Causly Server 1.6: Turn Any AI Agent Into Your Technical Co-Founder

In my first posts about Causly Server, the question was: what if your AI agent had hands?

It could touch your repository, your database, your deployments. That was the SDLC version of the idea, and it worked.

But while using it every day, I noticed something. The work of building a product isn't only code. It's also finding leads, researching the market, tracking a pipeline, writing content, and remembering what was decided in last Thursday's call.

So the question changed: what if your agent could work like a technical co-founder?

Causly Server 1.6 is the first version built around that question.

What Causly Server is now

Causly Server is an open-source MCP server (MIT license) that gives any MCP-compatible agent a set of tools to run the work around a software product. It now has 245 tools.

The original development tools are still there: files, Git and GitHub, Docker, Terraform, Postgres and MySQL, Supabase, Vercel, Sentry, Slack, Gmail, Notion, and an encrypted local secrets store.

On top of that, 1.6 adds the "co-founder" side:

Research: look things up and gather sources while the agent works.
Leads: find companies and contacts for outreach.
SEO and content: keyword ideas, page audits, content checks.
Pipeline: a local SQLite CRM for leads, deals, contacts and activity.
Meetings: transcribe a recording, pull out action items, set reminders, and prepare for the next call.
Calendar, Telegram and Linear: schedule, get notified, and manage issues.

Causly is still not an agent, an IDE, or a model. You bring the agent. Causly is the layer underneath that lets it do real work.

"Co-founder" is a metaphor, not a promise. The agent still needs your judgment. What changes is how much of the routine work it can carry.

Install in one command

The first posts said setup was manual: fork, clone, edit environment variables. That's no longer true.

npx causly-server init

init wires Causly into your client and sets safe defaults. It supports Claude Desktop and Cursor (init --client cursor), and any MCP client can connect to it.

After that, a few commands cover everyday setup:

causly-server setup slack # configure one service causly-server keys list # see which keys are set causly-server doctor --test # check every configured service causly-server login google # browser sign-in for Gmail and Calendar

Keys live in ~/.causly/.env and are hot-reloaded, so you don't restart anything after adding one. Node 22.13 or newer is required.

245 tools without flooding the context

More tools should not mean a worse agent. If you load every tool into the model's context, it spends its attention on the tool list instead of your task.

Causly loads tools in packs. By default only a small set starts: files, commands, project tools, plus research, leads, SEO, pipeline, reminders and meetings. That default tool list costs roughly 3.6k tokens. Loading every pack would cost around 36k.

For everything else, the agent calls use_tool. It can find a tool with search_tools and run it right away, without restarting the client. This matters because some clients, including Claude Desktop, don't refresh their tool list in the middle of a chat.

The goal is simple: the agent gets the capability it needs, when it needs it.

Safety: what changed in 1.6

Giving an agent real tools on your machine needs real limits. Before launch I went through a full security pass. These are the main changes:

Operator approval. Risky operations ask you first, through MCP elicitation. You can also set rules in ~/.causly/policy.json, for example to always allow some actions and always block others.
Workspace limit. init sets a workspace folder. File operations, Docker mounts and Compose files are kept inside it.
Minimal environment for child processes. Commands the agent runs no longer inherit every key you've configured.
Read-only mode. Set CAUSLY_READ_ONLY and the agent can look but not change anything.
Secrets stay secret. Credentials are redacted from logs and errors, and the local secrets store is encrypted with AES-256-GCM.
Prompt-injection awareness. Content read from Gmail and GitHub is marked as untrusted, so the agent treats it as data and not as instructions.
A readable audit trail. Every action goes to an activity log with size limits and rotation.

None of this makes an agent risk-free. It makes the risk visible and gives you control. I explain the thinking in more detail in a separate post on hardening an agent tool server.

How it's tested

An open-source tool is only useful if it works on machines other than mine. For 1.6:

More than 1,700 automated tests run on every change.
CI runs on Ubuntu, macOS and Windows, with Node 22 and 24.
I tested the installed package in Claude Desktop and Cursor.
I also tested it with a plain JavaScript MCP client and a LangChain JS agent.

Honest limits

I'd rather you hear these from me:

It runs on your machine. Causly Server OSS doesn't give your agent an isolated cloud environment. Your credentials, files and infrastructure are the ones the agent can reach. A managed version is in development, but the open-source server is what's available today.
macOS is tested through CI. I don't have a Mac to test on by hand, so I rely on the automated macOS runs.
Google sign-in can expire. While a Google app is in "Testing" mode, the token expires every 7 days. causly-server login google renews it in one step, and the docs explain how to publish your Google app to avoid this.
Some tasks use more tokens. Workflow tools that chain several steps cost more than a single call.

Open source, and what's next

Causly Server is MIT licensed and built in the open. It's listed in the official MCP Registry and on Glama, and it's on npm as causly-server.

Next, I'm focusing on what real use shows us: more client support, better setup for each service, and the managed version for people who don't want to run a server themselves.

If you try it, I'd like to hear what breaks and what's missing. Issues and pull requests are welcome, and there are good first issues open.