Rysh vs building it yourself: the boring 80% you'd rather not maintain
You're three sprints away from your own agent platform. You've been three sprints away for six months. Let's talk about why.
You're three sprints away from your own internal agent platform.
You've been three sprints away for six months.
I'm not mocking you — this is the most competent failure mode in software. Your team can build it. The prototype already worked: an agent loop, some tools, a Slack webhook, applause in the demo. Then someone asked "what happens if it runs rm -rf?" and you entered the part of the project nobody scoped.
This is the closing argument of our comparison series, because every other comparison secretly ends here. Not "rysh vs Warp" or "rysh vs LangChain" — rysh vs the platform your team would eventually build.
The demo is 20%. Here's the other 80%.
Everything below is what stands between "cool demo" and "thing the company actually leans on." None of it is glamorous. All of it is mandatory.
Tool plumbing. File edits, shell execution, git, code search, web fetch — each with timeouts, output caps, error surfaces, and edge cases. Rysh ships 46+ built-in tools that have already eaten those edge cases.
Approval policy. The difference between a toy and a coworker is knowing which operations wait for a human. Rysh gates destructive operations behind explicit approval — with a coloured diff preview before an edit touches disk — while read-only commands flow freely. You'd have to design that policy, build the UX for it, and defend it in a security review. We already did.
Secret handling. Your agents will read env files. Rysh redacts secrets from environment reads, and SecretNAT goes further: the model gets a token shaped like your key, never the value — verified on the wire. Would your v1 have that? Would your v3?
Multi-tenant isolation. The day team two joins, you're designing tenancy. Rysh isolates workspaces at the messaging layer with NATS subject ACLs — actual security engineering, not a tenant_id column.
Observability. "What did the agent do at 3am?" needs an answer a human can watch, not a log query. In rysh, every tool call renders live in a pane; the audit trail is the interface.
Channel connectors. Slack apps, IMAP IDLE, SMTP, an embeddable web-chat widget — each with auth, threading, reconnects, rate limits. Rysh humanoids ship with Slack, email, and website chat today (WhatsApp and phone are roadmap — we don't claim them as live, and your DIY plan shouldn't either).
Collaboration. Can a teammate watch your agent work? Take over? Rysh shares any pane — view or control — across the internet.
A UI. Your engineers will tolerate a CLI script. The moment anyone else touches the platform, you're building a product. Rysh is the product: a real terminal workspace where vim works, sessions detach and persist, and agents are panes next to your shell.
The line item nobody budgets: forever
The build estimate is never the problem — the maintain estimate is. Every one of those subsystems is now a codebase your team owns: through provider API changes, through the framework's next major version, through the departure of the one engineer who understood the approval flow. That's the real cost of DIY: not the six months, but the permanent tax on your best people.
| DIY platform | rysh | |
|---|---|---|
| Time to working demo | A weekend | An hour (brew install, one binary) |
| Time to governed, multi-user, multi-channel | Months, honestly | Day one |
| Approval gates + diff previews | You design & build | Built in |
| Secret redaction / SecretNAT | Usually "later" | Built in |
| Tenancy isolation | You architect | NATS subject ACLs |
| Channel connectors | One project each | Slack, email, web chat shipped |
| Who maintains it | Your senior engineers | Us — it's our whole job |
| Ownership | Total | Self-hostable, your keys, your data |
When building it yourself is right
Honest section, as always:
- Agents are your product. If the agent behavior is the thing you sell, own every line. Use a framework, go deep.
- Your constraints are genuinely exotic — air-gapped environments, custom models, compliance regimes no product fits.
- You're learning. Building an agent loop from scratch is the best education in this field. Do it once. Just maybe don't bet the roadmap on it.
And your DIY work isn't wasted with rysh anyway — rysh is an MCP client, so custom tools you've built can plug into the engine rather than being thrown away.
The actual question
Not "can we build it?" — of course you can. The question is: is an agent platform the best use of your best engineers, forever?
If yes — genuinely, godspeed, and publish what you learn.
If no: rysh is one self-hostable binary, running on Claude with your own API key, your infra, your data. (Self-hostable, not open source — we're precise about that.) The boring 80% is our entire job, so it doesn't have to be anyone's at your company.
You could be three sprints away. Or you could be done this afternoon.
We're recruiting design partners — teams who almost built this themselves. Bring us the requirements doc you already wrote; let's check it off together: become a design partner →
More in this series: Rysh vs LangChain and CrewAI · Is it safe to let an agent loose on your systems?