What is Belay?
Belay is a private experience layer for coding agents. It shows what keeps repeating, links useful lessons to the original evidence, asks the developer for approval, and carries relevant guidance into a future session.
Is Belay another coding agent?
No. Belay works alongside Claude Code, Codex, and Cursor. It does not replace them or provide its own general-purpose coding model.
Is Belay just memory?
No. Memory makes information available later. Belay goes further: it determines which lesson from prior work matters to the current task, who approved it, where it applies, and how the agent should verify its work.
Does Belay upload my coding sessions?
No. Belay Local stores coding-session data on the developer's machine. Doplex Labs does not receive the developer's transcripts, prompts, completions, code, or local Belay database through Belay. The only thing Belay sends is the usage ping described under Network boundary.
What is stored locally?
The encrypted local database may include session and project information, conversation text, tool calls and results, commands, verification evidence, per-session debriefs, and proposed or approved lessons.
The database is encrypted with a key held by the operating system for that user account: the macOS Keychain on a Mac, and the Windows Data Protection API (DPAPI) on Windows. The key never leaves the machine.
Belay scrubs recognized secret patterns before retaining transcript text. Developers should still keep credentials and other secrets out of agent conversations.
Does Belay make network calls?
Belay makes one kind of network call of its own: a de-identified usage ping. It sends an install event once and an active event at most once a day, containing a random ID that Belay generates, the Belay version, operating system and chip, which of Claude Code, Codex, and Cursor are installed, and a timestamp. It never contains paths, project names, prompts, transcripts, code, or anything that identifies the developer. Doplex Labs keeps the country of the request, not the address. belay telemetry off switches it off.
Some features can ask the developer's installed Claude Code or Codex to interpret selected local evidence under the developer's existing provider configuration and policies. Cursor Agent cannot run Belay's analysis, so debriefs and lesson proposals require Claude Code or Codex installed on the same machine. A Cursor-only machine gets sessions, evidence, and recurring-issue detection, but no debriefs and no semantic lessons.
When the developer requests analysis, selected evidence may be processed by their configured model provider. The local Belay database is not sent.
Does Doplex Labs train a model on my sessions?
No. Belay ships no model and does not upload sessions to Doplex Labs for training. Optional analysis follows the data policies of the developer's configured Claude Code or Codex provider.
Does Belay require an account?
No. Belay Local does not require a Belay account.
Can transcript text become an instruction automatically?
No. Conversation history, tool output, retrieved text, and AI-generated proposals cannot become active guidance without developer review and approval.
What does approval mean?
Approval authorizes a specific version of a lesson for a defined scope. The lesson can later be edited, paused, replaced, or removed.
Does Belay automatically modify my project?
No. Observed sessions and generated proposals do not automatically become project instructions. Any proposed configuration change is explicit, reviewable, and limited to supported agent-configuration files.
Which files does Belay read and write?
Belay reads the session history that Claude Code, Codex, and Cursor already keep on the machine. Cursor's agent transcripts are read from ~/.cursor/projects/*/agent-transcripts.
The configuration Belay writes is limited to the agent-configuration files it owns entries in. For Cursor those are two files: ~/.cursor/hooks.json, where quickstart installs monitor-only hooks, and ~/.cursor/mcp.json, where it registers Belay's local MCP server. Belay adds or removes only its own entries and leaves other hooks and other MCP servers in those files untouched. It also installs the /belay skill into ~/.cursor/skills/belay.
Cursor support is new in this alpha and has not had a clean-machine validation run yet.
Will Belay interrupt my coding agent?
No. If Belay is unavailable, the coding agent continues working normally.
What does Belay give a future agent?
Belay can provide a short Mission Pack containing a relevant lesson, why it matters, where it applies, where it should not apply, a recommended check, and references to the original evidence.
What is a session debrief, and where does it go?
After a session ends, Belay can ask your own Claude Code or Codex to write a debrief for you: where your prompting, scoping, or checks cost time or money, what an expert would have done at that moment, and the message to send instead. It sends that agent a scrubbed excerpt of the one session, with secrets removed and file paths reduced to file names, under the account and policy you already use.
The debrief is stored encrypted on your machine and compared only with your own other sessions in the same project. It is written for you, not for the agent: it is never delivered over MCP and never folded into a Mission Pack.
Cursor Agent cannot run Belay's analysis, so debriefs and lesson proposals require Claude Code or Codex installed on the same machine. A Cursor-only machine gets sessions, evidence, and recurring-issue detection, but no debriefs and no semantic lessons.
What do the preprint results show?
On 16 new tasks where a lesson from earlier work was useful, Belay-assisted agents completed 11. Agents without additional guidance completed 6, while agents with a concise human-written instruction completed 7.
The result demonstrates the potential for prior lessons to help future work and motivates broader evaluation across agents, tasks, and real development workflows. Doplex Labs will publish the methodology and agent-level results in a research preprint.
This evaluation measured approved lessons delivered to agents. Session debriefs have not been evaluated yet, and Belay does not claim that reading one makes the next session go better.
What ships today?
- Local Claude Code, Codex, and Cursor session acquisition.
- Encrypted local retention.
- Evidence-backed issue and session views.
- Per-session debriefs written by the developer's own agent.
- Reusable lesson proposals for developer review.
- Developer approval and lifecycle control.
- Delivery of approved guidance to supported coding agents.
- Recording of selected verification checks.