DEV Community 1h ago

Give your coding agent memory it can trust

Your coding agent starts every session knowing nothing about your repository. Then it guesses. Confidently. It does not know where authentication actually lives. It does not know that your billing webhooks verify signatures, or that nobody ever wrote down why. It reads a few files, builds a story, and starts editing. That is the problem I just shipped a fix for.

DevTime v0.1.2 turns local repository memory into an MCP server your agent can query. If you are new here: DevTime is a local-first Engineering Intelligence CLI. It scans a repository and helps it explain itself from evidence. No cloud, no telemetry, no code execution during scan, no AI required.

What Shipped in v0.1.2

One command: dtc mcp start

That runs a real MCP server over stdio, on your machine, exposing three read-only tools:

list_concepts - what the scanned repo supports, with confidence labels
explain_concept - the claims behind one concept, the evidence files, and what is still uncertain
get_context_pack - a governed context bundle to hand an agent before it touches code: supported claims, do-not-change-without-review paths, tests to run

So instead of your agent guessing where billing webhooks live, it can ask. And instead of getting a confident story, it gets claims with evidence paths and an explicit uncertainty list.

Try It in Two Minutes

Install with MCP support:

pipx install "devtime-ei[mcp]"

Scan your repo:

cd your-repo
dtc init
dtc scan

Hook it into Claude Code:

claude mcp add devtime -- dtc mcp start

Or any client that reads a config file:

{
  "mcpServers": {
    "devtime": {
      "command": "dtc",
      "args": ["mcp", "start"]
    }
  }
}

No repo to test on? DevTime ships a demo:

dtc demo init
cd devtime-demo-saas
dtc init
dtc scan

What the Agent Actually Gets

Here is the shape of an answer for a concept like Billing Webhooks:

Billing Webhooks is present and supported by behavior evidence. confidence 0.86
Evidence: src/billing/stripe-webhook.ts, tests/stripe-signature.test.ts
Uncertainty: no decision was found explaining key choices
Understanding Score: 58 / 100

Notice the last two lines. The tool tells the agent what it cannot prove. That is on purpose.

The Rules Did Not Change

Wiring a server is exactly the kind of moment where a tool quietly grows scope. This one did not:

Read-only. There are no write tools.
Local stdio only. There is no network listener.
No source code is returned. Claims, evidence file paths, and uncertainty. That is it.
Weak evidence produces uncertainty, not confidence.
If the repo was never scanned, the tools say so instead of making something up.

The main rule is still the same one DevTime started with: No claim without evidence. Agents are the exact audience that needs this rule. A human reviewer might notice when a summary smells wrong. An agent will happily build on it.

Where to Find It

DevTime is now listed in the official MCP Registry as: io.github.Shakargy/devtime

It is also on mcp.so and Smithery, and the package lives on PyPI as: devtime-ei

The Python import is still devtime and the command is still dtc. The name devtime was reserved on PyPI, so devtime-ei is the public distribution name.

Why I Keep Building This

AI tools write code faster than teams can understand it. The gap between "code exists" and "someone understands it" grows every sprint. Git remembers what changed. It does not remember what is true, what supports it, or what nobody ever decided. AI writes. EI remembers. That is the bet.

v0.1.0 was the scanner. v0.1.1 made it installable in one line. v0.1.2 hands the memory to agents.

What I Want Feedback On

Not "nice project". This:

DevTime got a concept wrong on your repo
A claim is too strong
An uncertainty is missing
The MCP setup failed on your machine
Your agent got a context pack that misled it

Every one of these can become a regression fixture. That is how the tool earns trust: wrong outputs become tests that can never break again silently.