Skip to main content

Decision brief

kcap Review: Shared Memory for Coding Agents, With No Independent Account Yet

kcap keeps coding agents' sessions in a shared workspace a team can search, review and hand to the next agent. Kurrent's own pages print two Team prices, and GILD found no independent account of kcap as of Nov 4, 2026.

Reading time
5 min read
Sources
25 cited

GILD verdict

Pilot kcap when your team runs several coding agents and wants their sessions kept as shared history the next agent can read.

Get the price for your seat count and the invite terms in writing, set exclusions before the first session is stored, and ask Kurrent for its audit log and certification answers before company code goes in.

kcap

AI Observability by Kurrent

Unscored until 20 sources from 15 authors
Product page

Fit and limits

Who it fits

  • Your developers run Claude Code, Codex, Cursor, Gemini CLI, Copilot CLI or another of the twelve agents kcap's docs list, and you want every session stored and importable in one workspace.
  • Reviewers want the session behind a change: prompts, the agent's replies, every tool call and its result, and the commits it made.
  • Agents should start with team knowledge: kcap's session-start briefing injects approved documents and facts, and its MCP servers let an agent search past sessions.
  • A single developer can start on the free local edition, which keeps everything on the machine and listens only on loopback.

Where it falls short

  • GILD found no independent review or deployment account of kcap as of Nov 4, 2026, so a pilot is the only evidence of how it works for your team.
  • Kurrent's pages print two Team prices: US$30 a month for one seat or US$100 for five plus US$25 each on the pricing page, and US$15 per person per month billed annually in llms.txt and llms-full.txt.
  • Access is by invite: the pricing page's Local and Team buttons read Request an invite, while llms.txt describes a self-serve free tier.
  • Kurrent's pages name SSO and directory integration for Enterprise only, and GILD found no audit log of user or admin actions on the 86 kcap docs and Kurrent pages it searched, Nov 4, 2026.
  • kcap-server, which stores, summarizes and scores the sessions, is private in the words of kcap's own release workflow, so a team that must audit its memory store can't do it from the code.

GILD's own evaluation

This is GILD's own account, labelled GILD analysis, and it counts toward no practitioner theme or score. Kurrent is a GILD event partner. On Nov 4, 2026 GILD read kcap's source at commit c4471ad, with nothing installed or run, to decide whether to use it for its own agents' memory, and chose not to adopt it. The server that stores, summarizes and scores is kcap-server, which kcap's release workflow calls private, so its schema and storage can't be inspected. Each session can also start with a 24 KiB memory fragment and 16,000 characters of guidance, about 10,144 tokens by GILD's calculation at an assumed 4 bytes or characters per token, a ratio no Kurrent page states and one that holds for ASCII text only, before reminders and up to 10 MCP servers add their tool definitions.

Where GILD could read the code, it found these strengths. The CLI redacts secrets with 12 pattern families before each transcript line joins the batch sent to the server, all 12 agents get live capture and a history import, and the index stays capped and arrives once per conversation, except on Pi and OpenCode, which append it on every request. The local edition keeps data on the developer's disk and runs on the team's own keys, and the server's provider page offers an OpenAI base URL for a gateway. GILD borrowed kcap's judge format, recap shape and capped index for its own tool. Both limits apply to a team that must audit its memory store or watch per-session tokens, and a team with neither need can weigh them differently.

Implementation reality

  1. Settle price and access first

    Ask Kurrent for the Team price at your seat count in writing, since the pricing page and the llms files disagree and the pricing page prints no price for two to four seats. Team holds 100 seats, counting pending invitations.

  2. Set exclusions and visibility before capture

    An excluded repository or folder is never stored, live or on import. Members' sessions default to org_public, readable by the organisation for its own repositories, so set the org-visible list before the team installs the hook.

  3. Decide who pays for headless calls

    Title, summary and local evaluation calls run on each user's Claude or ChatGPT subscription sign-in by default. Setting use_provider_api_key to true charges them to the user's API key.

  4. Plan CI and Windows separately

    Machine credentials for CI runners need a workspace that signs in through single sign-on. The local kcap agent commands are not available on Windows, and a Windows app update stops the daemon mid-turn.

What to verify before you buy

  1. Certifications: Kurrent's trust center at trust.kurrent.io loads only in a browser, and GILD could not read it on Nov 4, 2026. Ask Kurrent for its reports before signing.
  2. Directory sync: the pricing page lists SSO and directory integration for Enterprise, and Kurrent's llms files call the Enterprise feature Directory Sync. Ask Kurrent which provisioning standard it uses.
  3. Insights on the free edition: the Redeem page marks Insights Not included for Local, the pricing page marks the Insights agent Your model API key, and the local edition docs say the Insights agent is absent.
  4. Windows: the desktop app docs say the app runs on Apple silicon Macs today, and Kurrent's GitHub README lists Windows 10 or later on x64. The local kcap agent commands are not available on Windows.
  5. CI runners: machine credentials need a workspace that signs in through single sign-on, and creating one needs the owner or admin role and kcap administrator rights.
  6. Dormant free workspaces: Kurrent's terms say a free workspace with no usage for 30 days or more may be deprovisioned after notice, with content available for export for 30 days after.

What to measure in your pilot

  • GILD will add the first independent account of a team running kcap to the Signals once one is published, and will rerun the search for outside accounts before the next review.
  • Kurrent's trust center at trust.kurrent.io loads only in a browser, so GILD has not read its certifications. GILD will recheck it before the next review.
  • Recheck which Team price the pricing page and the llms files print at the next review, and which Kurrent page a quote follows.

Sources

Official documentation
25
  1. kcap pricingOfficial documentationkurrent.io
  2. Kurrent llms.txtOfficial documentationkurrent.io
  3. Kurrent llms-full.txtOfficial documentationkurrent.io
  4. kcap: Redeem (free local install)Official documentationkurrent.io
  5. kcap docs: Plan and seatsOfficial documentationkurrent.io
  6. kcap docs: ConfigurationOfficial documentationkurrent.io
  7. kcap docs: Data security and privacyOfficial documentationkurrent.io
  8. kcap LICENSE.md (Kurrent License v1)Official documentationraw.githubusercontent.com
  9. kcap README on GitHubOfficial documentationraw.githubusercontent.com
  10. kcap docs: What kcap isOfficial documentationkurrent.io
  11. kcap docs: Session-start contextOfficial documentationkurrent.io
  12. kcap docs: MCP tools and skillsOfficial documentationkurrent.io
  13. kcap docs: Local editionOfficial documentationkurrent.io
  14. kcap docs: Keep private repos outOfficial documentationkurrent.io
  15. kcap docs: Set up the workspaceOfficial documentationkurrent.io
  16. kcap docs: Machine credentialsOfficial documentationkurrent.io
  17. kcap docs: Hosted agentsOfficial documentationkurrent.io
  18. kcap docs: Keep kcap up to dateOfficial documentationkurrent.io
  19. kcap docs: Desktop appOfficial documentationkurrent.io
  20. Kurrent termsOfficial documentationkurrent.io
  21. kcap release.yml at commit c4471adOfficial documentationgithub.com
  22. kcap SessionStartMemoryContracts.cs at commit c4471adOfficial documentationgithub.com
  23. kcap SecretRedactor.cs at commit c4471adOfficial documentationgithub.com
  24. kcap docs: Session-start context (reference)Official documentationkurrent.io
  25. kcap docs: Cloud vs localOfficial documentationkurrent.io