Chase down truth.
Recursively.
Give your agent a research record it can resume, with checks tied to the evidence behind each finding.
Install MidcenturionFor decisions
you’ll revisit.
A good prompt can guide thorough research. Midcenturion adds saved evidence and acceptance rules enforced by code.
| The work | Prompting alone | With Midcenturion |
|---|---|---|
| Accept a finding | Ask the agent to verify its conclusion. | The record cannot accept it until required checks and any required review pass. |
| Change the evidence | Ask the agent to revisit affected conclusions. | Changing a linked source, candidate, or acceptance rule marks earlier acceptance stale. |
| Pick it up later | Carry the relevant context into the next conversation. | Resume saved sources, alternatives, check results, and open questions from the same local store. |
Use it for decisions that need repeated testing, revision, or a handoff. For a quick lookup, a prompt may be enough.
Perfect ranking.
Missing results.
Does fetching the nearest vectors, then filtering for access, return the best results a user can see?
Filtered vector search
210 synthetic cases
10 seeds · 3 correlations
“Fetch the top 100.
Filter. Return the best 10.”
We tested a fixed shortlist against exact nearest neighbors within each user’s eligible set.
Recorded synthetic test of retrieval correctness. It does not measure production speed or compare research agents.
Data & methodologyTry a question
Guided examples
Next check
Add it to
your agent.
Runs locally with your agent’s model and search tools.
MCP server + skill
Includes the source and tests.
Download v0.1.0Tested on macOS. Uses a Unix-style installer.
Linux requires Python with venv support.
Open a terminal in the folder containing the download.
tar -xzf midcenturion-0.1.0.tar.gz
cd midcenturion-0.1.0
python3 scripts/install_local.py
codex mcp add midcenturion -- \
"$HOME/.local/bin/midcenturion" serveRequires the Codex CLI. The installer copies the skill into your Codex skill directory.
Questions
Will it find better answers than a good prompt?
We haven’t demonstrated a general advantage in answer quality. Ordinary research and the original skill both passed our initial comparison. The verified benefit is the saved record and enforced checks. Read the evaluation notes ↗
Does this replace my agent?
No. The skill gives your agent research instructions. The MCP server stores the work and runs configured checks. Your agent still searches, reasons, and judges the evidence.
What happens when I close the agent?
Research stops; the investigation stays saved. An agent connected to the same local store can resume it.
What does “accepted” actually mean?
The finding met its configured checks and review requirements for specific inputs. Those checks can be narrow: matching a quote doesn’t establish its truth, and the agent can review its own work. Changing linked evidence makes that acceptance stale.
Do I need to host an MCP server?
No. Your MCP client starts the server on your machine. Your agent’s usual model and search costs still apply.