Coding Agent Cost Optimization#
Dosu's primary commercial wedge is reducing the cost of running AI coding agents (Claude Code, Cursor, Codex, Copilot) by roughly 50%, while also making them ~2x faster and ~3x more consistent . The savings come from a single root cause: without Dosu, agents re-research the codebase from scratch on every run, burning tokens. With Dosu, agents query an always-fresh knowledge base via MCP instead.
Dosu is AI-native knowledge infrastructure — not a coding agent itself. It makes existing agents cheaper and more reliable by ensuring they start every session with pre-synthesized context rather than rediscovering it.
The Mechanism#
Knowledge ingestion happens automatically: every PR triggers a knowledge review and suggested doc update; Slack/Teams tribal knowledge is captured before retention policies delete it . Sources include GitHub repos, Notion, Confluence, Slack channels, and GitHub wikis .
Context delivery to agents happens through the Dosu MCP Server, which exposes these tools to connected AI assistants:
| Tool | Purpose |
|---|---|
init_knowledge | Starting point — searches curated knowledge (approved answers, topics). Built-in knowledge covers 800+ public open-source projects. |
search_documentation | Searches raw docs (Notion, Confluence, GitHub wikis) when init_knowledge isn't enough. |
search_threads | Searches GitHub issues, Slack threads — for bug reports, past decisions, troubleshooting history. |
ask | Multi-step research agent across all selected data sources; returns answers with citations and confidence scores. |
save_topic | Lets agents write durable knowledge back to the store, closing the feedback loop. |
The recommended workflow for agents is: init_knowledge → list_available_data_sources → ask .
CLI Hooks (via npx @dosu/cli setup) integrate Dosu into Claude Code and other MCP-compatible clients. When a knowledge gap is detected (no prior knowledge on the topic), the CLI injects a nudge prompting the agent to call save_topic after completing work — creating a self-improving knowledge loop.
Benchmark Numbers#
All figures from internal benchmarks :
| Metric | Without Dosu | With Dosu |
|---|---|---|
| Cost/run | ~$0.84 | ~$0.42 (~50% reduction) |
| Runtime | ~2m34s | ~1m23s (~2x faster) |
| Consistency (repeatability) | ~25% | ~80% (~3x improvement) |
| Cost variance per run | $0.46–$1.74 | $0.23–$0.52 |
One reference customer (12-person team) cut Codex spend ~40% overnight via the MCP. Indeed reduced per-run runtime from ~3 min to ~1.5 min .
Target Customer#
Dosu leads with cost savings to AI-forward enterprises — companies that have already adopted coding agents and are feeling the pain of high token spend and inconsistent outputs . Company size is secondary; agent adoption is the non-negotiable qualifier. Buyer titles are VP Eng, CTO, or DevEx/Platform leads who answer to the CFO for coding-agent spend.
The cost wedge is evaluatable in ~2 weeks with no behavior change required — which is why it leads every pitch over Dosu's broader platform capabilities (Slack Q&A, doc review, doc search) .
Key References#
- Introduction to Dosu — product overview
- Does Dosu have an MCP Server? — MCP setup and tools
- 7 - MCP — MCP deployment configuration
- GTM & Sales Playbook — full benchmark numbers, ICP, objection handling
- Knowledge Contribution via MCP Tools — how agents write knowledge back via
save_topic
