Concepts
Everything behind your assistant
A complete map of the context and workflows that let your assistant take informed initiative through ContextStream's production MCP.
The production Rust MCP is how your assistant reaches ContextStream. The platform is what makes that connection useful: indexed source, durable project knowledge, written operating material, captured continuity, media, active work, and optional collaboration with explicit access boundaries.
That is useful when you work alone, not only when you work on a team. Your second laptop, next editor, or tomorrow's clean session can recover the same project reasoning. A shared workspace extends that continuity to other authorized people and agents; it is not a prerequisite for using the product.
The product model is a partnership. You choose the outcome and the authority you are granting. Your assistant chooses the relevant investigation and permitted next steps. ContextStream keeps the project scope, evidence, operating rules, and working state available so that you do not have to direct every lookup or reconstruct the project for each new session.
The four foundations
Find the work. Search indexed source by meaning, exact text, symbol, path, or pattern. Ask project questions with citations. Inspect dependencies, call paths, cycles, unused code, complexity, and change impact. Search the text and transcripts inside screenshots, recordings, audio, and PDFs.
Keep the reasoning. Record decisions, lessons, facts, preferences, constraints, terms, and significant project events. Capture session exchanges and snapshots when you want continuity. Resume the thread later from another supported assistant.
Give work a durable shape. Save plans, tasks, todos, runbooks, ADRs, postmortems, diagrams, reminders, tickets, incidents, releases, experiments, goals, sprints, reviews, risks, backlogs, discussions, and handoffs. Each kind keeps the fields and lifecycle its job needs.
Move context safely. Package selected context in a revocable capsule, hand work to a person or agent, save a proven workflow as a skill, and connect the repositories and services where useful knowledge already lives.
The partnership loop
For each task, managed instructions teach the assistant to orient itself, retrieve the context that matches the work, investigate the right evidence, act within its permissions, verify the result, and preserve only what should survive.
That means “fix the retry regression, keep the public API stable, and stop before deployment” can be enough direction. The assistant can find the project's retry decision, matching lesson, current implementation, affected callers, runbook, and open task without making you name those sources one by one.
It is bounded autonomy, not an unlimited agent. Ambiguous scope, conflicting evidence, missing capabilities, and consequential external actions still require a stop, a clarification, or explicit authority. The advanced guide to the agent partnership follows the complete loop and shows how to prepare a project for progressively more independent work.
Three ways it helps in real work
Resume your own work without reconstructing it.
You investigate an intermittent sign-out on your laptop, record why refreshes
are coalesced, save the missing 429 test as a lesson, and stop with one
clock-skew test failing. A snapshot preserves the exact state. The next morning
you open a different supported editor on your desktop and ask to continue.
The assistant can retrieve the decision, warning, open task, and stopping point before it searches or edits. No team is involved; the value is continuity across your own sessions, assistants, and machines.
Make a risky production change with the reasoning attached.
Before rotating a signing key, ask where verification happens, inspect every caller and the dependency impact, read the overlap-window decision, and follow the rotation runbook. Save a plan whose tasks cover dual-key verification, rollout, observation, signer change, and old-key revocation. Set a reminder for when work must begin.
Later, the source shows what changed, the decision explains why, the runbook says how to operate it, and the plan shows what remains. Those are different questions, so they stay as different records.
Turn scattered incident evidence into a safe handoff.
A support screenshot contains a timeout, an architecture recording explains the retry queue, and a pull request thread records why immediate retries were rejected. Index and search the media, link the source evidence to the incident, and create a handoff with the current state, open question, and exact next steps.
When another person or agent needs the material, make a narrow capsule. Keep the internal logs and customer identifiers in the workspace; send only the customer-safe status, necessary evidence, and controlled code context. The handoff tracks responsibility. The capsule controls what travels.
Complete capability map
Work with less repeated instruction
| Capability | What it covers |
|---|---|
| The agent partnership | How intent, automatic grounding, evidence selection, permissions, verification, and durable follow-through fit together |
| Instruction files | The managed rules that teach supported assistants when to orient, retrieve, verify, and persist context |
| Scope | How the current workspace, project, folder, account mode, and child project are resolved before work begins |
Find and understand
| Capability | What it covers |
|---|---|
| Code search | Meaning, exact text, symbols, paths, patterns, exhaustive matches, refactor navigation, cross-project search, and deeper investigations |
| Grounded questions | Cited answers from code, decisions, documents, sessions, and a curated answer base |
| Code health and relationships | Dependencies, callers, paths, blast radius, cycles, unused code, complexity, contradictions, drift, trends, and saved baselines |
| Images, video, audio, and PDFs | Upload or link an asset, follow processing, search OCR and transcripts, inspect results, and extract clips |
Remember and resume
| Capability | What it covers |
|---|---|
| Durable memory | Facts, preferences, constraints, habits, terms, goals, risks, decisions, notes, approvals, feedback, discoveries, achievements, and status updates |
| Decisions | A choice, its reasoning, its history, and what superseded it |
| Lessons | A mistake's trigger, impact, prevention, and severity, surfaced before it repeats |
| Past sessions | Ranked recall, transcript search, snapshots, chronological history, complete session reads, and older archived history when available |
Plan, document, and operate
| Capability | What it covers |
|---|---|
| Plans, tasks, and todos | Multi-step plans, linked execution tasks, progress, ordering, lightweight todos, due dates, and attachments |
| Documents and runbooks | Twenty-two forms of durable writing: operational, engineering, product, design, research, reliability, and reference |
| Diagrams | Process flows, service sequences, class and data models, timelines, mind maps, distributions, and other text-based diagrams |
| Tickets and operational tracking | Bugs, features, chores, epics, incidents, releases, experiments, goals and key results, sprints, reviews, risks, backlog views, and discussions |
| Handoffs | A sender, recipient, scope, summary, open questions, next steps, artifacts, optional context bundle, and completion state |
| Reminders | One-time and repeating reminders, priorities, snoozing, completion, and dismissal |
Share and reuse
| Capability | What it covers |
|---|---|
| Capsules | Scoped context bundles for yourself, teammates, external agents, onboarding, support, review, and debugging—with expiry, redaction, audit, and revocation |
| Skills | Reusable instructions and action bundles that can match automatically, run by name, import, export, share, version, and retire |
| Personal and team work | Personal or shared records, workspace membership, team views, and the same permission boundary across clients |
Connect and administer
| Capability | What it covers |
|---|---|
| Integrations and repositories | GitHub, GitLab, Bitbucket, Slack, Notion, Linear, Jira, and Figma; including pull requests, issues, reviews, notifications, links, automations, and webhooks where supported |
| Workspaces and projects | Folder association, project indexing, multi-machine settings, file and index history, duplicate-project cleanup, recent changes, and team membership |
| Privacy and indexing | Eligibility, ignore rules, secret filtering, transcript controls, de-indexing, deletion, and external sharing boundaries |
| Analytics and large exports | Hosted operational charts plus paged exports and aggregates over transcripts, decisions, lessons, and documents |
| Plans and usage credits | Account limits, metered ContextStream work, and the separate model bill from your assistant provider |
| Automation and MCP access | Hosted authentication, custom MCP clients, the optional local gateway, compatibility routes, and concurrency boundaries |
Why it is separate from your assistant
Assistants change and project knowledge should not disappear with them. You might switch editors, move to another machine, or work with a colleague using a different supported client. ContextStream keeps the authorized project context outside any one assistant so the MCP can bring it to the client in front of you.
The practical version: a decision recorded from Claude Code on Monday can answer your question in Cursor on Thursday. A runbook written during an incident can surface in a later terminal session. A PDF indexed from a design review can answer a question beside the code. If you share the workspace and project, the same authorized context can help a teammate too.
You mostly don't touch it directly
Your connected assistant reads and writes this as you work. The dashboard is there when you want to browse, edit, or audit it. In normal use, you state the outcome in English: “write the recovery runbook,” “handoff this migration,” “find the diagram of the ingest path,” or “what did that PDF say about retention?” The production MCP-backed instructions teach the assistant to choose the relevant capability instead of waiting for you to prescribe the retrieval sequence.
The MCP reference is intentionally separate. It lists the exact public domains, operations, fields, and availability rules for client builders. The guides above are organized around the work a person wants done.