One trust loop · three surfaces

Govern your AI agents.
Laptop to production.

The same trust loop — supervise, gate, record — across three surfaces: a free open-source CLI on your machine, a hosted Cloud control plane for your team, and a Gateway for agents across organizations. Model-agnostic. Bring your own keys.

Install the CLI Explore the platform
todo-app — agentic (the CLI) running
0 Runtime dependencies
2+ Harnesses · Claude Code, Cursor
100% Local & git-native
Apache-2.0 Open source
The lineup

One trust loop.
Three surfaces.

The same loop — swarm, gate, relay, eval — from a single developer's laptop to cross-org agent commerce. Start local and open source, scale to a shared control plane, extend across organizations. Same bundle, same gates, same audit by run ID.

01 · Your laptop Open source
agentic

A supervisory framework for agentic development.

Define your agent fleet, lifecycle, and gates once — then compile them into Claude Code, Cursor, and AGENTS.md. Supervise every run at the harness's own hooks and record it as a tamper-evident provenance ledger. Non-invasive: generated config is marked and drift-checked — you rewrite nothing.

  • Define → Compile → Supervise → Record
  • Works with Claude Code & Cursor
  • Git-native, local-first, zero-dependency
  • Hash-chained provenance ledger
agentic init · project · gate View source on GitHub →
02 · Your team Hosted
Agentic Cloud

Shared trust, across your team.

Everything the CLI records — now shared and enforced. A hosted trust ledger and versioned registry, behavior CI that blocks a regressing promote, team relay to Slack & PagerDuty, and promote policy with RBAC. Audit by run ID across every repo.

  • Shared ledger + versioned registry
  • Behavior CI · eval gates on promote
  • Team relay · Slack / Teams / PagerDuty
  • Promote policy + RBAC
agentic cloud push · promote
03 · Your ecosystem Roadmap
Agentic Gateway

Where agents transact — safely.

Secure agent-to-agent communication and commerce across organizations. Verifiable agent identity, scoped and revocable delegation, human approval on the money, and a bilateral, signed audit trail — built on A2A / AP2 with settlement.

  • A2A / AP2 transport & discovery
  • Scoped, revocable delegation + mandates
  • Human-in-the-loop on every commitment
  • Bilateral signed audit · settlement
agentic gateway · relay

One umbrella brand, split by surface: local, solo, single-repo → the CLI (open source). Shared, multi-party, cross-org → the Cloud.

01
Free · local · open source

agentic — the CLI

Govern the agents already running in Claude Code, guide them in Cursor — from your machine. Define your fleet, gates, and standards once; supervise every run; record it as a hash-chained ledger. Non-invasive: its own state lives in one directory, and the harness config it compiles is generated, marked, and drift-checked — you rewrite nothing.

Get the CLI ↗
curl -fsSL https://raw.githubusercontent.com/Agentic-CLI/agentic-cli/main/install.sh | bash Copied ✓ Zero dependencies · Python 3.9+ · or uv tool install / pipx install
Definebundle.yaml — roles · lifecycle · gates
Compileproject into .claude/ · .cursor/ · AGENTS.md
Supervisegate at the harness hooks · relay a human
Recordappend-only, hash-chained ledger

Non-invasive

One .agentic/ directory. Generated configs are marked; control runs at the harness's own hook boundary. Remove it and your setup still works.

Reusable packs

Pin well-written personas & standards from a git repo — agentic add, sha-locked, with per-repo overrides. Author once, reuse across every project.

Provenance you can verify

Every agent action recorded and hash-chained; agentic doctor verifies the chain. The trace becomes the trust artifact.

Repository ↗ Packs docs ↗
02
Hosted · team · early access

Agentic Cloud — the control plane

Everything the CLI records, now shared and enforced across your team: a hosted trust ledger & registry, behavior CI that blocks a regressing promote, structured relay, and promote policy with RBAC. The sections below are the Cloud platform.

Join the waitlist
Agentic Cloud · The gap

Code has CI. Agent behavior does not.

Teams ship prompt edits like config changes — without a behavior gate, an audit trail, or a mandatory human stop.

Without Agentic CLI Slack-as-workflow
@priya prompt PR merged — LGTM on the diff
@release ship to staging tonight
Users getting wrong account-deletion steps
…47 messages later… no approver log, no replay
No CI gate · no run ID · thread is the "audit trail"
With Agentic CLI Relay + eval
a3f9c1 Auditor 0.61 < 0.75 — awaiting human relay
eval/1.4.3 agentic eval — 19/25 golden · promote to prod blocked eval
trace Full DAG · tools · approver · replayable
Queue is system of record · Slack only notifies
Agentic Cloud · Platform

Decide what ships.
Prove it before prod.

Integration platforms connect. Memory layers remember. Frameworks run. Agentic CLI decides what's safe to release — and proves it with behavior CI, structured approval, and audit on every promote.

Govern

Behavior CI

Prod never advances on regressing behavior. Golden-suite eval runs on every change — promote stays blocked until your agent passes.

agentic cloud eval · trace

Approval queue

Low-confidence outputs land in a structured queue — draft, sources, and run ID attached so approvers decide in minutes, not Slack threads.

Promote & audit

Promote policy enforces who can ship and what evidence they need. Every approval and eval result is append-only and replayable by run ID.

Connect

Federated tool catalog

Give governed workflows access to every app your team uses — discover tools in plain language via MCP and A2A, load only what each run needs.

Managed connections

Connect each integration once. Credentials auto-refresh from your vault and never enter git or agent context — so production agents stay scoped and auditable.

Scoped permissions

Each workflow declares exactly which tools it may invoke. High-risk actions stay behind your gates until behavior and approval say otherwise.

Governed memory

Federate semantic memory from Mem0 — bundle-pinned RAG for docs, allowlisted writes, PII redaction, and recall in eval. Every memory read traced by run ID.

Run

Versioned registry

Pin every production release — agents, eval suites, approval rules, and tool permissions at workflow://org/name@prod. Roll forward with confidence, roll back with one command.

agentic cloud push · promote

Durable workflows

Git-native definitions with branching, retry, and pause-for-approval — governed agents resume exactly where the run stopped, not from scratch.

Operate & deploy

Run from CLI or fullscreen TUI. Schedule jobs, react to webhooks, notify Slack — same governed APIs whether your operator is in terminal or web.

agentic cloud tui · /commands

Integration platforms connect. Mem0 remembers. Agentic CLI decides they're safe to ship — and proves it.

Request early access
Agentic Cloud · The loop

The pattern.
Every workflow.

Every bundle enforces the same trust loop — swarm, gate, relay, eval. Author in chat or web; the registry stores the version. Below: in-app assistant at staging — eval must pass before prod.

workflow://acme/in-app-assistant@staging Cloud registry · pinned runtime ref

Request in → swarm collaborates → auditor gates → human intervenes if needed → eval proves behavior before promote. New domain? Same loop — new registry bundle.

planner 3 steps
executor 4 sources
auditor 0.61
relay a3f9c1
eval blocked
Stage 1 · Planner

Click a stage · use ← → keys · pattern is fixed · bundle is yours

Agentic Cloud · Surfaces

Artifacts platform teams trust

Three surfaces, one platform. Script with CLI · operate in agentic cloud tui · supervise in web. Command mode runs /workflow, /eval, /relay, /trace without leaving the terminal.

agentic cloud tuiOperate · /commands
agentic cloud evalGate promote
agentic cloud pushPublish bundle
agentic cloud traceReplay run ID
CLI for scripts · CI · pipes 
agentic cloud eval in-app-assistant@staging
agentic cloud promote in-app-assistant --to prod
agentic cloud trace e7c4b2 --format json | jq '.checks[]'
TUI command mode · press : or / 
/workflow list
/workflow promote in-app-assistant --to prod
/eval run @staging
/relay approve a3f9c1
/trace open e7c4b2
/registry push
Claude Code Claude Managed Agents EVE LangGraph MCP Mem0 A2A Temporal Agentic CLI

Connect, run, and govern from one CLI — model-agnostic, git-native, auditable by run ID.

agentic cloud tui j/k navigate · : command
a3f9c1 in-app-assistant · conf 0.54 · account-delete relay
b8e21d in-app-assistant · conf 0.71 · billing policy relay
eval in-app-assistant@staging · 19/25 · promote blocked eval
workflow://acme/in-app-assistant@1.4.3
runtime:
  model: claude-sonnet  # or gpt, gemini — model-agnostic
  key: vault://acme/prod/anthropic  # BYOK · never in git
relay:
  on: auditor.confidence < 0.75
  notify: [web_queue]  # + slack, email
eval:
  suite: ./evals/support-golden.yaml
  min_pass_rate: 0.85
.agentic/policy.yaml
promote:
  to: prod
  require: [eval_pass, rbac:release]
# prod pointer frozen until eval green + approver
Agentic Cloud · Compare

Integration connects. Frameworks run.
Only Agentic CLI governs.

Integration platforms wire agents to apps. Memory layers persist context. Frameworks orchestrate execution. Observability records what happened. Release governance — behavior CI, approval gates, promote policy, audit by run ID — ships here as one platform.

Tool Strength Connect Run Govern CLI
LangGraph / CrewAI Agent orchestration — ✓ — —
LangSmith Observability — ~ ~ —
Temporal Durable execution — ✓ — —
Braintrust Eval & experiments — — ~ —
EVE (Vercel) Build & deploy agents ~ ✓ ~ —
Claude Managed Agents Hosted harness ~ ~ — —
Mem0 Persistent agent memory ~ ~ — ~
Integration platforms App connectivity & workflows ✓ ✓ — ✓
Agentic (CLI + Cloud) Release system for agents ✓ ✓ ✓ ✓

Agentic (CLI + Cloud)

Release system for agents

Connect ✓Run ✓Govern ✓CLI ✓

Mem0

Persistent agent memory

Connect ~Run ~Govern —CLI ~

Integration platforms

App connectivity & workflows

Connect ✓Run ✓Govern —CLI ✓

LangGraph / CrewAI

Agent orchestration

Connect —Run ✓Govern —CLI —

LangSmith

Observability

Connect —Run ~Govern ~CLI —

Temporal

Durable execution

Connect —Run ✓Govern —CLI —

Braintrust

Eval & experiments

Connect —Run —Govern ~CLI —

EVE (Vercel)

Build & deploy agents

Connect ~Run ✓Govern ~CLI —

Claude Managed Agents

Hosted harness

Connect ~Run ~Govern —CLI —

“One registry standard — every squad pins workflow:// refs, the same relay queue, the same eval gate. Prod never moves when a prompt edit breaks golden behavior.”

Elena Staff engineer · AI platform · FANG-scale company · illustrative
Bring your own keys

Your keys. Your vault. Never in git.

Agentic CLI is not an LLM reseller. Bring your own Anthropic, OpenAI, Azure, or Bedrock keys — resolved from your vault at spawn, with short-lived leases. The control plane and the execution plane use separate credential classes.

  • Two credential classes Control plane (who can push / promote / resolve relay) → SSO (OIDC) + CI tokens. Execution plane (LLM + MCP auth) → your org vault, runtime only.
  • BYOK, injected at spawn Keys live in HashiCorp Vault or your cloud secrets manager, referenced via vault://. They never touch agentic.yaml, never hit git, never appear in a trace, never land on operator laptops.
  • Secrets & PII redacted before write Content-level detection across drafts, sources, envelopes, and inter-agent logs — not just known key names.
  • Append-only, tamper-evident audit Hash-chained log with signed export and WORM retention — an auditor can verify nothing was altered or dropped.
  • Least-privilege MCP tokens Short-lived and read-vs-write scoped. The allowlist controls which server; the token controls what it can do.
  • Compliance routing, fail-fast Declare compliance: hipaa or fedramp and the orchestrator refuses ineligible backends at spawn — fail-fast, not fail-at-audit.
.agentic/runtime.yaml
runtime_policy:
  compliance: hipaa        # refuses ineligible backends
  secrets:
    llm:
      anthropic: vault://acme/prod/anthropic
      openai:    vault://acme/prod/openai
    mcp:
      crm: vault://acme/prod/crm-readonly  # scoped token
  redact: [output, draft, sources, envelope]
  audit: append_only     # hash-chained · signed export
03
Cross-org · settlement · roadmap

Agentic Gateway — agents that transact

Secure agent-to-agent communication and commerce across organizations: verifiable agent identity, scoped & revocable delegation, human approval on the money, and a bilateral, signed audit trail — built on A2A / AP2 with settlement.

Get updates

Verifiable identity

Know the counterparty agent is really who it claims — signed agent cards, org-bound.

Scoped delegation

Grant bounded, revocable authority (e.g. "book flights ≤ $2,000") — never your raw credentials.

Human on the money

The relay pauses for a human before any commitment that spends or is irreversible.

Bilateral signed audit

Both sides get a signed record of what was proposed, agreed, and settled — provenance across the boundary.

Ship agents you can prove.

Early access for platform teams — behavior CI, promote policy, and audit by run ID on every release. Connect to your stack and run durable workflows on the same platform.

Model-agnostic · bring your own keys · early access opening soon