Provides evidence-retrieval tools for OpenAI Codex coding sessions, enabling local workspace investigations, bounded source-addressed excerpts, omission recovery, and optional TypeSafe Jev ranking without replacing the existing Codex coding model.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@Codex Jev check evidence_status, then find evidence for how config loading works"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Find the evidence. Keep your coding model.
Large repositories and noisy logs can fill a coding session with material it doesn't need. Codex Jev adds a retrieval layer: find candidate evidence locally, optionally rank sanitized excerpts with TypeSafe's Jev, and return bounded, source-addressed evidence with exact follow-up reads.
It does not replace your coding model, proxy inference, rewrite conversation history, or bypass Codex usage limits. Works without a Jev key in local-only mode.
Public beta. Offline safety and packaging tests pass; generalized Codex token, cost and speed savings are not yet established. Evidence-byte reduction is not an account-quota measurement. What is verified.
Capability | What you get |
Focused investigations | Query-prioritized code, tests, configuration and log evidence |
Progressive disclosure | Concise previews, source hashes, line ranges and exact reads |
Recoverable omissions | List retained, omitted and unscored candidates; inspect what was left out |
Controlled spending | Local-only by default; explicit paid activation, persistent caps and reservations |
Recovery boundaries | Private Git checkpoints, corruption checks and owner-acknowledged resume |
Honest measurements | Separate Jev, cache, bypass and fallback statistics; missing usage stays unknown |
Best suited to multi-file investigations and large diagnostics. Precise file reads, small edits, patches and test commands should keep using native tools.
Related MCP server: evidence-rag
Requirements: Node 22.13+, Python 3.11+, Git, ripgrep, and a Codex client with stdio MCP support. Recovery hooks additionally require a hook-capable Codex version. The tested deployment is a Linux execution host, including a remote host used from the macOS Codex app. Mac-local execution is unverified.
Run on the host where Codex actually reads your project. Use a permanent install directory outside the project; keep it after installation.
git clone https://github.com/Hyper-AI-Lab/codex-jev.git
cd codex-jev
git checkout v0.4.0-beta.1
npm ci
npm run build
python3 runtime/manage.py install \
--node "$(command -v node)" \
--workspace /absolute/path/to/your/git-project \
--codex-home "${CODEX_HOME:-$HOME/.codex}" \
--entrypoint distThe installer preserves your model, reasoning effort, authentication and unrelated settings. It adds an owned MCP entry, concise global guidance and recovery hooks. Existing conflicting configuration is rejected, not overwritten.
Reconnect the execution host or reload your Codex client so it reads the new MCP configuration. Review the generated hooks in the native Hooks UI and authorize them there. Installation is not hook trust. Continue the same task afterward.
Ask Codex:
Check evidence_status. Use search_workspace_evidence to investigate how this
project loads configuration. Show source ranges, recover any relevant omissions,
and distinguish local results from actual Jev selection.No paid requests occur from this quick start. To enable hosted ranking, follow the explicit key, spending-cap and retention-validation steps in Installation. The adapter is MIT-licensed; the hosted Jev service is separately billed.
MCP tool | Purpose |
| Scoped workspace investigation with bounded candidate selection |
| Query-related ranges from an eligible text or log file |
| Paginate candidate references, including omissions and unscored ranges |
| Read exact bounded ranges after hash and access revalidation |
| Loaded build, selection mode, budget, reservations and measurement coverage |
Selection currently bounds each request to 20 candidates, eight returned blocks and 48 KiB outbound. Critical overflow is recoverable through pagination and exact reads. Small packets bypass Jev; cache hits avoid a repeated paid call. Not selected does not mean nonexistent.
Workspace authorization, Git exclusions, sensitive-path denial and local redaction apply before sending evidence. Source text remains untrusted data.
Only bounded sanitized queries, requirements and excerpts go to TypeSafe, with opaque candidate identifiers, not full conversations or credential files.
One in-flight Jev request, transactional reservations and conservative unknown charges protect shared budgets. Quota responses halt covered operations.
Recovery verifies state before resuming; it never reapplies patches, repeats deployments or cancels unrelated processes automatically.
Regex redaction is not perfect. Do not enable hosted ranking for data you cannot permit a third party to process. Provider retention is governed by TypeSafe's policy.
Hooks are partial guards, not a security sandbox or universal interception layer. Already-running commands and hard process failures need explicit reconciliation. See Security and Recovery.
Installation and paid opt-in · Architecture · Recovery and rollback · Measurements and limitations · Contributing · Release notes
npm ci
npm run check
python3 -m unittest discover -s runtime -p 'test_*.py' -q
python3 -m compileall -q runtimeThe test suite uses synthetic fixtures and mocked providers; no API key, paid Jev call or native Codex inference is required. Packaging tests exercise isolated install, upgrade, rollback and uninstall without accessing real authentication.
Built by Hyper AI Lab, derived from James Cressler's Jev Codex Token Saver. Upstream MIT copyright is preserved. Full attribution. Independent project; not an official OpenAI or TypeSafe product.
Useful to your team? Share a reproducible, sanitized investigation or contribute a regression test. Correctness and transparent measurements matter more than headline compression percentages.
This server cannot be deployed
Machine-native research commons for agent evidence, discovery, rooms, and bounded research quests.
Read bounded illicit-economy evidence with provenance, privacy, availability, and federation intact.
Agent memory that refuses to guess: evidence-gated recall, exact-source reads, verifiable deletion.
Evidence-labeled Method search for AI Agents; optional contributor and Gateway discovery.