Early access

Your agent can act.
Now make it safe.

AgentBlocks is an API that gives your AI agents real-world capabilities — email, git, integrations — with deterministic approval rules, scoped credentials, and full audit trails.

Get your API key → Read the docs

agent.py

                # Add AgentBlocks to any agent in 3 lines

                from agentblocks import AgentBlocks

                ab = AgentBlocks(api_key="ab_live_...")

                # Your agent requests an action — AgentBlocks handles the rest

                result = ab.email.send(

                    to="team@company.com",

                    subject="Weekly standup summary",

                    body=agent_generated_summary

                )

                # result.status: "approved" | "pending_review" | "rejected"

                # Approval rules, rate limits, and audit logging happen automatically

// capabilities

Built-in integrations.
One consistent API.

Every integration follows the same pattern. Request an action, apply rules, execute safely.

smtp

Agent Email

Dedicated @agentmx.io address per agent. Send and receive with full approval flows. No impersonation risk.

gmail

Gmail

Read, triage, draft, and reply from your real inbox. Every outbound message hits your rules first.

git

Git & GitHub

Push commits, open PRs, manage branches. Short-lived tokens, not your SSH key.

slack

Slack

Post to channels, DM teammates, react to messages. Scoped to specific workspaces and channels.

cal

Calendar

Read availability, create events, send invites. Your agent becomes a scheduling assistant.

...

More coming

Databases, file storage, webhooks. Same pattern. Same safety. Open an issue to request yours.

// request lifecycle

Every action flows through
a deterministic gate

No magic. No probabilistic safety. A request either passes your rules or it doesn't.

Agent calls API

Your agent requests an action through the SDK

Rules evaluate

Deterministic rules auto-approve, flag, or reject

Human reviews

Flagged actions wait for human approval in dashboard

Safe execution

Scoped credentials, rate limits, full audit trail

// security model

Not a wrapper. A security layer.

Your agents never touch raw credentials. Every action is gated, logged, and reversible.

[✓]

Scoped credentials

Short-lived tokens with least-privilege. Rotated automatically.

[✓]

Deterministic rules

No LLM decides what's safe. You write rules, they execute exactly.

[✓]

Human-in-the-loop

Any action can require approval. Review from dashboard or Slack.

[✓]

Rate limiting

Per-agent, per-action, per-time-window. Runaway loops die immediately.

[✓]

Full audit trail

Every request, evaluation, and execution. Exportable. Queryable via API.

[✓]

Zero persistence

Agent credentials expire. No long-lived secrets stored in your codebase.

// developer experience

Built for how you actually work

Not another dashboard you'll never open. AgentBlocks fits into your existing stack.

Framework agnostic any llm

Works with LangChain, CrewAI, AutoGen, raw API calls — anything that can make an HTTP request.

One-line setup

Add the AgentBlocks system prompt to your agent. That's it. No infrastructure, no sidecar, no proxy.

Typed SDKs

Python and TypeScript SDKs with full type hints. pip install agentblocks and go.

Local dev mode

Auto-approve everything locally. Your dev loop stays fast. Rules enforce in staging and prod.

// get started

Ship agents that do real things

Stop building demos. Start building agents that send emails, push code, and talk to the world — with guardrails your team can trust.

Get your API key →

$ pip install agentblocks

Your agent can act.Now make it safe.

Built-in integrations.One consistent API.