For developers and operators building agent workflows
What is Contextful?
Contextful is a local-first context engine for AI agents and teams. Contextful collects source data, keeps versioned data in a store you control, and serves that data to agents.
Agents query rows, retrieve relevant material and use memory derived from stored data. Agents connect through Model Context Protocol (MCP) or HTTP. Developers and operators use the command-line interface (CLI) to run and inspect the engine.
Your application supplies prompts, business rules and the user interface. Contextful supplies data ingestion, storage and controlled reads.
When Contextful fits
Use Contextful when your agent workflow needs one or more of these capabilities:
- Context from changing sources. Collect documents or structured records, update the stored data and retrieve relevant material for an agent.
- Versioned data under your control. Keep source data in files, retain snapshots and sync a replica through object storage.
- Access controls shared by clients. Apply caller-specific table, row and column restrictions when agents or applications read data.
A local-first deployment reads from a local store. Source connections, model calls and optional object-storage sync use configured network services.
How Contextful works
- Connect a source. A connector reads a source through the engine's credential and network controls.
- Run a pipeline. A pipeline declares the ingestion work. Each run executes the pipeline and records progress, retries and failures.
- Commit data to the store. The store keeps versioned Parquet files and manifests describing committed data. Parquet is a column-oriented file format.
- Read from an agent or application. Clients query rows, retrieve relevant passages or recall memory under the caller's access rules.
For example, a research agent reads reports that a pipeline has collected. When reports change, the pipeline updates the store. The agent retrieves relevant passages from committed data under the caller's access rules.
Choose a guide by task
Read the guide that matches your next action. Each guide explains a flow and links to its exact rules.
| Your task | Read this guide |
|---|---|
| Connect a data source | Connectors and credentials Read before choosing a connector, binding credentials or declaring network access. |
| Run ingestion and keep data current | Runs, pipelines and derive Read when configuring pipelines, schedules, retries, recovery or historical imports. |
| Query data from an agent | Reading and memory Read before using SQL, ranked retrieval, MCP tools or synthesized memory. |
| Manage history, snapshots or sync | The store and its sync Read when choosing table keys, reading earlier data, setting retention or creating a replica. |
| Control who reads which data | Authority and enforcement Read before issuing credentials or restricting a caller's tables, rows, columns or inference zones. |
| Preserve source permissions and audit reads | Visibility, disclosure and accountability Read when mirroring source access, releasing aggregates, tracing reads or handling erasure. |
| Deploy on AWS | AWS Fargate reference deployment Provision Fargate tasks, S3 and RDS PostgreSQL from the OpenTofu configuration, step by step. |
| Deploy on Cloudflare | Cloudflare demo Run the engine, a scheduled pipeline and the operator console in one Container behind Cloudflare Access, with the store in R2. |
| Operate the console and team configuration | Operator surfaces Read when using Query and Admin, changing the control document or publishing a hosted interface. |
| Test or extend the engine | Assurance Read before changing engine behavior, adding evidence or interpreting a gate failure. |
Use the specification
Open the specification when a guide's explanation leaves a question about exact behavior.
- Contracts define operations, error cases and limits. Read a contract when implementing or checking a specific behavior.
- Contract cards provide a compact lookup of operations, errors and bounds.
- Architecture decisions explain the alternatives and costs behind a rule.
Read Corpus law before editing the specification. Use the milestone roadmap to understand the implementation order.