04 / 05Use case · agent builders

Give agents the context they need before they act.

ContextStream lets agents query decisions, lessons, docs, code context, and project history through MCP, Capsules, and scoped context layers.

Read MCP docs →
01 The agent context problem

Bigger windows aren't the answer. Right context is.

Agents that act without project history make plausible mistakes. They re-derive decisions, ignore guardrails, repeat lessons. ContextStream gives every agent a scoped, source-backed slice of the project before it commits to an action.

02 Lessons-aware guardrails

The lesson the team learned is the guardrail the agent sees.

◆

Lessons

Captured by humans, scoped, citable. Promote into guardrails on the surfaces they protect.

⊘

Guardrails

What an agent must not do. Returned with every traversal that touches the protected nodes.

⊡

Audit + recall

Every agent action records what context it saw, what it cited, and what it changed.

03 Project graph through MCP

Traversal, not retrieval.

Agents query the Atlas the same way ContextStream Code does: by node, by relation, by scope. Right context, not most recent.

Agent: "What touches auth/session.ts within 3 hops?"
→ contextstream.graph.traverse(node="file:auth/session.ts", hops=3)
✓ 7 nodes · 11 edges · 1 active guardrail
Agent: "Open the capsule from yesterday's handoff."
→ contextstream.capsule.open(id="cap_8f3c")
✓ opened · scope=project · expires in 7d
04 Scope and permission roadmap

Permissioned context, on the way.

◈

Agent-visible scope

Today: explicit subset of the workspace an agent can read.

⊘

Per-action constraints

Roadmap: agent actions guarded by lesson-derived constraints with audit.

★

Verified Capsules

Roadmap: cryptographic signing for handoffs that must be agent-verifiable on Enterprise.

Build agents that walk in with project context.

MCP server + ContextCapsules. Right context, not most recent.

Read MCP docs