One binary, two modes. The left column is the agent itself. The other two are what each mode adds, and the marks say where each feature applies. Nothing here is planned; it is what ships today.
What runs in both modes. This is the loop that does the work.
| Feature | What it does | Hosted | Local |
|---|---|---|---|
| Plan, act, verify | The full pipeline plans the change, makes it, then runs your tests and checks the result before it reports. The simple pipeline is one straight tool loop for models that cannot hold a plan. | ||
| Tools | Read, Write, Edit, Bash, Glob, Grep, Diff, Patch, Docker, SQL, LSP, clipboard and AskUser. The agent works by calling them; it never edits by guessing. | ||
| Permission prompts | Read, Glob, Grep and basic shell commands are allowed out of the box. Everything else asks, with the exact command shown. Answer once, or once for the session. | ||
| Snapshots and rollback | Every Write and Edit snapshots the file first. /rollback undoes the last task's edits. | ||
| Project trust | A directory is trusted once, explicitly. Project-level settings from an untrusted directory are stripped. | ||
| Sandbox | Tools are confined to the project tree. Credential directories (~/.ivy, ~/.ko, ~/.aws, ~/.ssh and the rest) are on a denylist the agent cannot read or write. Shell commands that pull secrets are blocked. | ||
| Sessions | Every turn is saved. -c continues the last session, -r shows a picker, --resume-id names one. | ||
| Memory files | Markdown memories under .ivy/memory, global and per project, loaded at session start. ivy memory list, show, edit, delete, reindex. | ||
| Skills | Reusable instructions the agent loads on demand. Create your own in the project; install signed ones from the store in hosted mode. | ||
| Slash commands | /model, /pipeline, /compact, /clear, /rollback, /plan, /thinking, /history, /memory and /help, inside the terminal. | ||
| Custom system prompt | --system-prompt prepends your text to the built-in prompt, so you can steer without losing the tool scaffolding. | ||
| Print mode and CI | --print runs one task with no terminal; --npa approves every tool call. Every flag is also an environment variable. | ||
| Inline terminal UI | Scrollback-friendly, streams tokens and tool calls as they happen, shows the plan and the context use. |
ivy --local, --ollama or --endpoint. The same agent on your own hardware, with no account and nothing sent to us.
| Feature | What it does | Hosted | Local |
|---|---|---|---|
| Any OpenAI-compatible endpoint | Ollama, vLLM, llama.cpp server, LM Studio, TGI, SGLang, NIM containers. If it serves /v1/chat/completions with tool calling, Ivy runs on it. | ||
| doctor | One probe per model with the full tool catalogue attached. Tells you whether the model can drive tools, and whether it did so natively or through Ivy's fallback. | ||
| Tool calls recovered from prose | Runtimes that put the call in the message text instead of the tool_calls field still work: Ivy extracts it, fenced or tagged, in streaming and non-streaming paths. | ||
| Emit gating | The done tool is withheld until a real tool has run, so a weak model cannot report a change it never made. | ||
| Anti-lie retry | A turn whose edits all failed is re-prompted to read the file and write it properly, up to twice, visibly. | ||
| Whitespace-tolerant Edit | A snippet that differs from the file only in indentation still lands, and only when the match is unique. | ||
| Several models | --model a,b,c. The first runs every step and judges; the rest are candidates on the full pipeline's plan step. | ||
| Endpoint auth and TLS | --api-key, extra headers for a gateway, --ca-cert for an internal CA, --client-cert and --client-key for mTLS, --insecure for dev. | ||
| Detached tasks | --detach runs a task in a background process that survives closing the terminal; --attach streams it again. | ||
| Cold-start tolerance | Waits up to five minutes for the first response header so a model loading into VRAM is not mistaken for a dead endpoint. | ||
| Local compaction | /compact asks the primary model for a dense summary and continues from it; /clear starts over. | ||
| No egress of its own | The only connection local mode opens is to your endpoint. The web tools still work if the agent uses them; on an air-gapped box they fail and it carries on. |
ivy with no flag. Frontier models through a Knockout account, plus everything a platform can do that a single machine cannot.
| Feature | What it does | Hosted | Local |
|---|---|---|---|
| Frontier models, routed per step | Anthropic, OpenAI, Google, Mistral, DeepSeek and xAI, direct. A catalogue picks the model size each step deserves and escalates on hard ones. | ||
| Multi-model consensus | Several models plan in parallel, a judge picks. Fallback and chain strategies too. | ||
| Worker fleet | Long sessions run on the compute fleet, not your laptop. Close the lid; the task keeps going. | ||
| Vault and BYOK | Your own provider keys, encrypted with a password we never hold. Unlock from the CLI or the dashboard. | ||
| Cross-device memory sync | Memory files encrypted through the vault and synced between machines. | ||
| Episodic memory | Past sessions indexed and searchable, with prior episodes offered to the agent when they match. | ||
| Workspace index and RAG | ivy workspace init embeds the repo so the agent can semantic-search it. Scratchpad offload for oversized tool output. | ||
| The Brain | A semantic store the agent reads and writes across sessions and projects, with conflict resolution. | ||
| Audit trail | Every tool call indexed and searchable in natural language. Export as JSONL. | ||
| Skills store | Signed, versioned skills installed and updated from the platform. | ||
| Project templates | ivy build starts a session from a configured template. | ||
| MCP for IDEs | ivy mcp-serve lets Cursor, VS Code, Windsurf, JetBrains or Claude drive the agent. | ||
| Encrypted snapshots | ivy sync push encrypts the working tree locally and uploads it; pull restores it anywhere. | ||
| Scopes | Tag sessions by category; per-scope retention, audit and privacy settings. | ||
| Dashboard | Sessions with transcripts, pipelines, skills, templates, the Brain, credits and settings, at ivy.fightclub.pro/dashboard. | ||
| Credits and cost | Pay per use from a wallet. Cost is shown per session and per task. Hard caps stop a runaway loop. | ||
| Self-update | ivy update checks for a newer release, verifies the signed manifest and replaces the binary. |