Context engine for AI coding agents
Your team's rules, decisions, and lessons live in one place. Your agent asks and gets back only the pieces that fit the task, each with a source. And you can run the same question in the console to see exactly what it got.
claude mcp add --transport http contexel https://contexel.ai/mcp
Sign in to the console to get your exact connect command. MCP is the open standard agents use to plug into tools.
The top matches from a real get_context response in the demo workspace. Pick a question to see what an agent gets back, with a source on each piece.
95% vs 46%
In a 2026 study of coding agents, documented decisions were followed 95% of the time when the relevant context was delivered with the task, and 46% of the time when it was not. Contexel's one job is delivering it.
Source: arXiv 2605.08112.
The problem
Agents drown in stale CLAUDE.md files, notes that leak from one project into
another, and rules that contradict each other with no way to tell which one wins. So you paste the
right background by hand, every session. Contexel turns that into a question your agent asks, instead
of a ritual you repeat.
How it works
Open the console, then point your agent at your workspace address. You do this once.
Say something like "what do we know about rate limiting?" The tools have plain-English names, and the service briefs your agent the moment it connects. Nothing to paste.
Sorted to the right project, with conflicts already resolved, a source on every piece, and trimmed to fit your agent's size limit.
When you would reach for it
Those facts live as small Markdown notes; your agent calls get_context and gets back only the ones that fit the task, with a source on each. The third session starts as sharp as the first one ended.
One shared workspace instead of a dozen slowly diverging CLAUDE.md copies. When rules disagree, precedence decides the same way for everyone. Change a rule once and every agent has it on the next call.
Paste the repo's existing CLAUDE.md into the console and it splits into routed docs your agent pulls on demand. Its first get_context call returns the maintainers' knowledge, with a source pointing back to where it came from. (Read-only GitHub repo mirroring runs from the CLI today; connecting a repo from the console is on the way.)
Each client is its own project (a realm), kept in a separate store rather than behind a filter, and tokens are scoped per project and fail closed. Kept apart by design, not by discipline.
Real answers: encryption at rest you can turn on, per-project tokens that fail closed, a tamper-evident audit log, and writes that land only through human-gated review as git commits. You choose who holds the encryption key (see below).
What it does
Type any question in the console and get back the exact context an agent would, each piece with its source. No more guessing whether your rules are actually reaching the agent.
Work, personal, one space per client. Each is a separate project (a realm) in its own store, not behind a filter that could slip.
When two instructions disagree, Contexel picks the one that should win (company rule first, then team, then your own) and shows you why.
Every piece of context says where it came from, and the bundle is packed to fit your agent's size limit. If anything had to be left out, it tells you.
Your agent can save a note (observe) or draft a lasting change (propose) that a person approves (review). Auto-save is off by default, and secrets are caught before anything is written.
Point Contexel at a flat CLAUDE.md and it maps into routed context. No rewrite, no new format. Meeting transcripts arrive as review-queued drafts, and everything passes secret redaction first.
Your content is plain Markdown that you own; the search index is built from it and can be rebuilt any time. Export whenever you want.
Encryption at rest
Turn on encryption and your documents and your search index are encrypted at rest under a key dedicated to your workspace.
By default Contexel holds the key so it can run indexing and search for you. That is custodial: we can technically access your data, and we say so plainly.
Switch to a key only you hold and supply it each session. It lives in your workspace's memory only while unlocked, never on disk. When the workspace is locked, Contexel has no way to decrypt your documents or your search index. Lose the key and the data is permanently unrecoverable.
Enable, rotate, switch to your own key, or export your data from the console. Backups taken before you switch may retain a copy until they expire.
Straight talk
Get started
Sign in with your browser using Google, GitHub, Microsoft, or your company SSO. No install. Your whole team shares one workspace, and every workspace is walled off from the others.
Open the consoleclaude mcp add --transport http contexel \ https://contexel.ai/mcp \ --header "Authorization: Bearer <YOUR_TOKEN>"
Mint a token in the console and this line shows up filled in, ready for Claude Code, Cursor, or any MCP client. Where browser sign-in (OAuth 2.1) is turned on, add the address with no token and sign in from your editor.
Then just talk to it. Ask "what do we know about rate limiting?" and your agent calls get_context. Say "remember this for next time" and it calls observe, and a person approves anything lasting through review. Not sure what to say? There is a help tool for that.
Hosted for your team. Free during the beta. Sign in and connect in minutes.