The control plane
for autonomous agents.
Agents aren't just answering questions anymore — they're logging into Stripe, writing to prod, moving real money. Ancha traces every call, holds the brake before anything expensive, and keeps a receipt for every dollar.
· · ·
Grok Agents, ChatGPT Work, Claude Agent SDK — autonomous agents are shipping to production. They send emails, approve refunds, write to databases, open PRs. Every incident now has a non-human cause. Every post-mortem starts with the same sentence: "we have no idea what it was thinking." Ancha is the layer that remembers, explains, and holds the brake.
Connect it. Trace it.
Hold the brake.
Everything you need to let an agent touch your real systems without paging the on-call engineer at 3am. Shipped as one CLI, one YAML file, one bin.
Wired in minutes.
25+ pre-built connectors. OAuth and scoped-token flows handled. Credentials live in Ancha's vault — not in your repo, not in the LLM's context window.
- Slack, GitHub, Stripe, Linear, Postgres
- Scoped tokens, least-privilege by default
- Build your own in ~40 lines of TypeScript
Every step traced.
A unified timeline per agent action: every reasoning step, every tool call, every side effect, every token spent. Replay from any point. Grep by user, by cost, by outcome.
- Reasoning · tool calls · side-effects · spend
- Replay & fork from any step
- OpenTelemetry-compatible exporter
Policies that preempt.
Define rules in YAML or code. Ancha evaluates before the side-effect happens: block it, cap it, or route it to a human for sign-off in Slack, email, or your own queue.
- Spend caps · scope rules · blast-radius limits
- Human-in-the-loop via Slack / email / webhook
- Dry-run mode for every new policy
Every thought.
Every call.
Every cent.
Agents fail in weird places: a tool call returns malformed JSON, an LLM decides to "retry just one more time," a cheap-looking loop quietly spends $40. Ancha stitches reasoning, calls, and dollars onto one timeline you can grep, replay, and attribute.
"We put Ancha on one customer-success agent and found a $180/day bug in a prompt loop we'd shipped a month ago." ENG LEAD · SERIES-B FINTECH · design partner
Policy as code.
Approval in the tools you already use.
Write a rule once, run it against every agent you deploy. Dry-run against last week's traces before enforcing. When a policy fires, humans see the full reasoning chain — not just a blind "approve/deny."
import { policy, approve, slack } from "@ancha/sdk"; export default policy({ name: "no-prod-refunds-over-100", on: "stripe.createRefund", // evaluate before the side-effect runs when: (call) => call.args.amount > 100_00, action: approve.via({ channel: slack("#ops-approvals"), include: ["reasoning", "recent_trace", "customer_history"], timeout: "15m", onTimeout: "deny", }), });
shopify-refunds-bot wants to refund $284.00 to customer #4821.
Wire your agent to the systems it needs — safely.
Scoped tokens, OAuth flows, and audit trails for every connector. Build your own in ~40 lines of TypeScript.
5 minutes.
Not a vendor rewrite.
Ancha slides in next to the agent you already have. It's a bin, a YAML, and an SDK — not a framework. No lock-in, no "port everything to our platform" tax.
- 1.Install the CLI. One brew/npm/go command. The sidecar auto-attaches to your running agents via a tiny shim.
- 2.Wire connectors. Pick from 25+ pre-built or write your own. Ancha handles OAuth, token rotation, and scoping.
- 3.Ship policies as you grow. Start permissive, tighten as you learn. Dry-run everything against last week's traces before enforcing.
The adult supervision
your agents need.
Autonomous agents are finally good enough to do real work. The next five years of software will be agents writing to your databases, moving real money, and sending emails that end up in a lawyer's inbox.
Which means every outage, every refund gone wrong, every mis-sent message — the root cause is now a chain of reasoning you weren't in the room for.
Ancha is the room. We log every step. We hold the brake pedal. We ship the audit trail. Your agents get smarter without your weekends getting worse.
Built in San Francisco by ex-infra engineers who got paged one too many times. Open-source core. Self-hostable from day one.
Free for solo.
Scales with your blast radius.
Start on the free tier, upgrade when you need SSO, dedicated residency, or SOC 2 reports for your customers.
- 100K traces per month
- All 25+ connectors
- Up to 3 policies
- Community Discord
- Unlimited traces & policies
- Slack / email / webhook approvals
- Replay & fork
- 14-day free trial · no card
- SSO · SCIM · audit log export
- SOC 2 · HIPAA · custom residency
- Self-hosted on your cloud
- Dedicated support + SLAs
Give your agents a conscience.
Install Ancha in five minutes. Watch what your agents are doing. Hold the brake before anything expensive.