Okto Pulse
A local-first SDLC workbench that governs AI-assisted delivery with explicit specifications, quality gates, and durable project memory.
- Source repo
- OktoLabsAI/okto-pulse
- Stars
- ★ 139
- Last updated
- 3d ago
- License
- NOASSERTION
- Primary language
- Python
- FA score
- 67/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude CodeClaude.ai (Partial support)
- Cost
- Free, no paid service needed
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A team using Codex, Claude Code, Cursor, or another MCP client wants requirements, acceptance criteria, and test scenarios established before code changes begin.
- Not a fit if
- Teams that need to offer the software to third parties as a hosted service
- Teams unwilling to adopt enforced specification, validation, and governance workflows
- Environments unable to run Python 3.11 or later
- Website
- oktolabs.ai/platform/pulse ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-17
- Source review
- 67/100 · Some gaps
What does this agent do, and when should you use it?
Okto Pulse is a local-first software-development workbench that organizes stories, ideation, refinement, specifications, sprints, tasks, tests, and bugs into a traceable workflow. It combines okto-pulse-core, which owns the domain model, governance gates, and Knowledge Graph contracts, with a community package that supplies SQLite, Okto Grafx, filesystem, scheduler, REST, React UI, and FastMCP implementations. Humans inspect and steer work in the browser while coding agents create and update the same structured artifacts through a local MCP server. okto-pulse serve runs the Web UI/API and MCP listeners in one Python process to preserve the embedded graph database's single-writer model. Its outputs include linked specifications and execution cards, validation evidence, task conclusions, code-investigation receipts, and searchable per-board knowledge graphs. It suits teams that want self-hosted process governance and project memory, but it does not replace the coding agent or execute source-code investigations itself.
The workflow begins with okto-pulse init, which creates local state under ~/.okto-pulse/, seeds a default board and agent, and writes a project-local .mcp.json. After okto-pulse serve starts, people work through the UI/API at http://localhost:8100, while coding agents call the Core tool catalog through http://localhost:8101/mcp. Pulse stores Stories, Ideation, Refinement, Specs, Sprints, Tasks, Tests, and Bugs as structured artifacts with statuses, lineage, and validation rules; transitions check specification coverage, resource readiness, test evidence, bug traceability, task conclusions, and independent review requirements. Its per-board Knowledge Graph consolidates specifications, bugs, decisions, and implementation conclusions so agents can retrieve related decisions, identify contradictions or superseded context, reuse bug lessons, and query Global Discovery across boards. The CLI also exposes okto-pulse code-traceability records, verify-pipeline, KG health diagnostics, schema migration, backfill, JSON-LD export, and local metrics management. The community runtime does not acquire, clone, browse, probe, or inspect source code; an external agent performs those actions in its own environment and submits bounded results.
- A team using Codex, Claude Code, Cursor, or another MCP client wants requirements, acceptance criteria, and test scenarios established before code changes begin.
- A long-running product team needs searchable local memory built from previous specifications, decisions, bugs, and implementation conclusions.
- An engineering lead wants AI-assisted work blocked until coverage, validation evidence, task conclusions, and architecture findings satisfy defined gates.
- A privacy-conscious team wants SQLite data, graph storage, uploads, and metrics to remain on its own machine or Docker volume.
- A feature owner needs to approve a specification before dividing a large change into sprints and reviewable implementation batches.
- An operator needs health endpoints, dead-letter metrics, and KG recovery commands for diagnosing an embedded project knowledge graph.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
pip install okto-pulsedocker run -d --name okto-pulse \
- Environment variables
MCP_HOST- Config files
.mcp.json- Runtime
- Python 3.11
- Install notes cover
- Docker
How do you install or deploy this agent?
Python 3.11 or later is required. Install the package from PyPI:
pip install okto-pulseInitialize it inside the project directory where the coding agent will work:
okto-pulse initOn first run, Pulse attempts to download all-MiniLM-L6-v2 into the Hugging Face cache for Knowledge Graph semantic search. If that download fails, the application still starts in deterministic stub mode, with semantic search disabled. No external LLM API key is documented as a requirement; local MCP access uses a Pulse agent key revealed once after creation or rotation.
How do you use this agent?
Start the Web UI/API and MCP server:
okto-pulse serveOpen http://localhost:8100, select the default board, and create either a Story or an Ideation. The default MCP endpoint is http://localhost:8101/mcp, and many agent tools can discover the generated .mcp.json when launched from the project directory. Its generated shape is:
{
"mcpServers": {
"okto-pulse": {
"url": "http://localhost:8101/mcp?api_key=dash_..."
}
}
}For Codex, create an agent or regenerate its key under Menu → Agents, choose Codex (CLI), and run the copied command:
codex mcp add okto-pulse --url "<Pulse MCP URL with agent key>"After rotating a key, copy the new command and restart the Codex session; the old key stops working, and board access must still be granted in Pulse. Regenerate agent configuration after changing the MCP port:
okto-pulse init --agentsA published container image is also available:
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:latestRetrieve the container's bootstrap API key with:
docker exec okto-pulse okto-pulse api-keyWhat are this agent's strengths and limitations?
- People and coding agents operate the same specification, task, test, and bug artifacts through complementary Web UI and MCP interfaces.
- Governance covers scenarios, business rules, API contracts, tasks, test evidence, bug traceability, conclusions, and architecture findings rather than merely tracking cards.
- SQLite, Okto Grafx, uploads, and metrics remain local by default, with a documented persistent Docker-volume deployment.
- The Knowledge Graph preserves lineage, prior decisions, contradictions, superseded context, bug lessons, and cross-board discovery data.
- MCP response projections named
summary,detail, andfulllet users trade response depth against agent context consumption.
- Enforced transitions and coverage checks add specification, evidence, conclusion, and maintenance overhead for teams that prefer unstructured rapid coding.
- Semantic search depends on downloading
all-MiniLM-L6-v2; without it, Pulse runs only with deterministic stub behavior and semantic search disabled. - The community runtime does not inspect source code itself, so an external coding agent must handle repository access, capability checks, and investigations.
- Embedded persistence enforces a single-writer model, and MCP execution has fixed writer, concurrency, and queue admission limits.
- The README identifies Elastic License 2.0 and prohibits offering the software as a hosted or managed service to third parties, while repository license metadata is
NOASSERTION.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Okto Pulse This agent | 67 · Some gaps | CLIFree | ★ 139 | 3d ago | Python | Codex · Claude Code |
| MCP Memory Service | 68 · Some gaps | MCP serverFree | ★ 2.1k | 1d ago | Python | ChatGPT · Codex · Claude Code · Claude.ai |
| MemoMind Memory Layer | 45 · Major gaps | MCP serverFree + model costs | ★ 603 | 3mo ago | Python | Claude Code · OpenAI API · Claude API |
| Agents Towards Production | 44 · Major gaps | Library / SDKFree + model costs | ★ 22k | 6d ago | Jupyter Notebook | — |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows local storage and loopback listeners as defaults, and documents one-time MCP key display, rotation, board grants, and shell-history exposure. The CLI describes a private one-time credential handoff, while DLQ redrive is permission-gated, confirmed in the UI, and transaction-bounded. These support ordinary controls for privilege, confirmation, data-flow transparency, sensitive data, external effects, and recovery, earning 2s. Dependencies have version bounds and a stated Authlib security rationale, but the static packet does not establish comprehensive auditing or ongoing vulnerability mitigation, so dependency security is not full marks. Authors, contact details, and a reporting path are present, but registry verification is absent; source attribution earns 1.
README, TOKEN-USAGE, and pyproject disagree on tool count or version: README says 0.4.0 and 285 tools, pyproject says 0.4.2, and TOKEN-USAGE says 338 tools. SECURITY still lists 0.1.x as supported, which conflicts with the current-version signals; self-consistency earns 1. Python requirements, dependency ranges, model-download fallback, and local storage are documented, earning 2 for dependency availability. Startup timeouts, saturation retry outcomes, database errors, and graph recovery limits have actionable messages, earning 2 for failure messages. This static review does not infer successful operation.
The materials target developers and teams using coding agents, describe a story-to-task workflow, cover several agent clients, and include Docker, local, and source setups; audience and scenarios earn 3. Core/Community ownership, MCP admission scope, permission boundaries, and unsupported direct SQL repair are described, but the sample does not expose the full tool set or all trigger rules, so capability boundaries and trigger precision earn 2 each. Python, Docker, ports, caches, and local data paths are documented, though environment-specific coverage remains limited, earning 2 for environment fit.
The README has a useful contents list, document index, and sections for architecture, setup, CLI, storage, and troubleshooting, earning 3 for information architecture. Installation, agent connection, Docker, and source steps are usable, but the packet does not include all referenced files, so install notes earn 2. Product names and commands are mostly stable, while version, tool-count, and supported-version conflicts reduce naming stability to 2. Examples and troubleshooting are present but FAQ coverage is limited, earning 2. Specific limitations cover model fallback, incompatible storage, lack of migration, graph recovery, and health-check scope, earning 3. The license text is explicit but includes hosted-service, competing-service, and branding restrictions, so license earns 2 rather than implying unrestricted open-source terms. Release notes exist, but the README and pyproject versions conflict, earning 1 for versioning and changelog. A security contact and reporting process are provided, but ownership and ongoing maintenance commitments are not fully established, earning 2.
The product describes structured work items, lineage, governance gates, and a shared MCP/UI workflow that can produce inspectable artifacts; since this is static review, output usability earns 2. The staged workflow, evidence gates, and durable project memory provide a clear claimed contribution beyond direct prompt-to-code work, earning 3 for marginal value. TOKEN-USAGE reports an estimated 50.8K-token fixed connection cost and notes large response payloads, alongside mitigation ideas; value is apparent but costs are material, earning 2 for cost-benefit.
README points to architecture, release, health, token-cost, and recovery documents, and gives concrete configuration examples, earning 2 for claim traceability. The packet itself contains version and count conflicts across README, pyproject, token documentation, and security policy, so cross-source corroboration earns 1. TOKEN-USAGE labels its figures as estimates and gives a measurement method; the historical E2E report distinguishes observations, inferences, and work not performed. That supports fact/inference separation, but these remain repository claims not independently verified here, earning 2.
- The packet disagrees on release version, MCP tool count, and supported security versions; treat these statements as unverified current-state claims.
- The MCP catalog and large responses may consume substantial context; review tool-loading strategy, data-directory permissions, credentials, and network binding before deployment.
- The license restricts hosted and competing services and imposes branding requirements; review the full terms before deployment or redistribution.
FAQ
Does Okto Pulse require a paid model or API?
Will it start without network access?
all-MiniLM-L6-v2 download cannot complete, Pulse starts in deterministic stub mode, but Knowledge Graph semantic search is disabled. Initial package and model acquisition still require a network source.Does connecting an agent automatically grant access to every board?
Does Pulse clone and inspect repositories on its own?
Can I offer it as a customer-facing hosted service?
Is Okto Pulse open source?
Is Okto Pulse safe to use?
Is Okto Pulse still maintained?
Who makes Okto Pulse?
What is Okto Pulse's official website?
Thanks — we use this to improve the page.