# Firekeep — self-hosted continuity for AI agents and teams ## Your agents shouldn't work like strangers. Firekeep gives your team and connected agents one self-hosted Keep for shared knowledge, live context, handoffs, and inspectable work history across sessions, tools, and machines. **Claude Code, Codex, Kiro, OpenCode, and other MCP clients** can work from the same durable context. Where a client exposes supported hooks, Firekeep adds automatic briefings and lifecycle guidance. [Install Firekeep](https://firekeep.ai/#start) · [See it in daily work](https://firekeep.ai/case-study.html) · [Download Studio for Windows](https://firekeep.ai/latest/studio/windows) · [Download Studio for macOS](https://firekeep.ai/latest/studio/macos) · [Read the technical docs](https://firekeep.ai/docs.html) · [View the source](https://github.com/kapella-hub/FirekeepHQ) Self-hosted · MCP-native · Personal + team knowledge ### One local gateway. One self-hosted Keep. Claude Code, Codex, Kiro CLI, OpenCode, and other MCP clients keep their own interfaces. On each connected machine, the local Firekeep gateway links them through MCP to a self-hosted Keep carrying: - **Knowledge:** decisions and procedures. - **Working state:** goals and progress. - **Coordination:** tasks and handoffs. - **Evidence:** Firekeep-mediated replay and outcomes. The local gateway handles MCP, identity, and routing. Automatic lifecycle behavior depends on the hooks each client exposes. The stock Keep server binds to `127.0.0.1` by default. ## Every session starts with more. Firekeep carries selected decisions, procedures, and working state between the agents and people you connect without replacing the tools they already use. - **Start informed.** Bring relevant decisions, project history, and procedures into the next session before the background has to be rebuilt. - **Stay oriented.** Goals, plans, decisions, and progress remain available through context compression, handoffs, and new machines. - **Finish with confidence.** Keep a useful lesson with provenance so the next teammate or agent can use it in context. Less reconstruction. Clearer handoffs. Each useful session improves the starting point for the next. Supported runtime hooks can deliver briefings and lifecycle guidance automatically; other connected MCP clients recall the same Keep through its tool surface. [How to share memory between Claude Code and Codex](https://firekeep.ai/claude-code-codex-shared-memory.html) ## Your knowledge becomes your team's memory. Start with your own projects and machines. Add teammates to the same complete product when context should travel across people, roles, and tools. - **Personal — carry less in your head.** Your decisions, active goals, selected folders, and hard-won fixes stay available across connected sessions and machines. - **Shared — give the team one thread.** Project context, procedures, handoffs, and evidence remain connected to the people and projects they came from. - **Organizational — let knowledge compound.** Chosen runbooks, internal documents, code structure, and operating history return inside everyday AI-assisted work. One place. More contributors. A richer starting point for every connected agent. Folders added through Docdex are member-private by default, even on a shared Keep; share with the workspace only when you intend to. ## Firekeep Studio — your AI work, in one clear place Firekeep Studio brings Codex, Claude Code, Kiro CLI, and Grok into one desktop workspace. Choose who leads, bring in a fresh reviewer, and turn important goals into Missions with visible checks and human acceptance. [Latest Windows installer](https://firekeep.ai/latest/studio/windows) · [Latest universal Mac installer](https://firekeep.ai/latest/studio/macos) · [Installation and quick start](https://firekeep.ai/docs.html#studio) The latest stable Studio supports Windows x64 and universal macOS for Apple Silicon and Intel. 1. Install and open Studio. 2. Choose the project workspace. 3. Pick Codex, Claude Code, Kiro CLI, or Grok as the primary agent and connect its provider account if needed. 4. Run `/doctor`, send the first prompt, then add a reviewer or create a Mission. - **Pick up the thread.** Named local sessions and a chosen workspace make the next turn easy to find. - **Keep roles clear.** One explicit primary leads. Fresh, read-only reviewers add perspective without taking over. - **Finish with evidence.** Missions pair the work with deterministic local checks, bounded repair, review evidence, and human acceptance. - **See every runtime clearly.** Runtime Center shows account, connection, review, and Keep status across supported agents. Studio uses your installed Codex, Claude Code, or Kiro CLI. Grok connects directly with an xAI key protected by your operating system. Studio is an optional desktop console alongside the Client Kit. Each runtime keeps its own provider-native continuation, and Studio schedules active runs one at a time. Runtime Center shows memory and hook status separately for each runtime. Grok stays provider-connected inside Studio rather than connecting to the Keep. Installers and source are public for inspection. [Latest release notes](https://firekeep.ai/latest/studio/release) · [Published checksums](https://firekeep.ai/latest/studio/checksums) · [Explore the Studio source](https://github.com/kapella-hub/FirekeepHQ/tree/main/studio) ## Your knowledge. Your infrastructure. Connected MCP agents call one local gateway. Behind it, your self-hosted Keep carries working knowledge, session state, procedures, coordination, and evidence across the tools you already use. 1. **Your agents.** Claude Code and Desktop, Codex, Kiro, OpenCode, and other MCP clients keep their existing interfaces. 2. **Local gateway.** One `firekeep` stdio gateway on each connected machine handles identity and routing, client configuration and lifecycle hooks, local Symdex code intelligence, and the Decision Board. Studio remains a separate optional console. 3. **Your Keep.** The self-hosted server provides Cortex for memory, retrieval, and skills; Bridge for sessions and replay; Relay for tasks and coordination; and Sentinel for environment events. - **You choose where it lives.** The server, datastores, embeddings, and default inference run on infrastructure you control. - **You choose what connects.** Model providers and optional integrations follow the services you configure. - **One shared doorway.** MCP connects the gateway to the Keep; supported client hooks add automatic briefings and lifecycle guidance. [Explore the self-hosted MCP memory server](https://firekeep.ai/self-hosted-mcp-memory-server.html) ### Knowledge, state, practice, and proof Each part stays connected to its project and provenance, so the right context can return without becoming a pile of disconnected notes. 1. **Knowledge — keep what the work has learned.** Decisions, facts, proven fixes, and selected documents stay connected to their source and relationships. 2. **Working state — return to where the work stands.** Goals, plans, decisions, progress, and file notes carry through compression, handoff, and tomorrow's session. 3. **Procedures — reuse the way a problem was solved.** Reusable Skills and document-derived drafts turn hard-won fixes into reviewable team practice. 4. **Coordination — give shared work clear ownership.** Tasks, messages, presence, handoffs, and fenced resource leases help agents and teammates coordinate. 5. **Evidence — see what happened and why.** Firekeep-mediated replay, declared outcomes, and evaluation receipts make important work inspectable. 6. **Domain context — bring the world around the prompt.** Docdex covers selected documents, Symdex maps code, and Maildex adds member-private, read-only mail recall. [Explore Dexes](https://firekeep.ai/dexes.html) ## We publish the passes. The misses. The unknowns. Firekeep is measured in real work across development and operations. Baselines stay baselines, uncertainty stays visible, and the methods and result data are open to inspect. - **LongMemEval-S retrieval benchmark.** Firekeep scored **97.7% Evidence Recall@10**: 459 of 470 scored questions retrieved an evidence-bearing session in the first ten results. This measures retrieval, not generated-answer accuracy. - **Five-month case-study snapshot.** The study's latest month contains 90 tracked sessions, 45 active skills, 12 indexed internal documents, and 845 commits. Weeks after a difficult deployment fix, the relevant procedure returned in a pre-flight briefing to an agent that had never seen the original incident. - **Published instruction baseline.** **18 of 32 sessions (56%)** met the frozen recall-before-answering predicate. A newly shipped action-declaration instruction began at **0 of 32** and later moved to **1 of 35**. Exposure receipts now separate exposed, not exposed, and unknown sessions; unknown stays unknown. [Read the five-month case study](https://firekeep.ai/case-study.html) · [Read the instruction study](https://firekeep.ai/instruction-compliance.html) · [Compare AGENTS.md and memory](https://firekeep.ai/agents-md-vs-memory.html) · [Read the benchmark report](https://firekeep.ai/longmemeval-benchmark.html) · [Inspect the current result data](https://firekeep.ai/benchmarks/longmemeval-s-20260828.json) ## Install Firekeep. Make tomorrow easier. Give your decisions, documents, procedures, and working context a durable place to live—then connect the agents and teammates who should build on them. macOS / Linux: ```sh curl -fsSL https://firekeep.ai/latest/install | sh ``` Windows: ```powershell irm https://firekeep.ai/latest/install.ps1 | iex ``` [Open the installation guide](https://firekeep.ai/docs.html#server) · [Read the technical docs](https://firekeep.ai/docs.html) Firekeep is free for self-hosted internal use by individuals and teams during early access. No Firekeep account or licence key is required. Firekeep is source-available under [BUSL-1.1](https://firekeep.ai/LICENSE.txt). Each release converts to Apache-2.0 four years after publication. Security contact: security@firekeep.ai