target, CVS, Best Buy.
β Microsoft Scout β first "Autopilot." Always-on, own Entra identity, runs across M365. Built on OpenClaw. Frontier preview only.
β 7 new AI models incl. their first reasoning model, in Azure AI Foundry.
β Surface RTX Spark Dev Box for local agent workloads (Very Cool)
The real product isn't the demos. It's the governance layer underneath β per-agent identity, policy checks, audit trails. An always-on agent acting unprompted in your tenant is the thing that scares security teams. That's the problem Microsoft is actually trying to solve.
Capability was never the blocker.
Deployability is.
https://t.co/BpmsKmRk7p
Anthropic's new Claude Opus 4.8 just quietly set a new bar for AI agents.
Here's what actually matters:
Honesty got a major upgrade. It's 4Γ less likely to let its own code bugs slip by unnoticed. It flags uncertainty instead of faking confidence. That's rare.
Agent performance? It's the only model to complete every case on the Super-Agent benchmark β beating GPT-5.5 at the same cost. 84% on browser agents. First to break 10% on the Legal Agent all-pass standard. Although not all these benchmark reflect real world experience!
Dynamic Workflows (Claude Code) now lets you spin up hundreds of parallel subagents in one session. Think: full codebase migrations, start to merge, unattended.
Same price as before. Fast mode is now 3Γ cheaper.
Oh, and Mythos-class models β the next tier above Opus β are coming to everyone in weeks, well thats whats being said
Worth checking out, not sure it will convince GPT 5.5 diehards, but seem decent upgrade from my experience
https://t.co/vDLkLMzMBR
"If you were starting your company today, would you build it this shape?"
That's the question Tom Blomfield (YC GP) closed his batch talk with. Almost nobody would.
His talk lines up beat-for-beat with the whitepaper we published a few weeks ago.
Companies are still structured like Roman legions. Hierarchies built so humans can relay information up and decisions down. That structure existed because moving information used to be expensive.
It isn't anymore.
AI isn't a better relay station. It's a different answer to the same question the Romans were asking.
The example that made it click:
YC built an agent to query their internal database/knowledge. A AI copilot.
Then they put a monitoring agent on top β watches every query, sees where it fails, and overnight writes the code, opens the PR, merges, deploys.
Next morning, the same query works or is improved!
That's not productivity. That's a company that improves while you sleep.
π₯ https://t.co/ungPI0sNkbπ https://t.co/M1C1bwHW1G
@Benioff@salesforce We gave our clients a CRM for free for their Agents and their people, opensource AI ontop Headless underneath
https://t.co/nOFJONzo64
Most CRMs were built for humans clicking buttons.
Salesforce, HubSpot, Pipedrive β great dashboards. But try giving one to an AI agent and it falls apart immediately. There's no API-first CRM built for the agentic era. So I built one. And its free
Headless CRM β open source, self-hostable, MCP-native.
What you get out of the box:
β 29 typed MCP tools (contacts, deals, pipelines, approvals)
β Full RBAC β give your agent exactly the permissions it needs
β Request human approval before any high-value action
β Works with Claude, Codex, Cursor, Windsurf
β SQLite for local dev, Postgres for prod. Zero config swap.
β Barebones UI, if you still want to track stuff Old School
Hand an agent a JWT and an MCP endpoint. It manages your entire pipeline by end of day.
Early beta β πͺ¨rough edges included. Would love brutal feedback from builders.
https://t.co/zQL25jgIeq
Most CRMs were built for humans clicking buttons.
Salesforce, HubSpot, Pipedrive β great dashboards. But try giving one to an AI agent and it falls apart immediately. There's no API-first CRM built for the agentic era. So I built one. And its free
Headless CRM β open source, self-hostable, MCP-native.
What you get out of the box:
β 29 typed MCP tools (contacts, deals, pipelines, approvals)
β Full RBAC β give your agent exactly the permissions it needs
β Request human approval before any high-value action
β Works with Claude, Codex, Cursor, Windsurf
β SQLite for local dev, Postgres for prod. Zero config swap.
β Barebones UI, if you still want to track stuff Old School
Hand an agent a JWT and an MCP endpoint. It manages your entire pipeline by end of day.
Early beta β πͺ¨rough edges included. Would love brutal feedback from builders.
https://t.co/zQL25jgIeq
Most CRMs were built for humans clicking buttons.
Salesforce, HubSpot, Pipedrive β great dashboards. But try giving one to an AI agent and it falls apart immediately. There's no API-first CRM built for the agentic era. So I built one. And its free
Headless CRM β open source, self-hostable, MCP-native.
What you get out of the box:
β 29 typed MCP tools (contacts, deals, pipelines, approvals)
β Full RBAC β give your agent exactly the permissions it needs
β Request human approval before any high-value action
β Works with Claude, Codex, Cursor, Windsurf
β SQLite for local dev, Postgres for prod. Zero config swap.
β Barebones UI, if you still want to track stuff Old School
Hand an agent a JWT and an MCP endpoint. It manages your entire pipeline by end of day.
Early beta β πͺ¨rough edges included. Would love brutal feedback from builders.
https://t.co/zQL25jgIeq
AI is not the root problem.
According to Eric Markowitz @ericmarkowitz, AI is a mirror reflecting a system that already treats human value as expendable.
Read the full article: https://t.co/gOw87Qei7O