Documentation
Eve + Krumtap deployment control plane
A practical guide for using Vercel Eve to help humans deploy and operate Institutional MPC Wallet instances—without giving an LLM direct authority over Kubernetes.
Initial architecture
The control boundary: agents coordinate and explain; the approved control plane executes.
Web chat, Slack, GitHub, scheduled reviews
Instructions, skills, typed tools, subagents, durable waits, evals
VaultSpec, planner, policy checks, approval gate, executor
Legacy configurator bundles, code, known issues, logs, inventory, image catalog
Source hash, spec revision, plan hash, image digests, policy result, trace, approval, remediation
Approved executor only: Helm, Kubernetes, secret provider, diagnostics, rollback
Vercel Sandbox is for untrusted parsing and dry validation. It must not receive production cluster credentials or become a bypass around the control-plane approval gate.
Deploy and operate wallet instances
Manual Helm recovery and day-two operations for instances created by Krumtap Deployer.
Use each instance’s generated helm-upgrade.sh when the Deployer is unavailable or after a reviewed recovery.
Start with Helm status, namespace events, pod describe, and current and previous container logs.
Restart a known workload, roll back a reviewed Helm revision, or uninstall deliberately. The guide marks destructive actions.
Tutorial: build the agent in five decisions
Use Eve as the human-facing deployment layer, not as a replacement for deployment policy.
1. Define a narrow agent contract
The agent may create or revise a draft, explain a plan, inspect evidence, and request approval. It cannot treat a chat message as permission to change a cluster.
2. Keep planning durable
A rendered plan is stored with its spec revision and hashes. The workflow parks during a human review and resumes with the exact approved plan—not a newly generated approximation.
3. Sandbox only untrusted work
Use a sandbox for parsing legacy bundles, chart rendering, schema checks, and dry runs. The real executor keeps Kubernetes and secret-provider access behind the approved control plane.
4. Require evidence for diagnoses
Known issues need a detector, source links, affected versions, owner, remediation, and confidence. Retrieval without runtime evidence is only a hypothesis.
5. Operate through the channels people use
Expose the same agent through web chat and Slack/GitHub, schedule drift reviews, run evals in CI, and correlate every run to its deployment evidence.
Recommended agent directory
This mirrors the existing Deployer agent while making deployment safety explicit.
agent/ ├── instructions.md # deployment rules and approval boundary ├── skills/ │ ├── legacy-import.md # mpa-wallet.yml → VaultSpec rules │ └── known-issues.md # source-backed diagnosis protocol ├── tools/ │ ├── import_config.ts # draft only │ ├── render_plan.ts # dry, content-addressed evidence │ ├── get_runtime_logs.ts # diagnostics evidence │ └── apply_plan.ts # requires human approval ├── channels/slack.ts # operator approvals ├── schedules/drift.md # daily drift / expiry review └── evals/import-config.ts # regression fixtures
Implementation pattern
Each tool maps to a Krumtap capability and returns evidence that can be reviewed later.
Interpret mpa-wallet.yml + bootstrap + values into a draft VaultSpec. Never resolve or copy secret values.
Render the immutable plan, Helm diff, secret inventory, image resolution, policy violations, and blast radius.
Match log fingerprints to known issues, then cite the code/config/runtime evidence and remediation.
Require human approval and invoke the existing exact-plan control-plane operation; capture the result in the ledger.
Where Eve is unusually strong
Integration density, not a novel agent loop.
AI Gateway: choose and fail over models without binding the deployment agent to one provider.
Workflows: park at approval, wait for a channel response, and resume after restarts with durable state.
Connect + channels: put scoped external access and operator interaction on a platform-owned path rather than scattering long-lived tokens.
Sandbox + observability: isolate validation work and correlate agent behavior with deployment results.
Competitive landscape
Choose the runtime based on the control point you need most.
Filesystem-first agent directory plus Vercel workflows, sandbox, connections, channels, schedules, evals, and deployment conventions.
Best durable workflow spine when cross-platform infrastructure control is the priority; agent and channel conventions are application work.
Best explicit state graph and interrupt/resume model for deeply custom agent topology; compose hosting, auth, and channels.
Strong TypeScript primitives for tools, handoffs, guardrails, HITL, sessions, and tracing; choose the workflow/runtime platform separately.
Continue with the Wallet instance wizard to create a draft, then render a plan before any apply operation.