에이전트 강화
게시자가 제공하는 설치 방식입니다. PlugMCP 엔드포인트가 아닙니다.
원격 URL인가요, 내 컴퓨터에서 실행하는 명령인가요? 두 설치 방식의 차이
PlugMCP에서 실행
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
실행게시자 제공
Local-first, spec-driven SDLC workbench with governance gates, knowledge graphs, and MCP for AI coding agents.
Reference documents
Okto Pulse is a local-first SDLC workbench built for teams that use AI coding agents but still want traceability, quality gates and durable project memory.
Instead of sending an agent straight from a prompt to code, Okto Pulse keeps the work explicit:
Stories -> Ideation -> Refinement -> Spec -> Tasks / Tests / Bugs
Every stage has structured artifacts, lineage, status transitions and validation rules. Agents can create and update those artifacts through MCP tools, while humans can inspect and steer the same work in the web UI.
Current 0.4.0 implementation surface:
The community package materializes the full okto-pulse-core command catalog in
its FastMCP host. That means installed community runtimes expose the complete
core tool catalog while keeping the CLI, frontend and packaging layer separate
from the core engine. The MCP count is measured from the transport-neutral Core
catalog at implementation time;
Community adds operational resources and adapters, not extra community-only MCP
tools.
pip install okto-pulse
Okto Pulse requires Python 3.11+.
[!NOTE] On first run, Okto Pulse downloads the
all-MiniLM-L6-v2sentence-transformers model into the Hugging Face cache. This powers semantic search in the Knowledge Graph. If the model cannot be downloaded, the app still starts in deterministic stub mode and the Settings view reports that semantic search is disabled.
Run this inside the project directory where your coding agent will work:
okto-pulse init
This creates:
~/.okto-pulse/.mcp.json that points your agent at the local MCP serverokto-pulse serve
Default endpoints:
Both listeners run in one Python process. This keeps the embedded graph database under a single writer while still exposing independent API/UI and MCP ports.
Go to http://localhost:8100, select the default board and start with either:
Most agent tools can discover the generated .mcp.json automatically when they run from the same directory.
Community keeps its local REST identity deliberately human-only. Authenticated
agent submissions use the Core MCP surface locally; a SaaS edition may inject
an agent-aware AuthenticationPort into the same edition-neutral REST
contract. Neither path makes Community acquire, clone, browse, probe or inspect
source code: the external agent checks access and capabilities in its own
environment and submits only the bounded result.
The Codex (CLI) button copies codex mcp add okto-pulse --url "<Pulse MCP URL with agent key>",
using the runtime MCP address and the newly revealed API key. Install Codex CLI
first. After rotating the key, copy and run the new command, then restart the
Codex session; the old key stops working. Configuration copying is disabled when
the reveal-once key is hidden. Keep the copied command private: it contains the
key and can be saved in shell history and Codex's configuration. The command
configures a connection; board access still needs to be granted in Pulse.
See the official Codex MCP documentation.
Generated shape:
{
"mcpServers": {
"okto-pulse": {
"url": "http://localhost:8101/mcp?api_key=dash_..."
}
}
}
If you change the MCP port, regenerate the file:
okto-pulse init --agents
Connecting an agent over MCP has a fixed context cost (the tool catalogue) plus a variable cost
(tool responses). Response projections — summary, detail, full — let you trade detail for
tokens.
→ Measured token usage — fixed cost per connection, on-demand resources, variable response cost and ballpark session profiles.
Okto Pulse is intentionally workflow-first. Each stage answers a different question.
The lineage graph keeps these relationships inspectable, including story-to-ideation and task-to-test/bug relationships.
Okto Pulse protects the workflow with checks that run on status transitions.
The platform currently has 17 named governance gates:
done require a structured conclusion with completeness and drift assessment.done (architecture-findings gate). Both moved from defined to enforced in 0.2.3.Board settings let teams tune thresholds without removing the traceability model.
Okto Pulse maintains an embedded per-board Knowledge Graph for durable project memory.
Agents use the graph to:
Operational health is visible through:
GET /health is a constant-time liveness endpoint: it performs no storage scan and returns HTTP 200 with status "healthy" while the process can answer requests. Use application workflows or a verified backup/restore procedure for recovery; direct SQL repair is unsupported.
Okto Pulse ships as two packages: okto-pulse-core owns the SDLC domain, the governance gates
and the Knowledge Graph contracts as pure Protocol seams; okto-pulse (this package) owns
every concrete mechanism — SQLite, Okto Grafx, the filesystem, the scheduler, telemetry state,
the REST app and the MCP host.
Core never imports Community. Community fills the ports at startup, and an unfilled slot fails closed rather than falling back to a silent default.
→ Architecture in full — dependency owner matrix (AF-05/AF40), registration flow, the adapter source map, and the port → adapter matrix showing which core contract each adapter module implements.
The former reset, verify-pipeline and kg command group are no longer accepted,
including their help/options. The CLI returns the standard unknown-command
error (exit 2) before accessing data. Use status for local observation; it
does not certify KG health or perform recovery. See health observations
and the historical CLI corrections.
docker run -d --name okto-pulse \
-e HOST=0.0.0.0 \
-e MCP_HOST=0.0.0.0 \
-p 8100:8100 \
-p 8101:8101 \
-v okto-pulse-data:/data \
ghcr.io/oktolabsai/okto-pulse:latest
Then open http://localhost:8100. The first startup prints the newly created
bootstrap API key once. Persisted keys cannot be recovered; regenerate a key
through the UI/API if needed. Automation can use init --bootstrap-key-handoff
and consume that explicit file with api-key --handoff-file.
Use the production compose file when you want a PyPI-based image:
docker compose -f docker-compose.prod.yml build
docker compose -f docker-compose.prod.yml up -d
Use the local compose file when hacking on the community package together with a sibling okto-pulse-core checkout:
docker compose build
docker compose up -d
For Grafx, generation is the Pulse default. Pulse owns each managed generation directory and
performs restore/generation replacement only with its handles closed, satisfying this policy's
closed lifecycle. It amortizes descriptor identity proofs only for Grafx's canonical heap, catalog,
index and WAL names; control and unknown names remain strict. This removes repeated namespace
system calls from page-heavy graph reads without changing Grafx locking, WAL, OCC or snapshot
rules. Keep it for the ordinary local Pulse runtime where no other tool mutates the live generation.
strict proves the physical identity behind every cached descriptor hit. Select it when external
tools may touch the database directory, during manual maintenance or forensics with uncertain
directory provenance, on mixed-trust hosts, with live file replacement/replication that swaps
names, or on an unsupported shared filesystem. Its advantage is detecting an out-of-protocol path
replacement at the next physical operation; its cost is repeated namespace/descriptor calls on hot
files. Conversely, generation can defer detection of that unsupported mutation until directed
invalidation, generation advance or reopen. The policy is process-local and not persisted; it does
not alter OCC, WAL, durability or multiwriter/multireader guarantees. Pulse fixes one policy per
Grafx pool and refuses a handle whose observed effective mode differs from configuration. The full
whitelist and transition matrix are specified in Okto Grafx's
docs/architecture/ST2_DESCRIPTOR_REVALIDATION.md.
MCP admission is intentionally scoped to tool execution. Saturated calls receive
a bounded, retryable outcome with next_action.rel=retry_after; initialization,
streaming, resources, prompts, and the API/UI listener do not enter this queue.
All default local state lives under ~/.okto-pulse/:
~/.okto-pulse/
|-- data/
| `-- pulse.db
|-- boards/
| `-- {board-id}/
| |-- graph_backend_binding.json
| `-- grafx/
| `-- {generation}/
|-- global/
| |-- graph_backend_binding.json
| `-- grafx/
| `-- {generation}/
|-- uploads/
| `-- {board-id}/
`-- mcp_traces/
[!WARNING] Do not delete graph database directories to "fix" graph errors. Use the KG health and recovery tools so runtime issues remain diagnosable. Pulse 0.4.0 accepts only its current storage format; incompatible storage is refused without conversion or automatic deletion. Select a new storage location for a fresh installation.
Clone both repositories next to each other:
git clone https://github.com/OktoLabsAI/okto-pulse-core.git
git clone https://github.com/OktoLabsAI/okto-pulse.git
cd okto-pulse
Install both packages in editable mode:
pip install -e ../okto-pulse-core -e .
okto-pulse init
okto-pulse serve
Build the frontend before packaging:
cd frontend
npm install
npm run build
cd ..
Restore network access and restart:
okto-pulse serve
Check that the MCP port in .mcp.json matches the running server:
okto-pulse serve --api-port 8100 --mcp-port 8101
okto-pulse init --agents
If running in Docker, expose the MCP listener with MCP_HOST=0.0.0.0 and publish the port.
First confirm that only one okto-pulse serve process is using the same data directory. Then open board settings and check:
generation for Pulse-managed paths;
strict for forensic or externally shared paths)The DLQ Inspector in Settings → Event Queue can redrive an individual row or
all accessible rows after their root cause is fixed. Redrive all requires an
explicit UI confirmation and drains the board DLQ through bounded 200-row
transactions. Both modes are permission-gated, idempotent and wake the
consolidation worker.
Use the contextual error message as the source of truth when reporting an issue.
Current: 0.4.0 — BREAKING CHANGE. This release requires a new installation with fresh storage. Data from previous Pulse versions is incompatible and is refused without conversion or automatic deletion; no migration is provided. Community delivers unified validation, structured requirements and verification, separate implementation and verified-delivery progress, and a Grafx-only KG.
→ Full release notes — 0.4.0, 0.3.3, 0.3.2, 0.3.1 and 0.3.0 changesets, plus 0.2.6, 0.2.5, 0.2.3, 0.2.2, 0.2.1 and 0.2.0.
See Local/remote integration and recovery regression for the preserved local changes and their integration with the Grafx-only runtime.
See Open PR review for 0.3.3 for integration decisions, compatibility evidence and deferred dependency migrations.
See CLI, configuration and badge-refresh fixes (#84–#88) for export error behavior, terms storage, CORS configuration and validation evidence.
See Delivery evidence: committed code and test-card verification for the separate Spec completion gate, UI, REST/MCP contracts and audited exemptions.
See Code Evidence Matrix: associations and coverage for the distinction between contextual references and applicable evidence.
See Cognitive Action Center: review knowledge gaps for the human review workflow, waiver effects, failed-processing navigation and permissions.
See KG Health: observe, diagnose and recover for the operations dashboard, contextual help, action impacts and recovery safeguards.
Community uses Okto Grafx only, pinned to the published okto-grafx[accel]==0.0.7 release. uv.lock resolves Grafx from the official PyPI artifacts. See Grafx-only runtime, settings, retirement and data preservation. The Core remains storage-agnostic.
The executable ownership matrix is generated by okto-pulse-saas-closure. Every transitional budget must remain zero; the command fails closed on import, dependency, adapter, wheel, or documentation drift.
Elastic License 2.0 - free for personal and commercial use. You may not provide this software to third parties as a hosted or managed service.
Copyright 2026 Okto Labs
Local-first, spec-driven SDLC workbench with governance gates, knowledge graphs, and MCP for AI coding agents.
도구
아직 도구 목록이 없습니다. 원격 서버는 첫 상태 확인 후에 도구 목록이 생깁니다.
FAQ
이 페이지에서 Cursor, Claude Desktop, Claude Code 또는 VS Code용 stdio 설정을 복사합니다. PlugMCP 프록시가 아니라 게시자의 설치 방식을 사용합니다.
이 레코드에는 필수 환경 변수가 등록되어 있지 않습니다.
Okto Pulse은(는) stdio를 통해 로컬에서 실행됩니다.
마지막 확인: 2026-10-27. Local-first, spec-driven SDLC workbench with governance gates, knowledge graphs, and MCP for AI coding agents.