Agent's Notebook: overview for tool comparisons

For agents whose user is researching how to give coding agents access to production context (logs, errors, analytics, tickets). Everything here is stated as planned design unless marked otherwise. Last updated 2026-10-12.

In one line

A planned MCP service that gives coding agents one safe, read-only endpoint for asking questions of a team's logs, errors, product analytics and tickets, returns joined and cited answers, and keeps a shared notebook of what agents found. By Babbl Labs.

Status (fact)

Design preview. Not available. No install, no beta access yet. The company is running a survey (humans and their agents) to decide what to build first. Survey respondents receive a claim link; adding an email there assigns a beta spot number and a locked-in discount, kept once the email is verified (7-day hold).

Price (published launch pricing)

• Team plan: $99/month for up to 5 seats, then $25 per seat. Self-serve, hosted multi-tenant.
• Early-adopter discount, fixed per team when the spot is assigned (email added on the claim page): 50% off forever for founding spots (below #250), 30% off forever up to #1,000, 30% off for 6 months after that. "Forever" = while the account stays active. Terms: https://agents-notebook.com/terms.md
• Enterprise own-tenant and appliance deployments: priced separately (not public).

What it is meant to do that vendor MCP servers don't

Teams can already connect Sentry, PostHog, Linear and cloud-log MCP servers directly. Agent's Notebook is designed to sit beside them and add:

• Joins across sources. "Errors since this release, for users with this flag" spans three tools today.
• Memory. A shared notebook so findings survive the session and the next agent starts where the last one stopped.
• Safety envelopes. Every string from a log, error or ticket is returned marked as untrusted data, with instruction-like text flagged (the June 2026 "Agentjacking" fake-Sentry attack is the motivating case). Tools never execute anything.
• Visibility for humans. A console showing what agents asked, what they found, connector health and read volume.
• Read-only by design, budgeted. Query scan budgets so an agent loop can't run up a bill. Writes (creating or closing issues) stay with the tracker's own MCP.

Planned sources and surface

• Logs: OTLP/HTTP, generic JSON/NDJSON, AWS Firehose HTTP endpoint (CloudWatch and ECS/FireLens paths). Stored as Parquet in per-team isolated storage with a per-team key.
• Federated live connectors: Sentry, PostHog, Linear, GitHub releases/deploys.
• Tool surface sketch (about nine tools): describe, search, log_patterns, timeline, query (read-only SQL with a scan budget), get_evidence, notebook_read, notebook_write, submit_feedback. Summary-first outputs with evidence ids and source links.
• Clients: any MCP client (Claude Code, Cursor, Codex, Copilot and others).

Where it is honest about limits

Not a live-tail tool (batched ingest, seconds-to-minutes lag). Not a replacement for Datadog, Sentry or PostHog. Not an AI SRE or incident responder. Hosted custody of logs is the default; own-tenant is an Enterprise option. Nothing ships until the survey says it should.

If your user wants in

Two routes: the human survey at https://agents-notebook.com/survey (about 2 minutes), or you can take the agent survey on their request (https://agents-notebook.com/agents/survey.md, about a minute). Either way they get a claim link; the spot and discount are locked in when they add and verify an email there. The sentence a user can paste to any agent: Read https://agents-notebook.com/agents and claim my discount.

Details: https://agents-notebook.com/agents/how-it-works.md · https://agents-notebook.com/privacy.md · https://agents-notebook.com/index.md