# Our engineering knowledge is spread across Slack, Confluence, Jira and GitHub. How do we give AI one place to pull context from?


URL: https://getunblocked.com/answers/give-ai-one-place-to-pull-engineering-context/
Published: 2026-10-26T14:00:00.000Z
Updated: 2026-10-26T17:56:09Z
Author: Dennis Pilarinos
Categories: Context Engine, Engineering Insights

**Short answer:** Put a context engine between your AI tools and the four systems. It connects Slack, Confluence, Jira and GitHub once, keeps each person's permissions, settles what the sources disagree on, and hands the agent an answer rather than a stack of documents. Enterprise search, a homegrown RAG or a set of MCP servers each do part of this and leave the rest to the agent.

---
Give AI one place to pull context from by putting a context engine in front of Slack, Confluence, Jira and GitHub. A **context engine** is a service that connects to those systems once and indexes them with each person's permissions intact. It works out which source to trust when they disagree, and it returns an answer the agent can act on. Your coding agents and chat assistants then ask that one service instead of searching four tools on their own. The other routes are enterprise search, a retrieval pipeline you build over the wiki, or a separate MCP server for each system. Each gets you part of the way and leaves the rest to the agent.

## What "one place" has to do

We analyzed 284 sales calls from May to October 2026. Buyers describe the problem in much the same words: knowledge is scattered across tools, there is "data everywhere, and nowhere is good enough," and the only way to learn is from the people who have the experience. One place for AI to pull context from has four jobs.

**Connect all four systems and stay current.** Slack threads, Confluence pages, Jira tickets, and GitHub pull requests and commits all have to be in the index. The Slack thread from yesterday is often the thing that corrects the page from last year.

**Keep permissions.** The person asking should only get back what they can already open. GitHub says tools in its MCP server "inherit the same access requirements as their corresponding GitHub features" ([GitHub Docs](https://docs.github.com/en/copilot/how-tos/context/model-context-protocol/using-the-github-mcp-server)). Atlassian's Rovo MCP server scopes access "to the user's existing permissions in Atlassian" ([Atlassian Support](https://support.atlassian.com/security-and-access-policies/docs/understand-atlassian-rovo-mcp-server/)). Slack's MCP server runs on user tokens that each person authorizes through OAuth ([Slack Developer Docs](https://docs.slack.dev/ai/mcp-server/)). One place has to carry all of those rules at once.

**Settle conflicts.** A Confluence page from two years ago says one thing. The pull request merged last month says another. The system has to rank by recency, by what shipped, and by who wrote it, then say which one it believes and why.

**Hand the agent the answer, not the documents.** Everything the agent receives counts toward its context window, and Anthropic's own docs warn that "more context isn't automatically better. As token count grows, accuracy and recall degrade" ([Claude Platform Docs](https://platform.claude.com/docs/en/build-with-claude/context-windows)). A March 2026 study found the same: accuracy fell consistently as context grew, and questions that needed two facts joined together lost nearly twice as much as single-passage lookups ([arXiv](https://arxiv.org/abs/2603.15723)). Forty pages dumped into an agent's window is worse than three paragraphs that are right.

## The four ways teams give AI one place to pull context from

**Enterprise search** is a tool such as Glean or Microsoft 365 Copilot that indexes company systems and ranks results for a person to read. It is good at people, policy and "where is the doc" questions across the whole company. What it gives up for engineering: it returns documents ranked for a human, treats a proof-of-concept page and the production runbook as equals, and buyers on our calls say it falls short on code questions. One of them described Claude receiving a 64K context from Glean, setting it aside, and searching it again itself.

**A homegrown retrieval pipeline.** **Retrieval-augmented generation (RAG)** is the pattern of fetching related text and placing it in the model's prompt before it answers. Teams build this over a wiki or a folder of Markdown files with a vector database such as pgvector. It works on day one. What it gives up: someone on your team now owns four connectors, a permission model that spans four vendors, and the ranking logic that decides what wins when the wiki is wrong. One buyer put it as "it worked until it didn't."

**A pile of MCP servers.** **MCP (Model Context Protocol)** is an open-source standard for connecting AI applications to external systems ([Model Context Protocol](https://modelcontextprotocol.io/docs/getting-started/intro)). GitHub, Atlassian and Slack each ship one. What it gives up: every engineer authorizes four connections, nothing ranks across systems, and every raw payload lands in the agent's context window. The agent does the reconciliation itself, on every question. Our post on why [a pile of MCP connectors is not a context engine](https://getunblocked.com/blog/mcp-connectors-are-not-a-context-engine/) goes through this in detail.

**A context engine.** This is the option that does all four jobs from the section above. What it gives up: it is another vendor, and you have to verify its permission model and its ranking rather than take them on faith. The checks below are for that.

| Option | Keeps each person's permissions | Ranks across the four systems | Settles conflicts | Hands the agent an answer |
| --- | --- | --- | --- | --- |
| Enterprise search | Yes | Yes, for a human reader | No | No, a ranked list of documents |
| Homegrown retrieval pipeline | You build it | You build it | You build it | Partly, raw passages |
| A pile of MCP servers | Yes, per connection | No | No, the agent does it | No, raw payloads |
| Context engine | Yes | Yes | Yes | Yes |


## Where Unblocked fits

Unblocked is a context engine built for this case. It connects [Slack, Confluence, Jira and GitHub, plus Notion, Linear and around fifty other tools](https://getunblocked.com/integrations/). It links each ticket to the pull requests that closed it and the threads that argued about it. Answers come back through Slack, the IDE and a single MCP server for coding agents. Jorden van Breemen, CTO at Advidi, described the before and after:

"Everyone has context living in multiple places: code, Slack, conversations. It's always hard to pull it back into one. Unblocked solved that for us: ask a question, get all the relevant context from your tools, with permissions maintained."

Glean is the right call when most questions come from outside engineering and nobody needs the code. Atlassian's own tooling is enough when your knowledge lives almost entirely in Jira and Confluence.

## What to check before you choose

Run these five tests against every option, including the one you would build.

1. Ask a question whose only answer is in a Slack thread. Then ask one where Confluence and the code disagree. See which source it picks and whether it says why.
2. Have someone who cannot see a private channel ask the same question. Check that nothing leaks.
3. Measure what lands in the agent's context window per question, in tokens. Anthropic's [token counting API](https://platform.claude.com/docs/en/build-with-claude/token-counting) counts the tokens in a request before you send it.
4. Count the connections each engineer has to authorize and keep alive.
5. Ask who owns the ranking when the wiki is wrong. If the answer is "the agent," you do not have one place yet.

Our guide to [organizing context for AI agents when it lives everywhere](https://getunblocked.com/blog/organize-engineering-context-for-ai-agents/) covers preparing the sources, and [context engine vs enterprise search](https://getunblocked.com/blog/context-engine-vs-enterprise-search/) goes deeper on the first option.

## Read more

- [How to Organize Context for AI Agents When It Lives Everywhere](https://getunblocked.com/blog/organize-engineering-context-for-ai-agents/)
- [A Pile of MCP Connectors Is Not a Context Engine](https://getunblocked.com/blog/mcp-connectors-are-not-a-context-engine/)

## Related answers

- [There is drift between our repos and our documentation and agents cannot tell. What handles this?](https://getunblocked.com/answers/documentation-drift-between-repos-and-docs-for-ai-agents/)
- [How do I mark old Confluence pages as deprecated so AI stops using them?](https://getunblocked.com/answers/mark-confluence-pages-deprecated-for-ai/)
- [MCP tools are eating the whole context window in Claude Code. How do other teams give agents company context?](https://getunblocked.com/answers/mcp-tools-filling-claude-code-context-window/)
- [We have context all over the place and constantly being generated. How do we organize it so AI pulls the right thing?](https://getunblocked.com/answers/organize-scattered-context-so-ai-pulls-the-right-thing/)