Context layer for agents

Context worth passing around.

Taco it once — so any human, agent or session can continue it.

claude mcp add --transport http contextaco https://mcp.contextaco.dev
Run once per machine, then /mcp in Claude Code and choose Authenticate. Your workspace is https://mcp.contextaco.dev.
evaluated byevaluated bysold tostudiedstudiedai-coding-assistantsCursor, Copilot, the agentsmetr-rct-202519% slower; felt 20% fasteraccenture-rct+26% PRs — a different metricoss-maintainers16 devs, 5 yrs on the repoenterprise-teamsAccenture engineersjunior-developersnobody has measured them
toolstudypopulation
How it works

Four verbs, and your agent does three of them.

Capture

Your agent writes what the session worked out — the decision, the evidence, the thing that failed — while it is still working.

Load

The next session, or the next person, starts holding all of it. No re-reading, no re-deciding.

Fork

When the intention diverges, take a copy and make it yours. The origin is recorded, and nothing you do can touch it.

Continue

The point of all of it. Work resumes instead of restarting.

The problem

Your agent forgets. Your taco doesn’t.

A session ends and what it worked out ends with it. A taco does not belong to a session — or to a client, or to one person. Your agent writes it while the work happens, and whoever comes next reads it in one call.

capturesthe agent that did the work writes it down
loadsany other client reads it, whole
forkswhen the question diverges, it becomes yours
Claude CodeCursor
captures, as the work happens
does-ai-make-developers-fastermira · 6 entities · 3 types · public
loads it whole, or forks it
a new sessionChatGPTCodex
The history

Every version keeps the reason it changed.

Every write is a checkpoint, and a checkpoint carries the sentence explaining it. Nothing is overwritten quietly: the version that was wrong stays readable, next to the reason it changed.

Which is the half that usually goes missing. A file tells you what someone decided. A history tells you what they tried first.

See how a taco changes →
7f2a91Nobody has measured juniors — added the population with no study on it3 entities2d
4c81deAccenture measures PRs merged, not time. Kept it, retyped it.2 entities5d
b93c07Every population reaches the tool. Only two reach it through a study.1 entity1w
7f2a91changedaccenture-rct
Accenture found +26%, so the METR result does not replicate.
Accenture measured pull requests merged; METR measured time end to end.
Two metrics, two questions. Neither result touches the other.

Never explain the same context twice.

Private by default. Publish when you choose, and only what you choose.

Connect an agent