A TOOLKIT FOR CONTEXT-AWARE AGENTS

Build agents that know your systems and can work with them.

Combine searchable knowledge, live tools, and controlled execution. Run agents in your own workflows—or bring Context Harness context into the AI assistant you already use.

Build an agent Use with your assistant

Explore the examples

Give your agents context, capabilities, and control.

Choose the knowledge they can consult, the tools they can call, and how their work runs.

Context they can search

Index selected files, Git repositories, and other configured sources. Retrieve relevant passages and source documents without loading the entire collection into every conversation.

Keyword search works without embeddings. Add semantic retrieval when it helps.

Tools for live work

Connect current information and useful actions through selected tools, including configured stdio MCP servers. Let a runtime agent look things up or request an action during a task.

Tool availability and execution depend on your configuration and permissions.

Execution you can inspect

Run bounded model-and-tool loops with explicit permissions, approvals, execution limits, and recorded activity. Submit durable tasks when an application needs background work.

Unattended workers deny operations that require interactive approval.

Run your own agents. Or bring your assistant.

The context foundation is reusable. Choose who owns the conversation and execution.

Own the execution

Configure a standalone agent, or embed the runtime in your application. Combine indexed context and selected tools, set execution policy, and inspect the resulting activity.

Build an agent Embed the runtime

Context Harness runs the model/tool loop. Your application supplies its domain rules, events, and result delivery.

Keep your existing assistant

Expose indexed context over MCP and use reusable profiles for starting instructions. Your compatible client can consult that context alongside the other tools it already uses.

Connect your assistant Explore profiles

Your client owns the conversation, model, edits, and permissions. Connecting MCP does not automatically activate a profile.

Put the pieces to work.

Different workflows use the same building blocks: relevant knowledge, selected tools, and a clear execution owner.

Illustrative workflows — not captured runs or bundled integrations.

Investigate a change across repositories

Task: “Add a field to order events and carry it into the data lake. What needs to change?”
Context: Source and guidance from the producer, shared event framework, lake contracts, and ingestion pipeline.
Tools: Search the configured index, then inspect current files with the tools available to your assistant or runtime agent.
Useful output: A change checklist covering the producer schema, framework validation, lake compatibility, and ingestion mapping, with source references to verify.

The assistant still needs to inspect current implementations, load applicable repository instructions, and test the change.

Set up multi-repo context

Prepare a response when a request arrives

Task: When a support escalation arrives, gather the relevant guidance and current ticket details, then prepare a response for review.
Context: Indexed runbooks and product documentation, combined with current details from a configured external tool.
Tools: A trusted application accepts the event and submits a durable task. The agent searches guidance, calls the selected live-lookup tool, and produces a draft.
Useful output: A reviewable response with supporting sources, unresolved questions, and a proposed next action. Publishing requires a separately authorized tool or application step.

You supply the event receiver, authentication, external integration, and result delivery. A worker must be operated separately; required interactive approvals are denied.

Build an application workflow

Start with one useful task.

Choose the context it needs and the tools it may use. Start with a standalone agent or connect your existing assistant, then expand the workflow as you verify it.

Build an agent Connect your assistant

Download for your platform

Keyword search needs no model or API key. Runtime agents need a configured model: the built-in generation adapters support OpenAI and local Ollama. Cloud models receive the context included in their requests.

Installation options Try the search demoBrowser retrieval only; no agent execution.

Before you build

Is this a context layer or an agent runtime?

Both. Context Harness provides searchable context and an optional runtime for executing agents with selected tools. Use that runtime directly or inside an application, or expose context to an existing AI client over MCP.

Can agents use live information and take actions?

Yes, through configured tools and permitted operations. The runtime includes support for selected stdio MCP tools. External-service compatibility depends on the server and authentication setup; direct remote MCP transports and OAuth are not provided by this client path yet.

Can a webhook trigger work?

Your application can receive and verify a webhook, filter the event, and submit a durable task. Context Harness does not provide a built-in webhook receiver. A separately operated worker executes accepted work; your application owns result delivery and publication policy.

Do I need a separate database service?

No. Context Harness stores its knowledge index and runtime history in SQLite, with no separate database service required. Configured models, clients, connectors, and tools may send data over the network. Local storage does not mean every operation stays on the machine.

Does search automatically refresh files or apply repository instructions?

No. Search uses the indexed material; sync configured sources when you need to refresh them. Including AGENTS.md or CLAUDE.md makes those documents searchable, but does not automatically apply their rules. The client or workflow must load the applicable guidance.

Do permissions and history mean every run can recover safely?

No. Eligible interrupted runs can resume after recovery checks, but not every run is resumable. MCP-backed runs currently do not support resume. Inspect external state before repeating an operation whose outcome is uncertain; recorded activity is not a guarantee of correctness or exactly-once effects.

Can I use Claude or another existing assistant?

A compatible client can use Context Harness over MCP while keeping its own model and execution controls. The built-in runtime does not currently include a native Anthropic generation adapter. Client integration and runtime provider support are different paths.

Will this reduce token costs?

Focused retrieval can reduce how much material you choose to include, but savings depend on the task, model, client, and tool responses. Embedding reuse and model-input token costs are separate. Compare representative workflows; no fixed reduction is promised.