Skip to main content

KurrentAI Observability

kcap

kcap, formerly Kurrent Capacitor, records coding agents' sessions into a team workspace as they run: prompts, replies, every tool call and its result, and the commits made. Teammates search and review them, and the next agent reads them back through kcap's MCP servers. Team is priced per seat, a local edition is free for one user, and access is by invite.

Disclosure Kurrent is a GILD event partner. GILD writes its reviews on its own. About GILD partners

The 30-Second Take

kcap has no practitioner score yet, because GILD found no independent review or deployment account of it as of Nov 4, 2026.

Kurrent's product records sessions from Claude Code, Codex, Cursor and nine other coding agents into a shared workspace, where a team searches and reviews them and hands the work back to agents.

Best fit
Engineering teams running several coding agents who want each session searchable by teammates and readable by the next agent.
Main constraint
Access is by invite, Kurrent's pricing page and llms files print different Team prices, and kcap-server, which stores the sessions, is private, so its storage can't be inspected from the code.
First move
Request a Team trial, set repository exclusions and session visibility, and get the price for your seat count in writing.

Why

It fits teams that want agent sessions kept as shared, searchable history, priced per seat. Kurrent's own pages disagree on the Team price and on Insights in the free edition, so get the price and terms in writing, and set repository exclusions before the first session is stored.

Practitioner sentiment, all practitionersUnscored until 20 sources from 15 authors

0 practitioners, 16 sources reviewed, Updated Nov 4, 2026How this is measured

Read the GILD Review
Save a product your organization uses.Keep up with kcap without adding it to your stack.Visit Website
Operating stack
Agentic Software Development

Comparisons

How does kcap compare with the alternatives?

  1. kcap vs Langfuse

    • AI Observability

    Both capture Claude Code sessions through its hooks. Pick kcap to keep every coding agent's sessions as team memory priced per seat, and Langfuse to trace and evaluate an LLM application you build, metered in units.

  2. kcap vs Braintrust

    • AI Observability

    Pick kcap to keep coding agents' sessions as searchable team memory priced per seat. Pick Braintrust to score and test an AI application you build, on plans priced by processed data and scores.

Buying and deployment

What engineering leaders should verify before a broader rollout.

Sourced from Kurrent documentation

How you buy it

US$30 a month for one seat

Or US$100 for five seats plus US$25 each. Kurrent's llms files print US$15 per person billed annually.

Source 1, Kurrent, kcap pricing: kcap pricingSource 2, kurrent.io, llms.txt: llms.txtSource 3, kurrent.io, llms-full.txt: llms-full.txt

Details

The pricing page lists Team at US$30 per seat a month for a single seat, and its structured data at US$100 a month for five seats plus US$25 per additional seat. Kurrent's llms.txt and llms-full.txt print Team at US$15 per person per month billed annually (US$18 billed monthly). Local is free for one user, Enterprise pricing is custom, and the Local and Team buttons on the pricing page read Request an invite.

Cost drivers

Per seat, up to 100

No overage or metered fee on 86 kcap docs and Kurrent pages (Nov 4, 2026).

Source 1, Kurrent, kcap pricing: kcap pricingSource 4, kurrent.io, Plan and seats: Plan and seatsSource 5, kurrent.io, Configuration: Configuration

Details

Team is priced per seat: active members plus pending invitations count as seats, Team holds 100 and refuses invitations at the limit. Title, summary and local evaluation calls run on each user's Claude or ChatGPT sign-in by default, or the user's API key when use_provider_api_key is true. On Team and Enterprise, server-side evaluations use BYOK or Kurrent AI. GILD found no overage terms beyond the seat price on the 86 kcap docs and Kurrent pages searched, Nov 4, 2026.

Where GILD searched for overage terms beyond the seat price, Nov 4, 2026
WherePages readMatches
kcap documentation and Kurrent's kcap pages, every URL in the stored index, including pricing, terms, privacy, llms.txt and llms-full.txt860
GILD search record, Nov 4, 2026
Where it runs

EU cloud or your machine

Enterprise can be self-hosted. kcap is Kurrent License v1, source-available.

Source 6, Kurrent, Data security: Data securitySource 1, Kurrent, kcap pricing: kcap pricingSource 7, Kurrent, LICENSE.md: LICENSE.mdSource 3, kurrent.io, llms-full.txt: llms-full.txt

Details

A cloud workspace is isolated per organisation and hosted in the EU on infrastructure Kurrent operates, with sign-in through WorkOS in the US. Enterprise is Kurrent-managed cloud or self-hosted. The local edition keeps data on the machine, though the CLI and desktop app send usage reports until turned off. kcap is Kurrent License v1, source-available: it permits use, copying, distribution and derivative works, subject to limits that bar offering it to third parties as a hosted or managed service.

What to verify

No audit log found

None on the 86 kcap docs and Kurrent pages GILD searched (Nov 4, 2026).

Source 1, Kurrent, kcap pricing: kcap pricingSource 6, Kurrent, Data security: Data security

Details

SSO and directory integration come with Enterprise only. Sessions are retained until a member or an admin deletes them, and Kurrent says workspace data is never used to train models. GILD found no audit log of user or admin actions on the 86 kcap docs and Kurrent pages it searched, Nov 4, 2026. Kurrent's data-security page says its own operational access is logged.

Where GILD searched for an audit log, audit trail or activity log, Nov 4, 2026
WherePages readMatches
kcap documentation and Kurrent's kcap pages, every URL in the stored index, including pricing, data security, privacy, llms.txt and llms-full.txt860
GILD search record, Nov 4, 2026

Signals

What is the current evidence saying about kcap?

What engineers report publicly about kcap, formerly Kurrent Capacitor, with each finding linked to its sources. GILD found no independent account as of Nov 4, 2026, so all 11 records listed are tied to Kurrent, 10 from its staff and channels and one post Kurrent solicited, and none counts.

Whose sources count

Practitioner sentiment, all practitioners

Unscored until 20 sources from 15 authors

0 sources from 0 authors so far.

How this figure is made

Leaders and teams have 0 accounts; GILD prints their sentiment from 20.

Current sentiment

Positive, mixed and negative sources

GILD found no independent review or deployment account of kcap as of Nov 4, 2026. Kurrent's staff and channels wrote 10 of the 11 records listed, a LinkedIn post came after Kurrent reached out to its author, and outside bug reports were written with coding agents, so none counts. The findings below rest on Kurrent's own statements.

Run a pilot on one production app to measure which failures it surfaces and how long tracing them takes.

Evidence field. All 16 sources, one mark each.

Arrange by

Each mark is one source: positive, mixed, negative, no opinion stated, or documentation (not counted).

By sentiment, All practitioners. Every source GILD read, grouped as the field draws it.
GroupSourcesPositiveMixedNegativeNo opinion stated and not counted in the scoreDocumentation, checked and not counted
Positive000000
Mixed000000
Negative000000
Not counted11000110
Documentation500005
Every source GILD read, by the quarter it was published, as the field's time axis counts it. Documentation is dated by the day it was checked.
QuarterSources
Q2 20262
Q3 20266
Q4 20263
Documentation, checked5

By sentiment

5 documentation pages check facts and are not counted.

Hover a mark for a quick look, click it or press Enter to open its source.Tap a mark to open its source, then step to its neighbors with Previous and Next.

What practitioners praise

Each mark is a source on it; filled marks praise it.

For all practitioners, no capability reaches 5 sources from 3 authors.

    Most common concerns

    Each mark is a source on it; filled marks raise it.

    For all practitioners, no concern reaches 5 sources from 3 authors.

      All 16 sources behind these findings
      1. kcap docs: What kcap isKurrent (Official), Product documentation, checked Nov 4, 2026
      2. kcap pricingKurrent (Official), Product documentation, checked Nov 4, 2026
      3. kcap docs: Data security and privacyKurrent (Official), Product documentation, checked Nov 4, 2026
      4. kcap LICENSE.md (Kurrent License v1)Kurrent (Official), Product documentation, checked Nov 4, 2026
      5. kcap docs: AI providers and modelsKurrent (Official), Product documentation, checked Nov 4, 2026
      6. kurrent-io/kcap issue #2226: kcap's SessionEnd hook and claude -p on WindowsLokhesh Ujhoodha of Kurrent (self-described: Kurrent's own team. The record rests on Ujhoodha's comment alone, since an AI tester wrote the issue body), GitHub issue, published Nov 2, 2026
      7. LinkedIn post by Norton Andreev on kcap's invite-only betaNorton Andreev, working on kcap's desktop app at Kurrent (self-described: Kurrent's own team, inviting developers to try the product), LinkedIn post, published Oct 28, 2026
      8. kurrent-io/kcap issue #1851: a hosted agent linked to the user's own terminal sessionNorton Andreev of Kurrent (self-described: Kurrent's own team, on its hosted agents), GitHub issue, published Oct 22, 2026
      9. kurrent-io/kcap issue #1155: large tool results lost to the redaction deadlineAlexey Zimarev of Kurrent, with a status note by Norton Andreev of Kurrent (self-described: Kurrent's own team, measuring its own redaction), GitHub issue, published Sep 25, 2026
      10. LinkedIn post: trying Kurrent Capacitor's handoff from Claude Code to CodexSami Maxhuni, senior full stack developer (self-described: An individual developer in Kosovo, writing a first impression after Kurrent's team reached out to Maxhuni), LinkedIn post, published Sep 14, 2026
      11. kurrent-io/kcap issue #859: a service-installed daemon took 13 minutes to startTony Young of Kurrent (self-described: Kurrent's own team, on one developer machine), GitHub issue, published Sep 10, 2026
      12. Capacitor: Token and Time Analytics for Coding AgentsKurrent, on its YouTube channel (self-described: Kurrent's own channel, on the product's analytics), YouTube video, published Aug 5, 2026
      13. A Token's Journey: What Kurrent Capacitor Really Does for Coding AgentsMadhu Murty, an author on Kurrent's blog (self-described: Kurrent's own team, on what the product captures), Medium post, published Jul 20, 2026
      14. When Your Coding Agent Doesn't ListenTony Young, on Kurrent's blog (self-described: Kurrent's own team, scoring one of its sessions with its own product), Kurrent blog post, published Jul 15, 2026
      15. Show HN: Capacitor, shared mem for Claude Code, Cursor and other coding agentsLokhesh Ujhoodha, Kurrent's Product and AI team (self-described: Kurrent's own team, on how it uses the product it builds), Show HN post, published Jun 30, 2026
      16. X post by Alexey Zimarev on why Kurrent built CapacitorAlexey Zimarev, Head of Product and Engineering at Kurrent (self-described: Kurrent's own team, on why it built the product), X post, published Jun 7, 2026

      Jun 2026 to Nov 2026.

      11 sources from practitioners, by month and sentiment.

      The full readExplore all kcap Signals4 praise themes, 2 concerns, 0 points of agreement and 0 disagreements across all three groups, with all 16 sources open to inspect.

      Where It Fits

      Where does kcap belong in the stack?

      Open Where It Fits: the Stack Field and the full map

      kcap is Core across Build, supports Validate, and touches Plan, Ship, and Operate.

      Business area
      Product & Engineering
      Function
      Software Development
      Operating stack
      Agentic Software Development
      Primary role
      Shared memory for coding agents
      kcap beside the rest of its stackThe Stack Field, every product on every layer
      Footprintkcap across the Agentic Software Development stack. Bar height is its role on each layer.

      1. Plan

        1. Work planning: Touches

          Work planning, touches. kcap's Linear app installs once per workspace, and its work items, a Team and Enterprise feature that the local edition also has, use the workspace's embedding provider for matching. Planning stays in the issue tracker.

      2. Build

        1. Codebase context: Core

          Codebase context, core. This is kcap's primary role: a hook in each coding agent sends prompts, replies, tool calls and commits to a shared workspace, and the agent searches past sessions through kcap's MCP tools. A session-start briefing injects approved documents and facts, capped at 16,000 characters by default.

        2. Coding: Touches

          Coding, touches. kcap records what Claude Code, Codex, Cursor and nine other coding agents do. The coding agent writes the code.

        3. Agent runtime: Supports

          Agent runtime, supports. Hosted agents run on a machine the team connects through the kcap daemon, which launches the agent CLIs installed there in a fresh git worktree, and the web app and desktop app follow the sessions. The coding agent's own interface stays where developers steer it.

        4. Tool execution: Supports

          Tool execution, supports. kcap's MCP servers run locally as kcap mcp <name> and give the agent tools over its own workspace, with reads filtered by session visibility and writes made as the user. Other tools and commands stay with the agent.

      3. Validate

        1. Debugging: Touches

          Debugging, touches. Kurrent says kcap lets a team revisit decisions and review the work an agent did from its stored session. Debugging the code stays with the developer's tools.

        2. Testing: Touches

          Testing, touches. kcap's evaluations ask the questions in the workspace's catalog about captured sessions, with a judge model set per vendor. Tests of the code stay in the test suite.

        3. Code review: Supports

          Code review, supports. The Team plan gives reviewers the session behind a change, which the pricing page calls session context to help teammates review code together. Review comments and merge checks stay with a code review tool.

        4. Application security: Touches

          Application security, touches. An excluded repository or folder is never stored, and members' sessions default to org_public, readable by the organisation for its own repositories. Scanning code for vulnerabilities stays with a security tool.

      4. Ship

        1. Source control: Touches

          Source control, touches. The GitHub App installs once per workspace with no personal token, and the hook records the commits an agent made. Branches and pull requests stay in GitHub.

        2. CI/CD: Touches

          CI/CD, touches. Machine credentials let CI runners, build boxes and agent sandboxes store sessions without a person signing in, on workspaces that sign in through single sign-on. Builds and deploys stay in the CI system.

      5. Operate

        1. Observability: Touches

          Observability, touches. Insights reports counts, tokens, cost, tool use and evaluation scores for agent sessions, and status.kcap.ai carries kcap's own status. Production monitoring stays with an observability tool.

      Market mapThe layers kcap covers, and who else owns each one.

      1. Codebase context & intelligence

        CoreBuild

      2. Agent interface & developer runtime

        SupportsBuild

      3. Tool execution & integrations

        SupportsBuild

      4. Code review & quality

        SupportsValidate

      5. Work planning & task decomposition

        TouchesPlan

      7 more layers on the full map. Core, Supports and Touches describe coverage. They say nothing about quality.

      What kcap does with a session

      What happens to a session kcap records, where does each step run, and what lands in the next session?

      Kurrent is a GILD event partner. GILD read kcap's code at commit c4471ad and Kurrent's docs on Nov 4, 2026, with nothing installed or run.

      One session, six steps

      Read where each step runs and what it holds back, and pick an agent to see the hook events kcap installs in it.

      Capture, redaction and the next session's context run on your machine. kcap-server stores the sessions. Summaries and scores run on your AI provider on the local edition, or on Kurrent AI by default in the cloud.

      1. 1CaptureYour machine
      2. 2RedactYour machine
      3. 3Storekcap-server
      4. 4SummarizeLocal edition: your AI providerCloud: Kurrent AI by default
      5. 5ScoreLocal edition: your AI providerCloud: Kurrent AI by default
      6. 6InjectYour machine
      Team price on the pricing page
      $30 per month for one seat
      Documented The pricing page says "Team is $30 per month for one seat, or $100 per month for five seats plus $25 per additional seat." Local is free for one user. Kurrent's llms.txt prints a different Team price, "$15 per person per month billed annually ($18 monthly)".
      kcap pricing Read Nov 4, 2026
      Memory index cap
      24 KiB
      Documented MaxFragmentBytes = 24 * 1024 in the CLI, which is 24,576 bytes.
      kcap SessionStartMemoryContracts.cs, line 114 Read Nov 4, 2026
      Guidance budget
      16,000 characters
      Documented The server's default for combined documents and facts from past sessions, which server operators can change.
      kcap docs: Session-start context Read Nov 4, 2026
      MCP servers
      Up to 10 per agent
      Documented kcap-review, kcap-sessions, kcap-flows, kcap-memory, kcap-workitems, kcap-plans, kcap-handoff, kcap-pages, kcap-analytics and kcap-knowledge. Grok Build gets none, and Pi gets them through a bridge.
      kcap docs: MCP tools and skills Read Nov 4, 2026
      GILD's estimate for the two capped pieces, before reminders and MCP tool definitions
      About 10,144 tokens
      GILD analysis GILD's calculation from the 24,576-byte memory fragment cap and the 16,000-character guidance budget at an assumed 4 per token, 6,144 plus 4,000. Both figures are caps and no Kurrent page prints a measured size. No Kurrent page states a ratio of bytes or characters per token, and the sum holds for ASCII text only.
      kcap docs: Session-start context Read Nov 4, 2026

      1. CaptureYour machine

        Hooks, a plugin or an extension in each of 12 agents start a watcher that tails the agent's transcript and streams it to kcap-server over SignalR. All 12 agents also have a history import.

        Limits Codex records nothing until its hooks are trusted. One watcher runs per live session. GILD has not run one to see which server it reaches.

        Sourceskcap HarnessId.cs Read Nov 4, 2026kcap WatchCommand.cs, line 702 Read Nov 4, 2026kcap repository tree at c4471ad Read Nov 4, 2026kcap README, line 520 Read Nov 4, 2026

        Hook events per agent

        Claude Code

        • SessionStart
        • SessionEnd
        • SubagentStart
        • SubagentStop
        • PreToolUse
        • PostToolUse
        • PostToolUseFailure
        • Notification
        • Stop
        • PermissionRequest
        • UserPromptSubmit

        Plugin hooks, a transcript watcher, and history import with kcap import --claude. PostToolUse also runs plan-read, task-list and plan-edit matchers.

        kcap Claude Code hooks.json Read Nov 4, 2026

      2. RedactYour machine

        The CLI passes each transcript line through SecretRedactor before it joins the batch sent to the server. Its 12 pattern families replace a match with [REDACTED]. Live capture and import use the same redaction, and commit messages are filed redacted.

        Limits A record over the 4 MiB bound is sent as a capture-loss marker. A secret in a format none of the 12 families matches stays unredacted, in GILD's reading of the regexes, and GILD's reading of the code, which GILD has not run, finds no setting that skips redaction.

        Sourceskcap SecretRedactor.cs Read Nov 4, 2026kcap SecretRedactor.cs, token prefixes Read Nov 4, 2026kcap TranscriptCapture.cs Read Nov 4, 2026kcap docs: What kcap installs Read Nov 4, 2026

      3. Storekcap-server

        Closed server The repo's own release workflow says "# kcap-server is private." Its schema and storage can't be inspected from the repo. kcap release.yml, line 442 Read Nov 4, 2026

        kcap-server stores the sessions: on the local edition with KurrentDB and Postgres on the developer's Mac, and in the cloud on Kurrent-operated infrastructure in the EU, one server per organisation.

        Limits The repo shows no compose file, Dockerfile or database version, and the README names no Postgres version, pgvector or pg_trgm.

        Sourceskcap release.yml, line 442 Read Nov 4, 2026kcap repository tree at c4471ad Read Nov 4, 2026kcap README, line 556 Read Nov 4, 2026kcap docs: Cloud vs local Read Nov 4, 2026

      4. SummarizeLocal edition: your AI providerCloud: Kurrent AI by default

        Titles and outcome summaries run on Kurrent AI by default in the cloud. On the local edition they run on the team's own provider key or, through the daemon, on its own claude -p with tools and MCP servers off or its own codex exec in a read-only sandbox with no network.

        Limits The CLI's setup request has no base URL, while the server's provider page offers an OpenAI base URL for a gateway, which GILD has not tested. Headless runs use the subscription login by default. Import titles send up to 500 characters of a session's first message and 300 of its first reply to the provider your claude or codex uses.

        Sourceskcap docs: Evals Read Nov 4, 2026kcap IntelligenceSetupRequest.cs Read Nov 4, 2026kcap docs: AI providers and models Read Nov 4, 2026kcap README, line 2735 Read Nov 4, 2026kcap ClaudeHeadlessHarness.cs Read Nov 4, 2026kcap docs: Cloud vs local Read Nov 4, 2026

      5. ScoreLocal edition: your AI providerCloud: Kurrent AI by default

        An evaluation asks 13 questions, scored 1 to 5 with pass, warn or fail by a separate headless judge with no filesystem or network tools. A question with too little evidence, or none that applies, gets no score. The CLI's default judge is sonnet. The server sets the judge per vendor, with defaults claude-sonnet-4-6, gpt-5-mini, command-a-03-2025 and google/gemini-2.5-flash.

        Limits A cloud workspace evaluates on Kurrent-managed AI by default. Up to 16 judges run in parallel per vendor, 4 by default.

        Sourceskcap README, line 758 Read Nov 4, 2026kcap README, line 748 Read Nov 4, 2026kcap docs: AI providers and models Read Nov 4, 2026kcap docs: Evals Read Nov 4, 2026kcap docs: Cloud vs local Read Nov 4, 2026

      6. InjectYour machine

        At session start the hook fetches the memory index from the server and hands the agent one context block: the index once per conversation, guidance from past sessions, and reminders. The index carries one slug and description line per memory, without the bodies.

        Limits The CLI caps the memory fragment at 24 KiB, and the server's guidance defaults to 16,000 characters. Pi and OpenCode append the index on every request. Up to 10 MCP servers load per agent, none for Grok Build, and Pi gets them through a bridge. The README wires the index for nine agents and leaves out Mistral Vibe, Grok Build and Muse Code (lines 179 and 196 to 198).

        Sourceskcap SessionStartMemoryContracts.cs, line 114 Read Nov 4, 2026kcap README, line 179 Read Nov 4, 2026kcap docs: Session-start context Read Nov 4, 2026kcap KcapMcpServers.cs Read Nov 4, 2026kcap docs: MCP tools and skills Read Nov 4, 2026

      GILD's own evaluation

      On Nov 4, 2026 GILD read kcap's source to decide whether to use it for its own agents' memory, and chose not to adopt it. The server that stores, summarizes and scores can't be inspected, and each session starts with context that costs tokens GILD watches. Those limits matter to a team that must audit its memory store or watch per-session tokens, and a team with neither need can weigh them differently. GILD borrowed kcap's judge format, recap shape and capped index for its own tool.

      Strengths

      Limits

      This is GILD's own account, labelled GILD analysis. It counts toward no practitioner theme or score, and Kurrent is a GILD event partner.

      Every stage, agent list and figure comes from kurrent-io/kcap at commit c4471ad and Kurrent's docs, read Nov 4, 2026. DeepSeek is not a supported agent. The repo at the pin shows no DeepSeek harness. HarnessId has no DeepSeek member, no path in the pinned tree of 5,881 files contains deepseek, and the one README mention of DeepSeek is a model value for an OpenCode hosted agent (KCAP_OPENCODE_MODEL, README lines 2184 to 2191).

      More on kcap

      1. GILD ReviewGILD's evaluation, with the buying facts
      2. Research1 piece
      3. Guides1 guide
      4. Where It FitsOn the Agentic Software Development stack
      5. Related Products4 products