# Firekeep

> Your agents shouldn't work like strangers. Firekeep gives people, teams, and connected AI agents one self-hosted Keep for shared knowledge, working state, procedures, coordination, and Firekeep-mediated evidence across sessions, tools, people, and machines. Shipped adapters configure Claude Code, Codex, Kiro, OpenCode, Pi, and Claude Desktop. Pi is hook-only: it ships no MCP client, so Firekeep gives it the lifecycle (briefing and recall reach its model context) but no direct-call tool surface. Any other MCP client can connect through a generic configuration; it gets the same tools and instructions, while lifecycle automation depends on the hooks each client exposes. Start personally, add teammates when context should travel, and let useful knowledge compound. Source-available under BUSL-1.1; self-hosted internal use is free for individuals and teams during early access.

Firekeep installs on infrastructure you control. Current server images target
`linux/amd64`: run them on an x86-64 Linux host, or through Docker Desktop with
amd64 container support on Windows or Mac. The client kit itself runs on Linux,
Windows and macOS. You, your team, and connected agents share one Keep: useful
decisions, working state, selected documents, procedures, and operational facts
survive across sessions, tools, people, and machines. Supported
lifecycle adapters open sessions with a briefing; generic MCP runtimes can call
the same session tools without hook automation. Recall can be nudged by recorded
session outcomes and explicit useful or not-useful feedback; current outcome
discrimination is limited. Since client 1.0.3 the Keep also volunteers: on
Claude Code every user prompt is checked against team memory and up to three
genuinely relevant, not-yet-seen memories are pushed into the model's context —
usually none; Kiro runs the same check but delivery over its userPromptSubmit channel
is unverified; pushed recalls are marked
`trigger: prompt-hook` in replay so the compliance measurement stays honest.
Since server v1.2.0, starting a session with a goal returns a prior-art block —
up to three team memories about similar past work plus other agents' in-flight
goals — on every runtime, marked `trigger: prior-art` in replay. Firekeep measured 97.7% Evidence Recall@10 on
LongMemEval-S (459 of 470 scored questions; retrieval, not generated-answer accuracy). Code and document understanding are on by default (removable in one
command); Docdex indexes the document folders a person chooses —
md/txt/pdf/docx/html/eml and conversation-shaped json (AI-chat exports,
role-labeled but not typed as claims) — into that recall, member-private by default (per-folder
visibility enforced as a hard recall filter on the server — private from other
members, not encrypted from the server operator). Maildex connects a mailbox
read-only (server-enforced IMAP EXAMINE, PEEK-only, no send/move/delete in
the code; app password member-owned in the Keep's vault, never on client
disk) and indexes recent mail member-private; provider-side deletions are
not mirrored until remove/re-add — stated in the product every time. The Keep backs itself up
nightly (cold snapshots, checksummed manifests, 7 nightly + 4 weekly retention,
`.env` included for bare-metal restore), and `firekeep backup pull` places a
verified copy on a machine you control; downloads are admin-only because
archives hold member-private data and the vault key. The instruction layer
measures itself: a live per-instruction
compliance table is computed deterministically from replay traces, with exposure
receipts and per-runtime slices, and everything unverifiable reported as unknown.
Governance is measured too — Enforced Runbooks (human-armed advise/require-ack/
block on matched commands where the client exposes an enforceable hook, opt-in)
and a Trust Ledger keyed by self-reported agent identity (visibility only) build on the
predict-then-act gateway.

## Install, source, and registry

- [Install Firekeep](https://firekeep.ai/#start): current macOS/Linux and Windows bootstrap commands for a new self-hosted Keep or a machine joining an existing one
- [Latest Firekeep Studio for Windows](https://firekeep.ai/latest/studio/windows): x64 Setup executable, resolved from the stable Studio release manifest
- [Latest Firekeep Studio for macOS](https://firekeep.ai/latest/studio/macos): universal DMG for Apple Silicon and Intel, resolved from the stable Studio release manifest
- [Public source](https://github.com/kapella-hub/FirekeepHQ): source-available product repository, issues, and release history
- [Official MCP Registry](https://registry.modelcontextprotocol.io/v0.1/servers/io.github.kapella-hub%2Ffirekeep/versions/latest): `io.github.kapella-hub/firekeep`
- [PyPI package](https://pypi.org/project/firekeep-client/): `firekeep-client`, the local gateway, adapters, hooks, and CLI
- [Release manifest](https://kapella-hub.github.io/firekeep-dist/latest/latest.json): current managed-installer version and release metadata
- [Latest signed checksums](https://github.com/kapella-hub/firekeep-dist/releases/latest/download/SHA256SUMS): current Client Kit artifact hashes with a companion [minisign signature](https://github.com/kapella-hub/firekeep-dist/releases/latest/download/SHA256SUMS.minisig)

## Docs

- [Install guide](https://firekeep.ai/docs.html): one bootstrap command for each platform — `curl -fsSL https://firekeep.ai/latest/install | sh` (macOS/Linux) or `irm https://firekeep.ai/latest/install.ps1 | iex` (Windows). Required setup asks for your agent identity and where your server is — "set one up on this machine" provisions the `linux/amd64` server stack on an x86-64 Linux host, or through Docker Desktop with amd64 support (on Windows the setup runs under WSL2 or Git Bash); the other answers take a join code, an address, or defer. It then offers a skippable prompt for another MCP client's rules file. Adapters for Claude Code, Codex, Kiro, and OpenCode are rendered either way (plus Claude Desktop and Pi, each auto-detected when its config directory exists); for any other MCP client, `firekeep install --runtime generic` prints a paste-in MCP-server JSON snippet (tools and instructions on connect, not the hook lifecycle).
- [Firekeep Studio guide](https://firekeep.ai/docs.html#studio): desktop installation, first run, Runtime Center, reviewers, Missions, essential commands, checksums, and updates
- [Claude Code + Codex shared memory](https://firekeep.ai/claude-code-codex-shared-memory.html): exact install and handoff flow for connecting both runtimes to one self-hosted Keep, including the lifecycle boundary between Claude Code hooks and explicit Codex MCP calls
- [Self-hosted MCP memory server](https://firekeep.ai/self-hosted-mcp-memory-server.html): product and architecture guide to one shared Keep for persistent memory, session state, skills, corpus, coordination, and Firekeep-mediated evidence, including current client hook boundaries and actual server requirements
- [ChatGPT + self-hosted MCP memory](https://firekeep.ai/chatgpt-self-hosted-mcp-memory.html): practical OpenAI Secure MCP Tunnel setup for giving ChatGPT a curated Firekeep memory, skill, and session surface without opening a public inbound port to the Keep
- [Product overview](https://firekeep.ai/): shared continuity across connected agents, with personal-to-organizational knowledge, working state, coordination, ownership, and evidence
- [Dexes](https://firekeep.ai/dexes.html): the dex concept — optional domain extensions that give the Keep understanding of a domain — plus the three that ship (Symdex for code, Docdex for document folders, Maildex for email) and the registry that decides which run

## Evidence

- [Instruction compliance study](https://firekeep.ai/instruction-compliance.html): we measured whether our agents follow their instructions — deterministic predicates over replay traces, honest exposure states, pre-registered experiments — and now grade whether the work itself succeeded (a self-report bound to verified identity, un-erasable), with a visibility-only detector that audits the grade against independent failure evidence
- [Case study](https://firekeep.ai/case-study.html): what changed when one team's agents got shared memory
- [AGENTS.md vs agent memory](https://firekeep.ai/agents-md-vs-memory.html): what a static instruction file structurally cannot do, and which layer to use instead
- [LongMemEval-S retrieval benchmark](https://firekeep.ai/longmemeval-benchmark.html): the current 97.7% Evidence Recall@10 result in plain language, with the exact protocol, acceptance gates, limits, methodology, and machine-readable result data

## Policies

- [Privacy](https://firekeep.ai/privacy.html)
- [License — BUSL-1.1](https://firekeep.ai/LICENSE.txt): self-hosted internal production use is free for individuals and teams during early access; production use outside that grant requires a commercial licence; each version converts to Apache-2.0 four years after release
- Security contact: security@firekeep.ai