v2.0 · MIT licensed · No telemetry

Ship features. Not context.

CodeFire is the desktop workbench for developers running AI coding agents. Wiki memory, a hybrid context engine, a multi-agent task board, a browser, and MCP — one place for Claude Code, Codex, and OpenCode to live, work, and hand off.

Team access
codefire://install-mcp?client=claude  ·  one-line MCP setup
CodeFire app — integrated browser, agent chat, and project workspace
Scroll
Workbench for —
CLAUDE CODE
CODEX CLI
OPENCODE
+ ANY MCP-AWARE AGENT
CLAUDE CODE
CODEX CLI
OPENCODE
+ ANY MCP-AWARE AGENT
Chapter I · The workbench

Stop briefing.
Start shipping with
your whole stack.

A new dev environment for the agentic era. Wiki memory your agents author, a context engine that retrieves only what's relevant, a Kanban that spans every project, and an in-app agent chat that runs Claude Code, Codex, and OpenCode against your code — with shared task handoffs between them.

Wiki memory
Living project knowledge agents read, write, and verify through MCP.
Context engine
Hybrid vector + keyword retrieval across code, docs, and commits.
Multi-agent tasks
Kanban + cross-CLI handoffs so Claude can pass work to Codex.
Agent chat in app
Real Claude Code, Codex, and OpenCode sessions — supervised, streamed, scoped.
Chapter II · The three engines

One app. Three
engines.

01 — Wiki Memory

Knowledge
your agents author.

One-click initialization seeds Overview, Stack, Architecture, Schema, Runbook, Roadmap, and Agent Guide pages from your repo. Agents read and update them through MCP. [[Wiki Links]], backlinks, status, and verification keep it from rotting.

Markdown native [[Wiki Links]] Backlinks Verified status
02 — Context Engine

Surgical retrieval,
not file dumps.

Hybrid vector + keyword search across functions, classes, blocks, docs, and git commits — exposed to your agent as context_search. Stale results never block; the index re-verifies in the background after sessions and on a scheduler.

Vector + FTS Git history aware Wiki + code ranked Background refresh
03 — Multi-Agent Tasks

Claude starts.
Codex finishes.

Per-project Kanban with priorities, labels, and notes — plus structured handoffs between agents. The sender drops a labeled task, the receiver picks it up from its inbox, posts progress notes, and closes with a result. No copy-paste, full context preserved.

Drag-and-drop from:claude → to:codex Shared task notes
The planner

Every project,
on one board.

The cross-project Planner is your morning dashboard. See what's in flight everywhere, pull in tasks from email and mobile capture, and ask the built-in CodeFire Agent what changed today across all your repos.

  • Global Kanban across every tracked project
  • Active task feed grouped by project
  • Inbox from Gmail and mobile capture
  • Ask the planner agent: "what changed today?"
CodeFire Planner — cross-project Kanban with active tasks, emails, and planner agent
CodeFire project Kanban with agent picker — Claude Code, Codex CLI, OpenCode
Agent chat

Pick an agent.
Start working.

Launch a real Claude Code, Codex CLI, or OpenCode session against the open project — supervised, streamed, and scoped to the folder you chose. Set permissions per session, share context across agents, and hand work off without leaving the app.

Claude Code Codex CLI OpenCode
  • Per-session permission mode: Ask · Plan · Auto-edit · Skip
  • External context folders you can attach on the fly
  • Output streams back into chat — replayable
Wiki memory

The brief
that survives
every session.

Architecture decisions, schema notes, runbooks, glossaries, agent guides — written once and read by every future session. CodeFire seeds the structure from your repo, agents author through MCP, and freshness metadata flags pages that drift.

  • One-click init: Overview, Stack, Architecture, Schema, Runbook
  • [[Wiki Links]] + backlinks for connected knowledge
  • Verified / draft / stale status — trustable by future agents
  • Local-first SQLite. Optional team cloud sync.
CodeFire Wiki — sidebar of pages, Agent Guide open in markdown editor
CodeFire — Codex CLI streaming a real git workflow in the agent chat panel
Watch agents work

Real CLIs.
Real output.

CodeFire spawns the actual CLI binaries you'd run in your terminal — no wrappers, no scraping. Tool calls, shell commands, and assistant turns stream into the chat panel as they happen, ready for you to approve, pause, or pivot.

  • Approval prompts surfaced inline — not buried in a TTY
  • Per-project defaults for agent, permissions, and context
  • Full session transcripts saved and searchable
MCP server

78 tools.
One protocol.

Every CodeFire surface — wiki, tasks, notes, sessions, context search, browser, terminal, agent handoffs — is exposed to your CLI through a single MCP server. Agents don't just read your codebase; they update the project's shared brain.

Wiki · 6 Tasks · 6 Notes · 5 Context · 1 Browser · 30+ Terminal · 3 Handoffs · 3 Sessions · 2
CodeFire — Codex agent running wiki_health through the CodeFire MCP server, returning a structured audit
Chapter III · And the rest of your dev stack

Browser. Terminal.
Git. Images.

CodeFire isn't only a memory layer — it's a desktop dev environment for the agentic era. The things you'd normally tab between, in one window your agents can also reach.

CodeFire — built-in AI image generation with media history
CodeFire — per-project settings with agent defaults and permissions
Persistent browser
30+ MCP tools for tabs, navigation, screenshots, cookies, storage, and network capture.
Tabbed terminal
node-pty sessions alongside your project views, with show/hide toggle and CLI orchestration.
Git intelligence
Status, diffs, staged changes, and branch context — exposed to agents and to you.
AI image gen
Text-to-image via OpenRouter for design comps, social, hero shots — saved to project media.
Per-project defaults
Pin a default agent, permission mode, and external context per project.
Notes & briefings
Pin architecture decisions and gotchas. Get AI-generated daily project briefings.
📥
Gmail + capture
Whitelist email senders into tasks. Send tasks from the CodeFire mobile companion.
Session history
Token usage, cost, tool calls, file changes — every run, observable.
78
MCP tools exposed
3
Agents, one workbench
0
Telemetry collected
MIT
Open source license
Chapter IV · Bring your own agent

Three first-class agents.
Any MCP client.

Claude Code, Codex CLI, and OpenCode run as first-class in-app agent chats. Any other MCP-aware CLI — including Gemini CLI — still gets full access to the 78-tool MCP server.

+ Gemini CLI and any other MCP-aware agent connect through the same MCP server.

Chapter V · Official release

Access opens
from the waitlist.

official release access Step 1

We are tightening up distribution while CodeFire moves toward its official release. Join the waitlist and we will send the right signed build for your platform when access opens.

connect your agent Step 2

One-click MCP install for any supported CLI — these deep links open CodeFire and wire up your config automatically.

Official release access will include platform-specific install notes and MCP setup guidance.

Built by developers who run three agents a day and got tired of re-briefing all of them.

Join Discord