# How do we get context into our coding agents in a consistent, quick way across the whole org?


URL: https://getunblocked.com/answers/consistent-context-for-coding-agents-across-the-org/
Published: 2026-10-26T17:00:00.000Z
Updated: 2026-10-26T17:59:32Z
Author: Brandon Waselnuk
Categories: AI Agents, Context Engineering

**Short answer:** Run one shared context source that every coding agent reads through one MCP connection, deployed by a platform team through managed settings so no engineer sets anything up. Keep rules files for repo-local conventions. Serve everything else (tickets, docs, Slack decisions, PR history) from that one source with the user's own permissions.

---
The quick and consistent way to get context into coding agents across the org is one shared source that every agent reads through one connection, deployed centrally so no engineer has to do anything. Rules files stay for the few conventions that belong in a repo. Everything else, such as tickets, design docs, Slack decisions and PR history, comes from a service the platform team runs once, with each person's own permissions enforced. **Consistent context** is the state where the same question gets the same answer from Claude Code, Codex or Cursor, whichever team or laptop asks. In brief, one source and one connection, with nothing for engineers to set up.

## Why giving each engineer their own setup stops working

The default setup is private by design, so it does not scale. Claude Code stores a local-scoped MCP server in `~/.claude.json`, and the docs say it "stays private to you" ([Claude Code MCP docs](https://code.claude.com/docs/en/mcp)). Codex reads a global `AGENTS.md` from `~/.codex` plus one file per directory from the repo root down, and stops adding files at 32 KiB combined ([Codex AGENTS.md guide](https://learn.chatgpt.com/docs/agent-configuration/agents-md)). Both are fine for one person; multiply by 60 engineers and you have 60 variants of what the agent knows.

The failure is quiet. For example, one team's agent knows the retry convention because someone wrote it down, and the next team's agent invents one. A digital-banking software company described it on a sales call as "a lot of independent experimentation going on, with mixed results and inconsistency across the teams." A second company's engineering team named the gap. Context "all lives inside each repository," and "there's no layer that would allow to make sure that everybody shares the same context." Rules files are repo-scoped on purpose, so cross-repo knowledge has nowhere to live.

## What consistent and quick across the whole org requires

Four things; drop any one and the per-engineer problem comes back.

One source, not one per repo or per person. It holds what the organization knows about its own systems: code, PRs, tickets, docs and the chat threads where decisions got made.

One connection. Claude Code, Codex and Cursor all speak the Model Context Protocol, whose spec defines hosts, clients and servers, so one server serves each the same way ([MCP specification](https://modelcontextprotocol.io/specification/2026-07-28)). Sam Younger, Engineering Manager at UserTesting, describes that setup: "Unblocked is the first MCP queried for everything we look up. It's not just checking the code — the code could be wrong. It pulls the Confluence docs, the feature planning documents, the Slack conversations. We can't use the Slack MCP for legal reasons, so Unblocked is the only way we can access that context. For an engineering manager trying to understand a platform fast, that's the difference."

Permissions kept. The agent should see what its user can see. Claude Code's managed settings let each user sign in to a provided server with OAuth, so the server answers as that person ([Claude Code managed MCP docs](https://code.claude.com/docs/en/managed-mcp)). Stack Overflow's 2026 survey, with more than 30,000 responses, found that 93% of developers say trusting AI output requires source attribution ([Stack Overflow 2026 Developer Survey](https://stackoverflow.blog/2026/10/06/the-results-of-the-2026-developer-survey-are-here/)). The shared source has to cite where each answer came from.

No per-engineer work. An HR-software company asked for exactly this: "how do we do that in a way that doesn't require every engineer to do something?" Claude Code supports it. `managedMcpServers` gives every user the listed remote servers, and `managed-mcp.json` takes exclusive control of the server set (both in the managed MCP docs above). A managed `CLAUDE.md` at a system path applies to all users and cannot be excluded ([Claude Code memory docs](https://code.claude.com/docs/en/memory#deploy-organization-wide-claude-md)). Push it once with MDM or Group Policy and a new hire's agent has the same context on day one.

## The three realistic options and what each gives up

| Option | What you get | What you give up |
| --- | --- | --- |
| Shared rules files in every repo | Stable repo-local instructions, no infrastructure | Cross-repo knowledge, freshness, every repo must adopt |
| MCP gateway or platform-team MCP | One config, one sign-in, an allowlist | The agent still searches and reconciles each source itself |
| Context engine behind one MCP | One cited, reconciled, permission-filtered answer | A vendor, an indexing window, a line item |


Shared rules files. Commit a `CLAUDE.md` or `AGENTS.md` with the build commands, naming conventions and the parts of the tree an agent must not touch. The Claude Code docs say to target under 200 lines per file; longer files consume more context and reduce adherence ([CLAUDE.md size guidance](https://code.claude.com/docs/en/memory#claude-md-files)). The [agent skills vs rules files vs retrieved context](https://getunblocked.com/blog/agent-skills-vs-rules-files/) post covers what goes where.

An MCP gateway or platform-team MCP. One endpoint fronts the Jira, Confluence, Slack and GitHub servers, and managed settings push it to everyone. **An MCP gateway** is a proxy that forwards tool calls to source servers; it does not read the sources for you. A March 2026 arXiv benchmark counted 121 tools across just five MCP servers ([arXiv 2603.20313](https://arxiv.org/abs/2603.20313)), and all of them land in the agent's context before any work starts. The agent still reconciles a March ticket against a June Slack thread on its own, so two agents behind one gateway can land on two answers. The [MCP gateway vs context engine](https://getunblocked.com/blog/mcp-gateway-vs-context-engine/) comparison covers where that bites.

A context engine. **A context engine** is a service that indexes code, PRs, tickets, docs and chat together, resolves conflicts between them, and returns one cited answer filtered by the asker's permissions. Reconciliation happens once, in the service, instead of once per agent per laptop, which is what makes "consistent" achievable. Unblocked is the one we build. In a controlled test we measured the same prompt on the same model with and without it, and tool calls fell from 80 to 29 ([same prompt, same model, different context](https://getunblocked.com/blog/same-prompt-same-model-different-context/)). The [one MCP for Claude Code, Codex and Devin](https://getunblocked.com/blog/one-mcp-for-claude-code-codex-devin/) guide has the config for each agent.

## A rollout order a platform team can follow in a week

1. Pick the shared source and write a one-page policy: what belongs in repo rules files, and what the agent asks the shared source for.
2. Deploy the source's MCP endpoint through `managedMcpServers` for Claude Code and the matching config for Codex and Cursor, with OAuth so each user authenticates as themselves.
3. Ship a managed `CLAUDE.md` with one instruction: query the shared source first for anything about why, who or history. Trim repo rules files to what is local and stable.
4. Pilot with two teams on ten real questions, fact-check the answers with the people who know, and compare across agents. Differences point at missing sources rather than at the engineers.
5. Turn on usage telemetry (`OTEL_LOG_TOOL_DETAILS=1` records which MCP servers and tools get called, per the [Claude Code monitoring docs](https://code.claude.com/docs/en/monitoring-usage)) and set a review date.

From then on the platform team owns the source; nobody else sets anything up.

## Read more

- [One MCP for Claude Code, Codex and Devin: Give Every Agent the Same Company Context](https://getunblocked.com/blog/one-mcp-for-claude-code-codex-devin/)
- [How to Centralize Context for Coding Agents Across Your Team](https://getunblocked.com/blog/centralize-context-for-coding-agents/)

## Related answers

- [Our wiki says one thing and the code does another. Which does an AI agent believe?](https://getunblocked.com/answers/wiki-says-one-thing-code-does-another-which-does-ai-believe/)
- [Everyone on the team connects their own MCP servers and context is not centralized. What fixes this?](https://getunblocked.com/answers/everyone-connects-their-own-mcp-servers/)
- [How do I give Claude Code or Copilot context from Azure DevOps, SharePoint and Teams?](https://getunblocked.com/answers/claude-code-copilot-context-azure-devops-sharepoint-teams/)