Engineering notes, product updates, and deep dives from the team building the open MCP platform.
Cron is a promise, not a delivery mechanism. Heartbeats — scheduled playbooks with structured reports, roll-up verdicts, escalation paths, and a reflexive self-check — are how autonomous work actually runs.
Approval fatigue is a design failure, not a safety feature. Policy before people, graduated autonomy bands, decision packaging, and an audit chain — how to keep humans in the loop by exception only.
A 500-tool MCP surface is a UI for a non-human user. Names are the primary interface, descriptions are error-prevention, and errors should teach the next call — hard-won patterns from production.
Agent memory can't be a pile of sticky notes. Bitemporal versioning — every fact versioned, every change reasoned, every belief queryable as of any date — turns 'the agent hallucinated' postmortems into queries.
The future isn't zero-human companies — it's supervised autonomy. Agents dispatch, execute, and report on their own; humans approve the irreversible. Here's the operating model that makes it safe.
We ran labeled recall probes against our live production memory store — 5,700+ real memory records (~2,000 active), not a synthetic corpus — three times over six weeks. The benchmark caught two real bugs and a store-pollution regression before we published a single number. Here are all of them, including the ones that embarrassed us.
Discover how FlukeBase is moving beyond simple CRUD to enable fully autonomous content management for AI agents.
AI assistants forget everything between sessions. Here's how a versioned memory system with semantic search gives agents true long-term context — across sessions, projects, and teams.
We wrapped the Google Workspace CLI as an MCP server — turning Gmail, Calendar, Drive, Sheets, and Tasks into AI-callable functions. Here's what that unlocks.
Press a hotkey. Speak your intent. Get a precise coding prompt. Voice-to-prompt development captures ideas at the speed you think them — and it's not dictation.
What if you could run 10+ production projects — across different tech stacks — on a single affordable VPS with shared CI, email, and DNS? Here's the architecture.
Instead of clicking through a dozen different web dashboards, what if every operation across your entire stack was a single function call? That's the MCP tool pattern.
Your meeting recordings sit in cloud storage collecting dust. What if every meeting was automatically transcribed, indexed, and queryable by AI — all on your own hardware?
Technical founders running multiple projects spend $300-500/month on developer SaaS. Self-hosted AI platforms are changing the equation — one binary, one VPS, zero vendor lock-in.