Your stack on Visca
Pick the agent your team actually uses. See, line by line, where the operational gaps are today — and where Visca closes them. The agent doesn't change. The substrate under it does.
Anthropic's official CLI. Top of the coding-agent leaderboards in 2026.
Identity
SigilClaude Code runs as your shell user. Sub-agents it spawns inherit the same UID. No cryptographic root, no per-session identity, nothing to revoke.
Every Claude Code session is issued a Sigil at startup, bound to the human invoker and the runtime. Sub-agents chain their Sigils to it. One call revokes the entire chain.
Credentials
WarrantAPI keys live in `.env`; MCP server tokens sit in `~/.claude/settings.json`. A prompt injection that can read the workspace reads every key.
Claude Code requests a Capability Grant per action — scoped to verb, resource, duration, and constraint. No reusable secret in any file.
Discovery
PlexusMCP servers configured by URL in `.mcp.json`. Hardcoded endpoints, no authentication discovery, no Sigil-rooted attestation.
MCP servers found by capability descriptor through Plexus. mTLS by default. URL changes don't break anything.
Declarative state
CastImperative installs (curl scripts), settings file edits, hand-edited MCP configs. Three machines drift into three slightly different setups.
One Cast plan declares the substrate; `visca cast apply` reconciles. Three machines stay identical.
Packaging
SealSkills and CLAUDE.md files committed to repos. Not content-addressed, not signed. Two clones can diverge silently over time.
Each Claude Code skill or profile is a Seal Bundle — content-hashed, Sigil-signed. Same hash, same behavior.
Runtime
HearthLocal CLI process. Token spend visible only in Anthropic's dashboard; no per-task budget guardrail at runtime.
Hearth schedules and budgets each Claude Code run — per-task token ceilings, model fallback, durable execution if the process dies.
Audit
ChronicleTranscripts in `~/.claude/projects/`. Tool calls scattered across MCP server logs. No cryptographic chain, no cross-tool correlation.
One Sigil-keyed, hash-chained record of every Claude Code action. Cross-tool queries are a single SQL.
Dev surface
PraxisCLI + settings.json + MCP servers + hooks + agent files + skill files. Five surfaces, five mental models.
One `visca` CLI, one set of primitives — Sigils, Bundles, Capability Grants, Audit Envelopes — from laptop to production.
How to read this
Every popular agent today, regardless of vendor, faces the same eight operational concerns: identity, credentials, discovery, declarative state, packaging, runtime, audit, and developer surface. Visca puts one substrate underneath that closes all of them.
The agent itself doesn't change. You keep running Claude Code, Codex, Cursor, OpenClaw, Hermes, or Manus exactly as you do today. What changes is the substrate underneath — so identity is cryptographic, credentials are ephemeral, audit is one chain, and the dev surface is unified.
See it in your environment
Visca Cloud is the fastest way to put a Sigil, a Warrant, a Plexus, and a Chronicle under the agent you already run.