Sagit
Sagit Code inside. Any model.
Your AI coding agent that remembers every project. Autonomous, always reversible. Says done only after it's verified.
Download for Mac ↓
Sagit Code inside. Any model.
Your AI coding agent that remembers every project. Autonomous, always reversible. Says done only after it's verified.
Download for Mac ↓Connect your tools. Oversee every run. Sagit Code now speaks Model Context Protocol — plug in any MCP server and its tools join the work mid-run, while status lights and live tab badges show exactly where everything stands.
Connect any Model Context Protocol server and Sagit Code uses its tools mid-run: query real databases, read issues, pull live data.
A visual panel to add servers: paste the same JSON you use in Claude Desktop or Cursor and it just works; one-click import finds configs you already have.
Connection lights, tool counts, readable errors, one-click reconnect; a broken server never blocks anything else.
Project tabs show a live dot: green while running, amber when waiting for your confirmation.
After making changes, Sagit Code runs your project's own type check; if anything broke, it fixes it before finishing.
Completed runs show at a glance that the project check passed; if issues remain, it says so honestly.
Long-running commands show a live elapsed counter, so you always know work is happening.
Works automatically on TypeScript projects with a local compiler; everywhere else, nothing changes.
Large files come in by the region, not the whole thing; the agent asks for exactly the lines it needs.
Older history is pruned and summarized automatically as sessions grow; the recent conversation is kept verbatim.
The agent searches its own session history; past decisions and pruned details are one lookup away.
Context budgeting adapts to each model's real window, from 192K to 1M.
Before Sagit Code modifies a file, the original is saved. Zero setup, invisible until needed.
One click reverts everything a run changed, or pick just the files you want; new files cleanly removed.
Restores are guarded while a run is in progress; every restore shows exactly what it will do first.
A usage bar shows how full the model's context is, before it becomes a problem.
Research roles run a full think-act-observe loop: search, verify, compute — live data in one round.
Stocks join crypto in the price feed; international tickers (like 000660.KS) handled automatically.
Fast-moving figures like market cap are calculated from live inputs, not quoted from stale pages.
Tool-carrying roles always run on mid-tier models or above.
Sagit Code keeps a local index of the whole project and finds relevant code across the codebase — no cloud, no upload, no vendor servers. Codebase understanding with zero privacy trade-off.
Ask for a change and Sagit Code locates the right files itself, even when you don't know where the logic lives — it searches, reads, then edits. Fewer steps, less back-and-forth.
The index builds itself on first use and refreshes only what changed. It respects your ignore rules and skips oversized and binary files, so it stays fast. No reindex button.
Each project gets a sagit.md; Sagit Code reads it on every run, so the agent always has the project's full context — plan, progress, decisions, gotchas.
The agent updates sagit.md after meaningful work and keeps it concise — a one-page brief, not a log.
A collapsible Project Memory panel lets you read and edit it anytime; it auto-refreshes after each run, and your edits are preserved.
Quit and reopen; every Sagit Code session continues with its full history — what you discussed, which files it touched, what you decided.
Each project tab keeps its own independent thread.
When a conversation grows too long, you get a plain "start a new session — this one stays saved", not a cryptic error.
Point it at a project and describe what you want — Sagit Code reads your files, writes code, and runs commands on any model you choose through ClawAPI. It asks when unsure, confirms before deleting, shows every diff, and stops the instant you do.
Drive Claude Code and OpenAI Codex side by side. Tab between projects on a single screen — no cd in a terminal — start, restart, and stop sessions with one click.
Claude Code reaches claude-fable-5, Anthropic's most capable model, through ClawAPI — the new default for Chat and Code Manager.
OpenAI Codex runs with your own ChatGPT subscription or API key (~/.codex/ config). Zero extra setup — bring the plan you already pay for.
Pick whichever matches how you already pay: Claude Pro / Max subscription (zero-touch), Anthropic API key (pay per token), or ClawAPI key (recommended).
One key reaches claude-fable-5, claude-sonnet-5, claude-opus-4-8, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gemini-3.1-pro, and more on the way.
Top up with USDC or USDT. No credit card. No Anthropic account required.
Auto-follows your system language. Manual override in Settings. Persists across sessions.
Install the DMG → open → paste your ClawAPI key → running Claude Code in five seconds.
Missions run in the background while you work elsewhere. A built-in review loop checks each result and reworks it until it holds up. Interrupted runs wait in the Archive — one click to resume.
Set a per-day spend limit for autonomous Curators in Settings → Curator Activity. Hot-reloads — no restart, no risk of a runaway dream.
autoDream, Skill, Memory, Knowledge — sharing scheduler, budget, and audit framework. They run in the background on their own cadence; every action lands in ~/.sagit/curator/audit.jsonl — review what they did and why.
The 100-mission cap is gone. Per-mission directory layout (~/.sagit/missions/<id>/) handles 10K+ missions; v1.3 history auto-migrates on first boot.
Manual 🔒 toggles plus 24h auto-lock on hash drift. Skills and memory partitions you've hand-edited won't get overwritten by a Curator.
Multi-role delegations now run concurrently via Promise.all. Measured 3.0× wall-time speedup on a 3-delegate batch.
CEO, CTO, Researcher, founding_engineer, founding_pm, software_engineer, product_manager and 32 others. Pick any role for any mission. Same org chart whether you're founding-stage or production-scale.
Settings → Roles. A 39-row table where you pick the model for each role, set custom fallback chains, reset to defaults. Power user paradise.
When Anthropic has issues, Sagit auto-tries GPT-5.6 Sol (OpenAI's strongest reasoning model) before falling further. high: Fable 5 → GPT-5.6 Sol → Gemini 3.1 Pro → Opus 4.8.
Scheduled tasks promoted to a first-class sidebar item. Daily briefs, BTC trackers, weekly standups — set once, run forever.
Every agent is a .skill.md file — system prompt, model tier, tool access. Read it. Fork it. Hot-reload on save. Your prompts, not a black box.
Auto mode picks the right agent. Need more? It upgrades to Expert — Planner decomposes, Workers execute, Reviewer ships.
Sagit reads its own mission transcripts and writes the lessons into typed memory files (learnings.md, decisions.md, tech-stack.md). The brain that grows with the work.
Every skill gets a 1-5 quality score after each mission. Underperformers surface as rewrite candidates — your prompts get sharper as your company gets older.
Drop a URL or a doc. Sagit compiles a structured wiki your agents reference automatically. The company's shared brain.
The 4th autonomous Curator. Feed Sagit URLs, docs, or notes — dedicated knowledge layer with LLM cleanup (URL dedup, tag normalization, stale marking). Lock anything you don't want touched.
Full-text search across missions, memory, and knowledge. BM25 ranking. Sub-millisecond. Your agents can search too — via the memory_search tool.
Bridge Sagit's memory to Claude Code, Codex, Gemini CLI, or any MCP-compatible tool. Your company brain becomes the context layer for every AI tool you use.
Drop PDF, Word, or PowerPoint into the Knowledge Wiki. Pure-JS converters extract text — no Python install, no external services. Searchable, reference-able, instant.
Export reports as PowerPoint or Word — Sagit-branded, chart-ready. One click from the dashboard.
Anthropic, OpenAI, Gemini, Ollama — add any OpenAI-compatible endpoint. Your key, your spend, every token visible.
Data stays in ~/.sagit/. Config, memory, knowledge — all local JSON and markdown. No cloud lock-in.
One click brings your skills, memories, and API keys from OpenClaw. Your OpenClaw files are never touched.
Get the DMG from the latest release. ~160 MB, signed & notarized.
Drag Sagit.app to Applications. Double-click to open.
Paste your ClawAPI key in Settings. Open any project. Running Claude Code in five seconds.