01
Create your AgentPM workspace
Start by creating an account. Then create an org for your project and add a data source so AgentPM knows where your coding-agent sessions should land.
Create account// agentpm guide for vibe coders
Keep the context when models run out.
AgentPM gives vibe coders agent continuity: move between agents, sessions, and models while keeping the full story of what happened, why it happened, and what to do next.
# agent continuity# model handoff# context recovery# notebook search
getting started
01
Add dataCreate the org and copy the install command.
02
Install the agentCapture sessions while you code.
03
Install the skillLet the next model ask what happened.
* CLAUDE CODE * CODEX * CURSOR * GROK * OPENCODE * NEVER LOSE THE THREAD
// the vibe coding tax
The code moves fast. The context gets left behind.
Vibe coding feels amazing until a model runs out of tokens, a session ends, or you want a different agent to continue. AgentPM keeps the work trail intact so the next run can start from reality, not vibes alone.
TOKEN LIMITSThe model was doing great, then the window filled up and the useful context fell out.
SESSION HOPSYou switch from Claude to Codex to Cursor and have to retell the same story again.
HIDDEN RISKSThe agent made a smart choice, skipped a check, or left a warning somewhere in the scrollback.
UNFINISHED WORKThe real next step is buried in a previous session instead of sitting in front of the next agent.
// the guide
Set up a continuity loop once, then keep coding.
The flow is simple: create an account and org, add data, install the agent where work happens, install the skill, and ask your coding agent to retrieve context before it plans.
02
Copy the org install command
From the org landing page, click Add data and keep Install directly to this org selected. Copy the command AgentPM gives you.
03
Install the agent
Paste the org-scoped install command into the terminal on the laptop, dev box, or server where your coding agents run.
terminal
$ curl -fsSL 'https://agentpm.dev/install.sh?org=my-org&code=oic_...' | bashDownloading agentpm...Installing local agent...Starting AgentPM capture...ok: agent installed and connected to my-org
04
Install the skill
Install the AgentPM conversation skill so Claude Code, Codex, Cursor, Grok, OpenCode, or the next tool can ask AgentPM what happened before.
terminal
$ agentpm skill installinstalled skill=agentpm-conversations
05
Pick up work from another agent
Start the next agent inside the repo you are already working in. Before it plans or edits, have it ask AgentPM for the current checkout's recent work, risks, and open tasks.
claude code / codex / cursor / opencode
you
Check AgentPM to pick up the conversation I left off in Codexregarding the OAuth callback work. Tell me what we already figured out, what is still open,and what I should do next before editing.
agent response
I'll check AgentPM for the prior Codex thread before I plan. AgentPM found the handoff:- OAuth callback fix was started in Codex- auth-state parsing was already investigated- preview env verification is still open- don't redo the cookie serialization path Next best move:run the callback tests, then verify preview env config.
06
Review your work in the UI
Open the AgentPM notebook to search conversations, inspect turns, and verify the tool calls behind a handoff.
07
Gain insight and understand
Use insights to see decisions, risks, outcomes, and advice without rereading every token of the previous session.
// agentpm skill
Ask the next model to check AgentPM before it guesses.
The habit is simple: when you switch agents, reopen a stale repo, or come back to a feature, tell the coding agent to recover the AgentPM thread first.
agent prompt
I'm in Claude now. Check AgentPM and pick up the conversation I left off in Codexabout the checkout refactor. Before you edit anything, tell me:- what we already decided- what is still open- the next safest step
agent response
I'll check AgentPM for the prior Codex thread first,then continue from that handoff. AgentPM found the Codex handoff:- checkout state was moved out of the page component- discount validation still needs a failing test- webhook retry behavior was not verified- don't redo the cart serialization investigation Next best move:add the missing discount test, then run the checkout flow locally.
The next model starts from the conversation you already had, not a blank prompt.
// useful commands
Real handoff commands for real coding-agent moments.
Once AgentPM has captured a few sessions, use it exactly where context usually breaks: starting a new agent, debugging an old issue, checking the receipts, preparing a PR, or handing work from one model to another.
Pick up the exact thread
you
Check AgentPM for the Codex conversationwhere I was working on the checkout refactor. What was the last real state before that session ended?
agent
AgentPM found the Codex thread. I'll recover the prior conversation,including the decisions, open tasks, and last known blocker. Next best move:continue from the last verified checkout state.
Avoid repeating old debugging
you
Before debugging the auth callback again,check AgentPM for earlier Claude and Codex sessions. What did we already rule out,and what evidence ruled it out?
agent
AgentPM found earlier investigations. I'll look for prior investigationsso we do not repeat solved paths. - do not redo solved auth-state parsing- check the evidence before changing code
Recover the reason behind a change
you
Check AgentPM for why the previous agentmoved checkout state out of the page component. Was that a user decision, an implementation shortcut,or a fix for a specific bug?
agent
AgentPM found the decision trail. I'll find the conversation where that decision was madeand recover the reasoning behind it. Next best move:read the source turn before refactoring further.
Find the last known failure
you
Check AgentPM for the last session about preview deploys. What command failed?What was the error?What did the agent try right before context ran out?
agent
AgentPM found the failure trail. I'll recover the exact failure trailfrom the prior conversation. - preview deploy verification is still open- last failed command needs review
Recover user instructions
you
Search AgentPM for anything I told the last agentnot to change during the auth work. List the constraints, warnings, and scope boundariesbefore you touch code.
agent
AgentPM found prior scope boundaries. I'll recover your prior instructionsso I do not violate the earlier scope. - do not change unrelated auth flows- keep the fix scoped before editing
Merge context across models
you
I worked on this feature in both Claude Code and Codex. Check AgentPM across both conversations.What did each agent learn,and are there any contradictions?
agent
AgentPM found Claude and Codex threads. I'll combine the previous model threadsinto one handoff before planning. Next best move:compare decisions before choosing an edit path.
Write the PR from the work history
you
Before I open this PR,check AgentPM for the conversations behind this branch. What was the intent,what was verified,what was skipped,and what should reviewers know?
agent
AgentPM found the branch history. I'll build the PR context from the actual prior sessions,not just the diff. Next best move:include verified checks and skipped risks in the PR.
// see it yourself
Use the web UI when you want to investigate.
The terminal handoff is fast, but sometimes you want to look with your own eyes. AgentPM gives you a searchable notebook, conversation evidence, tool-call context, insights, and advice so you can inspect the full trail.
NOTEBOOK
Open the notebook when you need the receipts.
The notebook gives you the session itself: prompts, replies, tool calls, files, commands, and the surrounding timeline. It is the place to inspect what really happened instead of guessing from memory.
- Search conversations by project, phrase, or decision.
- Open the exact turn behind a summary.
- Review commands and tool calls in context.
INSIGHTS
See decisions, risks, and loose ends without rereading everything.
AgentPM turns long agent sessions into searchable decisions, outcomes, risky moments, unresolved questions, artifacts, and follow-ups. Each item links back to evidence so the next agent can trust it.
- Find why the previous agent changed direction.
- Surface risky choices before they become bugs.
- Bring unfinished work into the next session.
ADVICE
Turn repeated lessons into reusable guidance.
Advice surfaces the patterns that keep showing up across sessions, then lets you turn them into instructions the next agent can actually use.
- Promote hard-won lessons into standing guidance.
- See repeated risks across a run or work streak.
- Set up follow-through without digging through old chats.
// start without drama
A simple habit for every new session.
You do not need a heavy process. When a session is long, stale, or moving to a different model, make AgentPM the first command before the next agent plans.
01
Ask for continuity
Run AgentPM before the new model plans.
02
Pull the real state
Recover recent work, decisions, risks, failed checks, and open tasks.
03
Plan from evidence
Use the retrieved context to continue instead of rediscovering the same facts.
04
Inspect for yourself
Open the web UI when you need the full notebook, tool calls, and source turns.
// next run
Keep coding fast. Keep the thread.
Let agents explore, edit, and ship. AgentPM keeps the context, evidence, insights, and advice close enough for the next model to continue with confidence.