Clean handoffs for coding agents

AI agents should know where the last one stopped.

Stop re-explaining your work to every new session. Kontinuo leaves a note the next agent can actually use — what was done, where it stopped, what comes next. It runs on your machine and works with the tools you already use.

Install Read the docs

Codex Claude Code Cursor Antigravity ACP editors

<1%

Of a new session's context. A compact handoff costs almost nothing to read.

Models run by Kontinuo. Your sessions and code never leave your machine.

Auto

Checkpoints capture on their own when your editor supports it. Nothing to remember.

The gap

Agents do the work. Session boundaries erase the map.

You hit the context limit mid-refactor. The new session starts blind.

You switch agents — Claude to Codex, Cursor to Claude — and spend ten minutes re-explaining what you did this morning.

The next agent redoes finished work, misses parked work, or trusts state that is three commits stale.

How it works

A handoff note good enough for a skeptical agent.

Kontinuo does not run agents, call models, or merge code. It never sends session data off the machine. It is the small local layer that keeps the handoff honest.

Capture

Transcript intent plus workspace facts

Kontinuo reads supported agent sessions, Agent Client Protocol streams, and git workspace state, then writes one normalized checkpoint for the current repo lane.

Verify

A checkpoint you can verify safely

Each handoff records evidence, git HEAD, changed files, dirty state, and a workspace fingerprint so a reader can spot stale state before acting.

out Drift since checkpoint: 29 files changed across 12 dirs; HEAD moved c31d5ec -> bc602c8; working tree dirty

Resume

Served over MCP, readable by humans

The next MCP-capable agent calls handoff_read and gets the goal, stopping point, deferred work, verification notes, and the exact next action.

Local install

Install Kontinuo

Kontinuo installs as a local CLI and MCP server for AI coding agent handoffs.

Requires macOS or Linux. No cloud account is required. Progress prints on stderr so scripted output stays clean.

Restart or reload your agent host so it starts the Kontinuo MCP server.

Install the CLI

Install the single Kontinuo binary locally.

curl -fsSLhttps://kontinuo.dev/install | sh

kontinuo setup --host claude--cwd . --config.mcp.json--store ./.kontinuo

Verify the local store

Confirm the local store is reachable before handing work off.

kontinuo log --cwd .

From the terminal

Set up, capture, and inspect without leaving your shell.

The CLI handles setup, capture, inspection, and verification. The MCP server keeps the agent in flow: read the last checkpoint, work from it, then write the next one back.

$ kontinuo version

out kontinuo dev

$ kontinuo setup --host claude --cwd . --config .mcp.json --store ./.kontinuo

out wrote MCP config for claude: /repo/.mcp.json

out store: /repo/.kontinuo

out next: restart or reload your agent so it starts the Kontinuo MCP server

$ kontinuo show --cwd .

out # Kontinuo Handoff (compact)

out Stopping point: Renderer, CLI flags, and MCP read params done and verified.

out Next action: Inspect the changed-file set, then commit and push.

out Files changed: internal/checkpoint (6), internal/cli (2), internal/mcpserver (3)

Hands-free capture

Checkpoints that write themselves.

When your editor supports the Agent Client Protocol, Kontinuo watches the session as it runs. No instructions in your prompts, nothing for the agent to remember — the checkpoint is just there when the next session starts.

kontinuo acp --cwd . --inject compact-- codex

Protocol-level capture

ACP support observes the Agent Client Protocol stream, so Kontinuo can capture clear intent without betting the handoff on private transcript layouts.

Context at the right moment

When you run an agent through Kontinuo, opt-in context injection can place a compact handoff into the first prompt before the agent starts guessing.

Less brittle than scraping alone

Transcript adapters still help, but ACP gives editors and agents a cleaner path: the handoff travels through the protocol that already connects them.

The standard

Built to degrade, not break.

If a transcript adapter misses, Kontinuo falls back to workspace-only capture. If ACP is available, Kontinuo can observe protocol signals instead of leaning on private transcript layouts. If work is deferred, it stays parked.

Local-first store outside the repo by default.

Open JSON checkpoint schema with Markdown rendering.

Secret redaction across provider token formats.

Parked work stays parked until it is reopened.

FAQ

Plain answers for agent handoffs.

What is Kontinuo?

Kontinuo is a local-first MCP server and CLI for verifiable handoffs between AI coding agents. It creates checkpoints from agent sessions and workspace state so another tool can resume from a clear stopping point.

How does Kontinuo help AI coding agents resume work?

Kontinuo records the goal, exact stopping point, next action, verification evidence, changed files, and workspace fingerprint. The next MCP-capable agent reads that checkpoint before it starts work.

What is an MCP handoff?

An MCP handoff is a checkpoint served through the Model Context Protocol. Kontinuo exposes 10 MCP tools for reading, writing, capturing, importing, and resolving handoffs, including handoff_read and handoff_write. Agents can use them without parsing another agent's private state.

Does Kontinuo run agents or call models?

No. Kontinuo does not run agents, call models, or merge code. The agent reasons and edits; Kontinuo records and serves the local handoff state.

Does Kontinuo send session data off-machine?

No. Kontinuo keeps checkpoints and session-derived state local. The default store lives outside the repo, and the MCP server reads from that local store.

Which agent hosts does Kontinuo support?

Kontinuo supports Claude Code, Cursor, Antigravity, Codex, oh-my-pi (OMP), ACP-capable editors, and generic MCP hosts that can start a local stdio server. Transcript adapters cover known host formats, while ACP support gives protocol-level capture where the editor path supports it.

What does ACP support add?

ACP support lets Kontinuo observe Agent Client Protocol traffic instead of relying only on transcript scraping. That gives cleaner intent capture and lets a compact handoff reach the first prompt before an agent starts guessing.

Start local

Leave better notes for the next agent.

Add the MCP server to your agent host and let every session end with a checkpoint another tool can trust.