Skip to main content
Glama

Jezero

A shared, searchable library of your agents' past work.

ci License: Apache-2.0 Python 3.12+ MCP

Your agents forget. Their sessions pile up as gigabytes of transcripts on several machines, and the decisions in them are written down nowhere else. Jezero collects those sessions, turns each one into a short summary plus atoms (decisions, rationale, fixes, exact errors, commands that worked, facts, risks), and serves it all, read-only, to every agent you run over MCP.

Documentation · Tutorial · Connect an agent · Reference · Changelog

What it looks like

You ask an agent about something another agent did weeks ago, on another computer. You do not mention Jezero; the plugin taught it to look:

you    > What did we end up doing with analytics on the shop site?

agent  · search("analytics shop site")
           [412] 2026-09-03  Shop site: remove the analytics script
           [431] 2026-09-10  Shop site: analytics back, only after consent
agent  · later(412, "analytics")
           [431] newer, same subject

agent  > It changed once. On 3 Sep the script was removed [412]; on 10 Sep it came
         back, loaded only after the visitor accepts the consent banner [431].
         That is the current state.

Each item is a summary with atoms, the date it happened and its project:

[431] Shop site: analytics back, only after consent        2026-09-10 · shop
summary   The analytics script returns, loaded only after consent. No cookies before.
atoms     decision  Load analytics only after the visitor accepts the banner.
          why       Consent is required before any tracking.
          fix       The banner test ran before the script loaded; wait for the event.
          command   npm run test:consent

(Synthetic example. Jezero never ships real transcripts.)

Related MCP server: agent-librarian

Features

  • Every agent, every computer. Connectors for Claude Code, Codex, OpenCode, Hermes Agent, Antigravity and Paperclip tasks; plugins upload sessions from other computers.

  • Used by habit. The plugins teach agents to search Jezero before working, to check whether a decision changed later, and to cite what they found, without being asked.

  • Write once, read many. A small model summarises each session once; searches are plain full-text queries that take milliseconds and no GPU.

  • One SQLite file. No database server to run; a backup is one file (how).

  • Read-only by design. Agents can search but never write; each key sees only the sources it is allowed to.

  • Private by default. Secrets are redacted before anything is stored, keys are made on each computer, and the config never holds a secret.

  • An app for people. A terminal app to search, watch activity and keep every agent's plugin up to date.

Status: pre-alpha. Everything above is implemented and tested; email comes next.

Install

Requires Python 3.12+ on Linux or macOS.

uv tool install jezero      # or: pipx install jezero, or: pip install jezero
jezero --version

To run it as a server from PyPI (config, systemd units, and the plugin bundle downloaded from the GitHub release with its SHA256SUMS), see Run Jezero as a service. Upgrade with uv tool upgrade jezero.

Get started

From a checkout, with uv:

git clone https://github.com/MarsGotta/jezero.git && cd jezero
uv sync
printf '[storage]\npath = "data/jezero.db"\n\n[server]\nhost = "127.0.0.1"\nport = 8765\nrequire_auth = false\n' > jezero.toml
uv run jezero check-config
uv run jezero init
uv run jezero import-jsonl examples/items.example.jsonl --source demo
uv run jezero search "migration rollback"

That jezero.toml is the tutorial's minimal local library; jezero.example.toml documents every field. The tutorial takes it from there: serve the library over MCP and connect an agent, in about fifteen minutes.

Connect your agents

Each computer makes its own key and the server approves it once (how). Then the plugins install straight from this repository:

claude plugin marketplace add MarsGotta/jezero && claude plugin install jezero@jezero
codex plugin marketplace add MarsGotta/jezero && codex plugin add jezero@jezero
python3 ~/.jezero-src/jezero-client/jezero_client.py setup --agent codex   # Codex's MCP
hermes plugins install MarsGotta/jezero/plugins/hermes --enable
agy plugin install https://github.com/MarsGotta/jezero/tree/main/plugins/antigravity

These follow main. To pin a release, add its tag: MarsGotta/jezero#v0.1.0 for Claude Code, MarsGotta/jezero@v0.1.0 for Codex, tree/v0.1.0/plugins/antigravity for Antigravity; Hermes Agent pins only to a commit SHA (--ref). The Jezero app installs the plugins from its own release tag by itself. OpenCode and any other MCP client: see Connect an agent.

The Jezero app

jezero with no command opens a terminal app that searches the library the way an agent does, shows what each source has been doing, shows the exact context an agent receives, and installs or updates the plugins on this computer.

uv tool install jezero
jezero ui --demo        # look around with made-up data

More in Use the Jezero app.

Documentation

Read it at jezero.marsgotta.com, or right here on GitHub in docs/:

Contributing

Bug reports, ideas and pull requests are welcome: start with CONTRIBUTING.md. To report a vulnerability, see SECURITY.md.

Roadmap

  1. Gmail connector (read-only scope, per thread, newsletters skipped).

  2. A folder watcher for agents without hooks (Cursor) on other computers.

  3. Optional vector search with sqlite-vec.

The name

Jezero is the Martian crater that once held a lake. Its layered sediments keep the record of what happened there, and the Perseverance rover sets samples aside in it so they can be picked up later. That is the whole idea of this project.

License

Apache-2.0

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables agents to run cross-runtime, cross-machine BM25 full-text search over historical conversations from Hermes, Google Antigravity, OpenClaw, Codex, and Claude Code, and to retrieve any past session's full transcript by ID. Queries return only ranked matching snippets, so recall costs no extra tokens or embeddings.
      MIT
    • A
      license
      Not graded
      quality
      A
      maintenance
      Gives AI agents a shared persistent memory by ingesting Obsidian vaults, code, transcripts, web clips, and session or bug-fix captures, then serving token-optimized context briefings, keyword and semantic search, and cross-source synthesis. Enables multiple agents to read and write the same growing knowledge base so each session builds on the last.
      10 npm
      MIT
    • A
      license
      A
      quality
      B
      maintenance
      Enables MCP-capable coding agents to semantically search, store, and inspect a single local, cross-tool index of session histories and project docs. Agents can retrieve what other tools and prior sessions already discovered.
      3
      MIT