Agents Overview

Build, deploy, and operate Claude agents that use agntdata APIs, integrations, and your workspace database as tools.

agntdata Agents lets you ship a working Claude agent without writing any infrastructure. You describe what you want in plain English; a meta-agent (“the Builder”) configures and deploys it; the deployed agent runs on Anthropic’s managed runtime and pulls every tool it needs through agntdata’s MCP gateway.

Two surfaces, one workspace

The Builder

A meta-agent that interviews you in natural language, proposes a configuration, and applies it after your approval. Lives at /dashboard/chat in the dashboard.

Deployed Agents

Production agents you can chat with from the dashboard, fire on a schedule, route through Slack, or wake up via an inbound webhook.

What an agent gets

Every deployed agent fetches its tools through agntdata’s MCP gateway:

https://api.agntdata.dev/runtime/agents/{deployed_agent_id}/mcp

When the runtime calls tools/list, the gateway emits:

• Data API tools (data_<source>_<op>) — every enabled data API in your workspace.
• Connection tools (connection_<vendor>_<op>) — your configured BYOK + OAuth proxies.
• Workspace database tools (agnt_db_*) — optional, gated by config.
• Utility tools (agnt_credits_*, agnt_usage_*, agnt_webhooks_*, escalate_to_human).
• Skill tools (agnt_typescript_run_skill, agnt_skill_manifest) — if the agent has skills attached.
• Workspace files (agnt_files_*) — read and upload files scoped to the workspace.
• Memory tools (agnt_memory_*) — opt-in per-agent durable memory.
• Observability tools (agnt_observability_*) — opt-in, for monitor and nightly-report agents.

See Tools for the full registry and what each family does.

The agent lifecycle

            (you describe → Builder proposes → you approve)
   draft ───────────────────────────────────────────────────► deployed
                                                                 │
       chat in dashboard ◄────┬──── deployed agent ────┬──────► Slack threads
                              │                        │
              schedule fires ─┘                        └─ webhook deliveries
                              │
                              └─────► agent_session ───► event stream + audit
                                                                 │
                                                                 ▼
                                           usage_logs + agent_compute_usage

Every conversation — dashboard chat, Slack thread, scheduled fire, webhook delivery — runs inside an agent session. Sessions get a full event history and per-turn cost attribution.

Core concepts

• System prompt — the agent’s persistent role and rules. Edited via the Builder’s propose_agent_patch tool (with your approval) rather than direct DB writes.
• Tools — discrete capabilities the agent can call. Each invocation is logged to usage_logs (data wallet) and agent_compute_usage (AI wallet).
• Sessions — every conversation gets an agent_sessions row pointing at Anthropic’s canonical event history.
• Triggers — schedules, inbound webhooks, and the coming-soon queue. See Triggers.
• Permissions — mark sensitive tools as always_ask so the agent has to call escalate_to_human before using them.
• Escalations — the agent asks a human via escalate_to_human; the question lands in your Inbox (and a Slack channel, if connected).
• Folders — sidebar organization for when you’ve built more than a handful of agents.
• Memory — optional pgvector-backed memory so the agent learns across sessions.

How it works end to end

1. Describe the agent

   Open Agents → Build with AI and tell the Builder what you want. It pulls in your workspace profile and recent design memories so you don’t repeat yourself.

2. Approve a proposal

   The Builder drafts the system prompt and tool list and surfaces a proposal card with Approve / Reject.

3. Deploy

   On approve, agntdata mints an agent-scoped API key, registers the agent with Anthropic’s runtime, and exposes the MCP gateway URL the runtime uses.

4. Talk to it

   Chat with it from the dashboard, mention it from Slack, fire it on a cron, or POST to a webhook.

5. Iterate

   Open the Builder for the same agent any time and ask for changes. The Builder reads recent sessions and proposes patches with diffs.

Next steps

The Builder

How the meta-agent designs and ships agents.

Triggers

Schedules, webhooks, and the queue.

Sessions

Lifecycle and audit.

Tools

The full tool registry.

Workspace database

Durable state for agents.

Permissions

Human-in-the-loop control.

Folders

Sidebar organization.

Slack

Route Slack DMs and channels to agents.