PYTHON
python
A Python transform. Receives upstream artifacts as loaded objects (dataframes arrive as pandas DataFrames) and emits new artifacts.
FEATURES
Everything in the box.
podblocks is one platform, not a stack you assemble. Here is what it actually does.
HYBRID GRAPH
Seven block types, one graph.
Code blocks and agent blocks live in the same DAG and pass data through one typed artifact contract. A dataframe leaves a Python block and arrives in a SQL block as a DuckDB table, or in an agent block as a schema plus a sampled head, never the full payload.
SQL
sql
A SQL statement executed by DuckDB against upstream dataframe artifacts, which appear as tables named by the producing node.
AGENT
agent
A single AI agent with a role, goal, and tools, compiled to a one-agent CrewAI crew.
CREW
crew
A multi-agent group as one node, sequential or hierarchical, defined in agents.yaml and run via CrewAI.
SUBPOD
subpod
A whole pod embedded as a node, with its own poset; only exported boundary events surface in the parent.
BROWSER
browser
A Playwright browser automation script, run in a sandboxed subprocess with a timeout.
MARKDOWN
markdown
A documentation node, rendered in the editor and never executed.
CAUSAL POSETS
Every run explains itself.
Each run produces a causal event poset following our goRapide engine. This means we know not just what happened, but what caused things to happen, which makes the agent smarter.
CONSTRAINT GATING
Rules that change the run.
Declarative constraints over the causal stream (never, must_match) enforce five policies mid-run: halt, retry, route to a remediation block, require human approval, or log. Enforcement runs in-process on the worker at function-call latency.
CAUSAL OVERLAY
See the why on the canvas.
The UI overlays the runtime causal graph onto your design-time DAG. Crew nodes expand into their internal subgraph of agent, task, tool, and LLM events. Click any event and the inspector answers root-cause and descendant queries.
TRIGGERS & RESIDENT PODS
Pods that wake other pods.
Pods run to completion or stay resident and react continuously. Triggers fire on a schedule, a signed webhook, a manual click, or a causal pattern over the live event stream. A violation in one pod can wake a resident remediation pod, and the combined record renders as one connected graph.
PRICING
Free for 30 days.
It's free for 30 days, then the service is just $14.99 per month for single users. Team pricing is available for teams that need to build agents together.
DUAL-VIEW EDITOR
Canvas plus notebook, with a real preview loop.
Click any node to open it in a Monaco editor. Run Block executes just that block against pinned upstream artifacts, cached until a source changes. Crew previews run under a token budget. Every save writes canonical YAML, so git diffs stay minimal, and editing files directly on disk works too.
PROVIDERS & SECRETS
Anthropic, OpenAI, or local. Keys never leak.
Models are provider/model references resolved against an encrypted vault at dispatch time. Run Anthropic, OpenAI, or any OpenAI-compatible endpoint (Ollama, vLLM, NVIDIA NIM). Secrets are write-only through the API, ride only inside sealed run dispatches, live only in worker memory, and are redacted from every event, preview, and log. A permanent CI test enforces it.
OPERATIONAL POSTURE
One binary. No telemetry. Apache 2.0.
The control plane is a single static Go binary with the UI embedded. It cross-compiles to Linux and macOS, amd64 and arm64. It collects no telemetry in any component. The hosted tier runs the same binary and never executes your code.
The differentiator deserves its own page.
Constraints, the causal overlay, and pods waking pods all come from one idea.
Read about causality