Where Coral came from.

Coral was extracted from Phoebe, the team's AI SRE product.

• Team built a self-hosted Phoebe just as Claude Code matured
• Concluded: future of agents is compose-your-own
• Unbundled Phoebe's data-retrieval layer → Coral, standalone
• User reaction was strong enough to pivot the whole company
• Open-sourced four weeks ago. 4.9k stars in five weeks.

The thesis, in the CEO's words.

"The best agent needs 12 different sources of data and a lot of them sensitive. It's a different kind of strategic and security iteration than having Datadog have your logs and GitHub have your repo and Slack have your communications. The future of agents will gravitate towards compose-your-own — to preserve strategic flexibility and to own the asset. Companies need to decide if they want the value of this usage to accrue to them or accrue to a vertical AI provider."

— Matt Henderson, May 27 2026

The pitch, in one paragraph.

Your agent's effectiveness depends on having access to twelve sensitive sources at once. Vertical AI products want that data and the derived intelligence to accrue to them. Coral is the agent-agnostic data runtime that keeps both with you.

Apache 2.0. Local-first. Compose your own agent on top — Claude Code, Cursor, whatever ships next — and switch harnesses without losing what your usage has taught the system.

Owned infrastructure. Owned intelligence. Owned future.

The architectural lead is perishable.

Five-way intersection nobody else hits.

• Apache 2.0 + local-first + Rust + MCP-native + cross-source SQL JOINs
• 3.4× cost efficiency, 31% accuracy gain vs direct provider MCPs (Claude Opus 4.6, n=82)
• 4.9k stars in 5 weeks — top 1% velocity for new dev tools
• 27 bundled + 75+ community sources shipped in month one

But Steampipe has 150+ plugins. MindsDB has 39k stars. Composio has 28k (and a recent enterprise-trust setback). The window to lead is 6–12 months wide.

The wedge.

More cost-efficient than direct provider MCPs on real coding tasks. No other vendor publishes a comparable benchmark. We can own the metric definition.

What we're racing against.

• Steampipe — 150+ plugins, AGPL, could relicense + add MCP
• MindsDB — 39k stars, ships MCP today, could pivot agent-cost narrative
• Composio — SF darling, MIT, 28k stars; recent security incident weakened enterprise trust
• Memory plays — Memori, MEM0, Cognee competing for the same "agent infrastructure" mind-share
• Anthropic — could ship first-party MCP aggregator
• Vertical AI vendors — Datadog Bits, Honeycomb MCP, Cursor — want to own the data and the derived intelligence

Matt's framing: "100 other smart teams in adjacent spaces. We need to be the winner or one of the winners."

Memory positioning.

Coral is shipping a memory feature. Memori, MEM0, and Cognee are all funded and pitching "we are your memory layer."

Matt's framing: "Either bring your own memory tool and use Coral with it, or use Coral for both."

• "Bring your own memory" reference architectures — Memori × Coral, MEM0 × Coral, Cognee × Coral
• "Coral as memory" reference architecture — head-to-head, same dataset
• Memory in the quarterly benchmark — public leaderboard, four configurations
• Strategic argument: agent-agnostic + non-model-vendor — same reasoning that applies to data also applies to memory

Why DevRel is the lever.

Coral does not need more engineering hours. Coral needs:

• Community-led plugin velocity to close the breadth gap
• Public reference customers to neutralize "zero logos"
• A defining content narrative before incumbents pivot

All three are DevRel programs. None are engineering programs.

One north star.

Active weekly developers running ≥3 sources per user.

PMF (installed) + activation (configured) + stickiness (multiple sources). One number, all the answers.

12-month targets.

• 20k GitHub stars (4× current)
• 250+ source specs (2.5× current)
• 5,000 weekly active developers
• 20 named customer references, 5 quotable publicly
• 50 Coral Cloud paying teams by Q4 2026 ($500K ARR)
• 8,000 newsletter subscribers

Seven initiatives.

• 1. Source Spec Marketplace — attacks W1 (breadth)
• 2. BYO Agent Reference Architectures — attacks S8 + O5
• 3. Token Efficiency Benchmark Series — attacks S2 + T3
• 4. Coral Co-Pilot Cohort — attacks W3 (customers) + W2 (Cloud pipeline)
• 5. Source Spec Bounty — attacks W1
• 6. Community + Content Engine — attacks W3 + W8 (brand)
• 7. Strategic Channel Integrations — attacks T6 + T5

Coral Co-Pilot Cohort.

12 weeks · 8–10 companies · 4 batches/year.

• White-glove from founders. 50% Cloud-for-life discount.
• In exchange: case study, quote, co-presented talk.
• Public Demo Day livestream — manufactures social proof at scale.
• Targets: Vercel, Linear, Stripe, Cursor, Replit, Cognition, Decagon, Sierra, Modal, Browserbase.

Source Spec Bounty.

Annual budget. Targets ~100 new community specs at 70% acceptance.

• T1 $200 · T2 $500 · T3 $1,000 · T4 $2,000
• $5K top quarterly contributor · $10K grand prize
• Quarterly themed Hack Weeks

Quarterly Benchmark Series.

Coral is the only vendor publishing token-efficiency benchmarks for agent workloads. We turn that into a publishing cadence.

• One major drop per quarter + ad-hoc on model releases
• Open-source harness, fully reproducible
• Compares against Composio, Anthropic ref MCPs, Steampipe, MindsDB, vendor MCPs
• Public leaderboard at withcoral.com/benchmarks

BYO Agent reference architectures.

Be the default data layer for the top 5 AI coding agents.

• Claude Code — co-marketed with Anthropic
• Cursor — featured Marketplace entry
• Continue.dev — OSS-friendly natural ally
• Sourcegraph Cody / Amp — code-graph × ops-data
• Goose, Aider, Devin, Cline — long-tail acquisition

Weekly cadence builds trust.

• Monday — newsletter draft
• Tuesday 11am PT — Office Hours (recorded)
• Wednesday — community "How I Coral" post
• Thursday — YouTube source spec walkthrough
• Friday — Discord recap + bounty leaderboard

Phase 1 — Foundations · Mo 1–3.

• Source Spec Cookbook v1
• Claude Code + Cursor reference architectures
• Benchmark drop on Opus 4.7
• Discord investment, weekly office hours launch
• Cohort #1 applications open
• Bounty program launches with $50K initial pool

Milestone: 8k stars · 130 sources · 1.5k WAD

Phase 2 — Coral Cloud GA · Mo 4–6.

• Coral Cloud / Teams launches
• Multi-tenant SaaS · audit logs · RBAC · shared credential vault
• AWS Marketplace listing
• Backstage plugin GA
• Cohort #1 Demo Day livestream
• Second benchmark drop

Milestone: 12k stars · 180 sources · 3k WAD · 25 Cloud teams

Phase 3 — Scale · Mo 7–12.

• AI Engineer World's Fair + KubeCon presence
• Anthropic co-marketed launch
• Plugin SDK v2 + community marketplace
• Cohorts #3 and #4
• Gartner Cool Vendor / Forrester Wave inclusion targets

Milestone: 20k stars · 250 sources · 5k WAD · 50 Cloud teams · $500K ARR

The real GTM: managed services.

Matt: "The revenue is in infrastructure or services. Customers need a developer tools team that knows their shit with agents. So we'll end up providing managed services that are like having an on-demand agent developer tools team."

DevRel manufactures the proof the managed-services tier sells:

• Reference implementations double as a "look what we can build for you" portfolio
• Cohort case studies with quantified ROI become sales narratives
• Public office hours let prospects verify the team's depth
• Quarterly "Build Days" — live-streamed enterprise co-builds
• DevRel engineer as bench overflow — 20% time on paid services, partial revenue from day one

What we deliver.

The same operating system in any of three engagement options.

• Always-on: strategy, content, community, benchmarks, cohorts, bounties, events, analyst, brand
• Project-based: reference architectures, cookbook, Cloud GTM, Backstage plugin
• Tooling: OpenClaw agent · MetricsBoard · Cohort Portal · Bounty Tracker

Three engagement options.

Matt: "We might end up concluding that we really need to be sort of all-in from the get-go." Pick the container that fits.

• A · Full-time, all-in. Close out Nebius over 60 days, join as Head of DevRel / DevRel Engineer. Standard SF comp + equity.
• B · 90-day paid pilot → full-time conversion. $155K for 90 days with defined OKRs, convert at the gate. De-risks both sides.
• C · Fractional with concrete OKRs. $35–50K/mo retainer + programs. Operating system, lighter commitment, scale post-Cloud.

Substance is identical in all three. Only exclusivity and ramp differ.

Why opencolin.

Same operating system, four prior programs.

• CrewAI — multi-agent orchestration DevRel
• Memori — agent memory + SF residency + MCP marketplace
• Glean — enterprise AI search dev programs
• Perplexity Sonar — 90-day growth + ecosystem vision
• Nebius (current) — agent + AI infrastructure DevRel; Tavily complement available for joint events

Plus our internal AI DevRel OS — OpenClaw agent, MetricsBoard, Cohort Portal — and a fellows network across SF, NYC, London, Berlin, Bangalore, São Paulo.

Next steps.

1. Matt + James align on which of the three engagement options fits Coral now
2. 30-minute follow-up with James to brief him on this deck + the live site
3. If pilot path: sign 90-day pilot ($155K) — kick off Source Spec Cookbook + Cohort #1 prep in week 1
4. If full-time path: confirm comp + equity, plan 60-day Nebius wind-down

collin@dabl.club