Skip to content

Blog

Ideas for the future of work

Product updates, engineering deep dives, and practical guides on building a calmer, self-hosted, AI-native workspace.

OneCampAI GovernanceEU AI Act

AI Governance Is Mostly Boring Plumbing

The AI Act's transparency rules took effect on 2 August 2026. ServiceNow bought a company to watch what agents do at runtime. Only 23% of organisations report significant ROI from agents. All three are the same story: nobody can show what their AI did. Here is what governing an AI workspace actually consists of, mechanism by mechanism, including the four places where my own answer is still 'not yet'.

Akash Hadagali
OneCampMCPGo

We Opened OneCamp to Outside AI Agents, and Every Gate We Added Found a Bug

Claude Desktop, Cursor and anything else speaking MCP can now reach a OneCamp workspace — through an authority rule that is an intersection with a real person's live permissions, never a union, re-evaluated per call. Gating all thirty tools turned into an audit I hadn't planned, because writing an authority rule beside one that already shipped is how you find out the shipped one was wrong. Four times it was: a doc author locked out of their own doc, an agent able to DM a ghost identity people are refused, an event loop that walked past the delegation gate, and a write that could never be retried.

Akash Hadagali
OneCampGoNextJS

Agents That Can Ask Each Other For Help — Without Laundering Permissions

Your specialist agents couldn't talk to each other. Not by oversight — by a deliberate mute that stopped one agent replying to itself forever. This release replaces that blanket mute with a budget, so an agent can pull in the right specialist for part of a job. The interesting part is the constraint: every hop carries the person who started the chain, so a chain with no human at its root is refused. Also in here: why I decided not to implement A2A, and deleted the half I'd already built.

Akash Hadagali
OneCampGoNextJS

AI Teammates You Can Interrupt: Stop It, Steer It, and Actually Watch It Work

Handing work to an AI agent has had one uncomfortable property from the start: once it goes, it's gone. You watch a spinner, you can't change your mind, and if it's heading the wrong way you wait for it to finish being wrong. This release fixes that. You can stop an agent from the thread you're standing in, send it a correction while it's still working, and see what it's doing without refreshing. Long conversations no longer die at the model's context limit. And stopping an AI answer actually stops paying for it.

Akash Hadagali
OneCampGoNextJS

OneCamp AI Teammates Reach Your Whole Workspace: External Databases, Documents, Voice, and Governed Code

Last week OneCamp's agents learned to open pull requests. This stretch was about reach: an agent can now query your real production databases (Postgres, MySQL) read-only through a plan you can inspect and re-run, read the PDF and Word doc you dropped in chat, answer from your workspace search with clickable citations, take dictation and translate a message in one click, and build a little interactive tool you can actually click. And the coding agent grew up — it now works on any repo you can reach, on a model of its own so it isn't starved by chat, with destructive actions held for approval and protected branches it will not touch. All self-hosted, all on the model you choose, all bounded by the same permissions and budgets.

Akash Hadagali
OneCampGoNextJS

OneCamp AI Teammates Open Pull Requests: A Coding Agent That Ships on Your Own Infra, Not Someone Else's Cloud

This week OneCamp's agents learned to write code and open a pull request, the same way you'd hand the job to a teammate: @mention one, describe the change, and it clones the repo, edits, builds, tests, and opens a PR for a human to review. The interesting part isn't that it can, everyone's agent can now. It's where it runs: on infrastructure you own, through the model you choose, inside a network-locked sandbox, with the same approvals, budgets, audit trail, and a reliability scorecard that learns from whether your PRs actually merged. Here's how it works, and an honest read on where it stands next to Spotify's Honk and Anthropic's Claude Code.

Akash Hadagali
OneCampGoNextJS

OneCamp AI Teammates Get Senses and Skills: Vision, Data Analysis, Sandboxed Code, and Agents That Watch Your Work

Last week I gave OneCamp's agents a place in the room and the governance to make handing one to a team responsible. This week was about senses and skills: native tool-calling that works on modest local models, an agent that can see the image you paste and knows what day it is, answers straight from your tables with charts it draws inline, runs real Python in an isolated sandbox, follows a pull request through CI, keeps standing instructions you just tell it, and watches your work on a routine, all self-hosted, all bounded by the same permissions and budgets.

Akash Hadagali
OneCampGoNextJS

OneCamp AI Teammates: Agents That Live in Your Channels, With Governance You Can Audit

The next chapter for OneCamp's agents: instead of a separate chatbot, an agent becomes a real, named member of a channel or a DM you can @mention, with its own model and permissions. You can now assign it a task and trust it to finish, on a durable engine that survives restarts, resumes the same work, posts live progress, and pauses to ask when it is stuck. Plus the governance that makes that safe to give a team, approval modes for writes, per-agent and per-channel spend caps, a saved-test eval harness, a tamper-evident audit log, reliability dashboards, and a week of cross-surface AI that meets you where you already work.

Akash Hadagali
OneCampGoAI

Putting a Budget on AI: Workspace and Per-User Token Caps for a Self-Hosted Workspace

How I added enterprise-grade cost control to OneCamp's AI: metering and gating every model call at one provider chokepoint, daily token budgets for the whole workspace and per individual user, a per-run token cap and wall-clock deadline for agents, exact-or-estimated streaming usage, fail-open Redis metering, admin-editable hot-reloaded limits, and a usage panel, all self-hosted.

Akash Hadagali
OneCampGoAPI

Making OneCamp Programmable: A Public API, Scoped Tokens, an MCP Server, and a TypeScript SDK

Turning a self-hosted workspace into a platform: scoped, hashed personal access tokens with per-token rate limiting and audit, a /v1 API that reuses the app's own permission-checked business layer, OneCamp as an MCP server so any AI client can call its tools, an MCP client so agents can use external tool servers, a TypeScript SDK, and shareable templates, all without the API ever exceeding what its owner could do by hand.

Akash Hadagali
OneCampGoNextJS

OneCamp Tables: Notion-Style Databases Your AI and API Can Actually Use

Why a table in OneCamp is its own first-class entity rather than a blob inside a doc: a jsonb row model that survives schema changes, fractional ordering, grid/board/calendar views, a relation field that links rows to real tasks and projects, AI that builds a whole table from one sentence, and a permission model that mirrors the rest of the workspace, all self-hosted.

Akash Hadagali
OneCampGoNextJS

OneCamp Collaborative Boards: A Miro-Class Whiteboard On Your Own Infrastructure

How I built a real-time, multiplayer whiteboard into OneCamp: Excalidraw bound to Yjs over the same collaboration server the docs use, live cursors, an AI that draws diagrams and UI mockups, pinned comments with mentions, snapshot-based mass-delete recovery, and an access model that mirrors the rest of the workspace, all self-hosted.

Akash Hadagali
OneCampGoNextJS

OneCamp AI Agents: From a Chatbot That Answers to a Coworker That Does the Work

How I turned OneCamp's workspace AI from a question-answering assistant into an agent that creates and updates tasks, drafts docs, manages projects, and chains reads into actions, all while respecting the exact permission model a human user has, and never acting without confirmation.

Akash Hadagali
OneCampGoAutomation

Workflows in OneCamp: Slack-Class Automation Without the Sprawl

Welcome messages, keyword auto-replies, turning a message into a task, lightweight moderation — the automation every team eventually wants. I built a Workflow Builder for OneCamp that does it without a single extra runtime, and that can never do something its creator couldn't do by hand. Here's the design.

Akash Hadagali
OneCampGoPython

The AI That Sits In Your Calls: How OneCamp Turns Meetings Into Notes, Decisions, and Memory

Every call you have is full of decisions and action items that evaporate the second it ends. OneCamp puts an ambient AI assistant in the room — it transcribes live, then writes the recap, posts it where the call happened, and remembers the commitments. Here's how I built it on a self-hosted LiveKit stack.

Akash Hadagali
OneCampGoAI

AI Nudges in OneCamp: A Workspace That Remembers What You Promised

Most software waits for you to remember. OneCamp's AI Nudges flip that — a quiet engine that notices the commitment you made in a Tuesday call, sees it's overdue, and taps you on the shoulder. Here's how I built it, why it stays out of your way, and the privacy line I refused to cross.

Akash Hadagali
OneCampGoNextJS

OneCamp Connectors: I Gave the Workspace AI Access to Gmail, Google Calendar, and GitHub

How I built per-user OAuth connectors for Gmail, Google Calendar, and GitHub so OneCamp's AI can read emails, check your calendar, and surface pull requests — all on your own infrastructure, without a single byte of your data touching a third-party aggregator.

Akash Hadagali
OneCampGoNextJS

I Built a Full Slack-Style App Platform for OneCamp. Here's Every Decision I Made.

How I designed and built a slash-command and app platform for OneCamp — a data-driven external-app webhook system, built-in first-party apps, a curated marketplace, and a per-app encrypted secret store — without requiring a backend deploy every time a new app is added.

Akash Hadagali
OneCampGoNextJS

Hardening a Self-Hosted Workspace: Zero-Trust Ingestion and Client Resilience

How I built a zero-trust ingestion layer for OneCamp—implementing magic-byte file signatures, zip-slip and zip-bomb decompression protection, ClamAV socket scanning, and designed a visibility-aware exponential backoff polling engine for client-server resilience.

Akash Hadagali
OneCampGoNextJS

How I Built a Self-Hosted, AI-Agnostic Workspace with Active Memory Layering

How I re-engineered OneCamp to support hot-swappable local and cloud LLMs, implemented a self-calibrating EWMA token budget to prevent context truncation, and designed a structured workspace memory layer spanning Postgres, OpenSearch, and DGraph GraphRAG.

Akash Hadagali
OneCampGoNextJS

How I Built a Universal Import Engine Supporting 8 Different Productivity Platforms

How I engineered a high-throughput, chunk-based, multi-database migration engine for OneCamp that seamlessly unifies, de-duplicates, and ingests workspace data from Slack, Notion, Asana, Jira, Linear, Trello, ClickUp, and Todoist.

Akash Hadagali